<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[Semi Doped]]></title><description><![CDATA[The Daily Brew of Semiconductors. News and analysis from Vik Sekar and Austin Lyons.]]></description><link>https://daily.semidoped.com</link><image><url>https://substackcdn.com/image/fetch/$s_!pSMR!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd1d09112-7a08-4b61-9599-274d200a77cf_1254x1254.png</url><title>Semi Doped</title><link>https://daily.semidoped.com</link></image><generator>Substack</generator><lastBuildDate>Sun, 22 Nov 2026 23:13:07 GMT</lastBuildDate><atom:link href="https://daily.semidoped.com/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Semi Doped]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[semidoped@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[semidoped@substack.com]]></itunes:email><itunes:name><![CDATA[Semi Doped]]></itunes:name></itunes:owner><itunes:author><![CDATA[Semi Doped]]></itunes:author><googleplay:owner><![CDATA[semidoped@substack.com]]></googleplay:owner><googleplay:email><![CDATA[semidoped@substack.com]]></googleplay:email><googleplay:author><![CDATA[Semi Doped]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[Sunday Pub Quiz]]></title><description><![CDATA[In the early 1900s, Nottingham chemist Frederic Kipping made a compound with the formula Ph&#8322;SiO.]]></description><link>https://daily.semidoped.com/p/sunday-pub-quiz-937</link><guid isPermaLink="false">https://daily.semidoped.com/p/sunday-pub-quiz-937</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Sun, 22 Nov 2026 13:05:25 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/05451887-6295-428e-b168-ca1ce48dfd08_250x239.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In the early 1900s, Nottingham chemist Frederic Kipping made a compound with the formula Ph&#8322;SiO. Its formula matched benzophenone, Ph&#8322;CO, a ketone, so he named it by analogy. Kipping knew the analogy did not hold because his compound formed a polymer, and the ketone does not. The name stuck anyway, and it is why writers still confuse a bathroom sealant with the material chips are made from. </p><p>What word did he coin?</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Thanks for reading! Subscribe for free to receive new posts and support our work.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div><h3>CLUES:</h3><blockquote><p>The word differs from the name of the element by a single letter.</p></blockquote><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!k60G!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png 424w, https://substackcdn.com/image/fetch/$s_!k60G!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png 848w, https://substackcdn.com/image/fetch/$s_!k60G!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png 1272w, https://substackcdn.com/image/fetch/$s_!k60G!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!k60G!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png" width="462" height="441.672" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:239,&quot;width&quot;:250,&quot;resizeWidth&quot;:462,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!k60G!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png 424w, https://substackcdn.com/image/fetch/$s_!k60G!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png 848w, https://substackcdn.com/image/fetch/$s_!k60G!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png 1272w, https://substackcdn.com/image/fetch/$s_!k60G!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff509821d-d55d-4f28-9ad4-4cc182196cd5_250x239.png 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div><figcaption class="image-caption">X crystallizes in a <a href="https://en.wikipedia.org/wiki/Diamond">diamond</a> <a href="https://en.wikipedia.org/wiki/Cubic_crystal_system">cubic crystal structure</a> by forming <a href="https://en.wikipedia.org/wiki/Orbital_hybridization">sp<sup>3</sup> hybrid orbitals</a>. By Ben Mills - Own work, Public Domain, https://commons.wikimedia.org/w/index.php?curid=2722584</figcaption></figure></div><blockquote><p><a href="https://en.wikipedia.org/wiki/J%C3%B6ns_Jacob_Berzelius">J&#246;ns Jacob Berzelius</a> discovered the the material; chemists today call these compounds siloxanes, after their silicon-oxygen backbone.</p></blockquote><h3>SPOILER ALERT: ANSWER BELOW</h3><div class="pullquote"><h4>Answer: <a href="https://en.wikipedia.org/wiki/Silicon">Silicon</a>. It is often misspelled as <a href="https://en.wikipedia.org/wiki/Silicone">Silicone</a>.</h4></div><div class="captioned-button-wrap" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/sunday-pub-quiz-937?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="CaptionedButtonToDOM"><div class="preamble"><p class="cta-caption">Thanks for reading! This post is public so feel free to share it.</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/sunday-pub-quiz-937?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://daily.semidoped.com/p/sunday-pub-quiz-937?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p></div><p></p><p></p>]]></content:encoded></item><item><title><![CDATA[TIL: Ethernet Ran at 2.94 Megabits Because a Clock Didn’t Fit]]></title><description><![CDATA[A Hawaiian radio network, a failed PhD and a paper found on a friend&#8217;s couch led to the cable that connects almost every computer.]]></description><link>https://daily.semidoped.com/p/til-ethernet-ran-at-294-megabits</link><guid isPermaLink="false">https://daily.semidoped.com/p/til-ethernet-ran-at-294-megabits</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Sat, 21 Nov 2026 13:30:46 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!Qz6v!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Ethernet&#8217;s first speed, 2.94 megabits per second, was set by a part that didn&#8217;t fit on a circuit board. We&#8217;ll get to the board later. The story starts with a Stanford professor killing time at Honolulu airport.</p><p><a href="https://en.wikipedia.org/wiki/Norman_Abramson">Norm Abramson</a> was on his way to give a paper in Tokyo, and his plane stopped in Honolulu just long enough for him to decide it was <a href="https://archive.computerhistory.org/resources/access/text/2020/12/102746750-05-01-acc.pdf">a pretty nice place</a>. On the way back he changed his flight, rented a surfboard at the Moana on Waikiki and learned to surf at thirty-three. About a year later he was teaching at the University of Hawaii, and surfed every day for thirty years (without a wetsuit).</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!Qz6v!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Qz6v!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg 424w, https://substackcdn.com/image/fetch/$s_!Qz6v!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg 848w, https://substackcdn.com/image/fetch/$s_!Qz6v!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!Qz6v!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!Qz6v!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg" width="500" height="700.4103967168263" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1024,&quot;width&quot;:731,&quot;resizeWidth&quot;:500,&quot;bytes&quot;:432918,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/jpeg&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://daily.semidoped.com/i/222159405?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Qz6v!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg 424w, https://substackcdn.com/image/fetch/$s_!Qz6v!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg 848w, https://substackcdn.com/image/fetch/$s_!Qz6v!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!Qz6v!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17e41c5e-da00-45ff-954e-b8c26fb8a86f_731x1024.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">By Norman Abramson - Norman Abramson, Public Domain, https://commons.wikimedia.org/w/index.php?curid=6850114</figcaption></figure></div><p>The university&#8217;s campuses sat on islands <a href="https://www.eng.hawaii.edu/wp-content/uploads/2020/06/abramson1985-Development-of-the-ALOHANET.pdf">within about 300 kilometres of Honolulu</a>, and phone lines to its main computer were poor and expensive. In 1968 his group started planning a radio network instead, and the rule they settled on sounds almost careless: any terminal could transmit whenever it liked, and if two packets collided, each waited a random time and tried again. And just like it is today, memory was expensive, so sending and receiving at once needed a second 704-bit buffer at about $300 a terminal, so they made do with one.</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Thanks for reading! Subscribe for free to receive new posts and support our work.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div><p><a href="https://en.wikipedia.org/wiki/ALOHAnet">ALOHAnet</a> went live in June 1971. Abramson never patented it, because he was &#8220;<a href="https://archive.computerhistory.org/resources/access/text/2020/12/102746750-05-01-acc.pdf">too busy surfing to worry about that sort of thing</a>,&#8221; and published instead, which is how his paper ended up in a blue 1970 conference volume beside a foldout couch in Washington.</p><p>That couch belonged to <a href="https://en.wikipedia.org/wiki/Steve_Crocker">Steve Crocker</a>, an <a href="https://en.wikipedia.org/wiki/ARPANET">ARPA</a> program manager, and <a href="https://en.wikipedia.org/wiki/Robert_Metcalfe">Bob Metcalfe</a> slept on it when ARPA work brought him to town. In May 1972 Metcalfe had <a href="https://ethw.org/Oral-History:Robert_Metcalfe">failed his Harvard PhD defense</a> and had to tell his parents not to come to Harvard Yard. The committee decided his thesis on building the ARPANET was engineering rather than theory. And most engineers without PhDs are aware of the short end of the stick often enough: theory is credited with nuance, while engineering is accused of brute-forcing things into working, however amazing the result. He went to <a href="https://en.wikipedia.org/wiki/PARC_(company)">Xerox PARC</a> anyway.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!UEck!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg 424w, https://substackcdn.com/image/fetch/$s_!UEck!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg 848w, https://substackcdn.com/image/fetch/$s_!UEck!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!UEck!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!UEck!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:null,&quot;width&quot;:null,&quot;resizeWidth&quot;:344,&quot;bytes&quot;:126008,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/jpeg&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://daily.semidoped.com/i/222159405?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!UEck!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg 424w, https://substackcdn.com/image/fetch/$s_!UEck!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg 848w, https://substackcdn.com/image/fetch/$s_!UEck!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!UEck!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F856ec5c1-c3f9-4d25-a51e-264e968cbbd0_355x486.jpeg 1456w" sizes="100vw"></picture><div></div></div></div><figcaption class="image-caption">By Andreu Ve&#224;, WiWiW.org - Own work, CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=115665925</figcaption></figure></div><p>So he was on the couch, jetlagged, when he took the blue volume off the shelf <a href="https://archive.computerhistory.org/resources/text/Oral_History/Metcalfe_Robert_1/Metcalfe_Robert_1_2.oral_history.2006.7.102657995.pdf">to read himself to sleep</a> and found Abramson&#8217;s paper, or so Metcalfe tells it. <a href="https://elists.isoc.org/pipermail/internet-history/2020-December/006772.html">Crocker remembers</a> showing him the paper himself. Either way, it was Crocker&#8217;s copy.</p><p>The math annoyed Metcalfe because Abramson had modelled infinite users who kept typing no matter what, but the terminals didn&#8217;t work like that: you sent in a card image, a little light went off, and you waited for the computer&#8217;s reply to switch it back on. He redid the model with a few users who wait for the light, and noticed the original never said how long a terminal should wait before trying again; set that too short and the channel went unstable with almost any traffic. Harvard got its theory, and he got his PhD in June 1973. </p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/til-ethernet-ran-at-294-megabits/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://daily.semidoped.com/p/til-ethernet-ran-at-294-megabits/comments"><span>Leave a comment</span></a></p><p>At PARC, the <a href="https://en.wikipedia.org/wiki/Xerox_Alto">Alto</a> and laser printer needed a network, and <a href="https://en.wikipedia.org/wiki/Charles_Simonyi">Charles Simonyi </a>had sketched one called SIGNET, for <a href="https://archive.computerhistory.org/resources/text/Oral_History/Metcalfe_Robert_1/Metcalfe_Robert_1_2.oral_history.2006.7.102657995.pdf">Simonyi&#8217;s Infinitely Glorious Network</a>, built to move 50 megabits per second when terminals ran at 300 bits per second. Butler Lampson moved Simonyi to a text editor and handed SIGNET to Metcalfe, who had photographed the building&#8217;s wiring closets, rats&#8217; nests of cable running every terminal to one room. He wanted one cable down the hall, and computers that listened for a quiet line, talked, and backed off when they collided.</p><p>Hawaii&#8217;s radio was no use: the ALOHA modem was about six feet tall, and his network had to fit on a card inside an Alto. He ordered a thousand-foot spool of coaxial cable, took it to the basement and started sending pulses down it, and was struggling with connectors when a Stanford graduate student in a yellow button-down shirt said he knew how to do that and did it. That was <a href="https://en.wikipedia.org/wiki/David_Boggs">David Boggs</a>, and the two were partners from then on, later sharing the patent with <a href="https://en.wikipedia.org/wiki/Butler_Lampson">Butler Lampson</a> and <a href="https://en.wikipedia.org/wiki/Charles_P._Thacker">Chuck Thacker</a>, who designed the Alto.</p><p>On May 22, 1973, Metcalfe wrote a memo renaming the Alto ALOHA Network the Ether Net, after the medium physicists once believed carried light, until <a href="https://en.wikipedia.org/wiki/Michelson%E2%80%93Morley_experiment">Michelson and Morley</a> looked for it and found nothing. The first two machines were called Michelson and Morley in their honor.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!2JJV!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!2JJV!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg 424w, https://substackcdn.com/image/fetch/$s_!2JJV!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg 848w, https://substackcdn.com/image/fetch/$s_!2JJV!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!2JJV!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!2JJV!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg" width="728" height="486.75125208681135" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:801,&quot;width&quot;:1198,&quot;resizeWidth&quot;:728,&quot;bytes&quot;:179443,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/jpeg&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://daily.semidoped.com/i/222159405?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!2JJV!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg 424w, https://substackcdn.com/image/fetch/$s_!2JJV!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg 848w, https://substackcdn.com/image/fetch/$s_!2JJV!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!2JJV!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cf4e446-abca-42d2-834c-cddd20fffb37_1198x801.jpeg 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">By Robert Metcalfe - https://broadbandlibrary.com/bob-metcalfe-lays-down-the-law. Archived, Public Domain, https://commons.wikimedia.org/w/index.php?curid=197114262</figcaption></figure></div><p>As for the board, it filled up until Boggs had to hang an error-checking chip off an extension at the edge, which left no room for a clock. The Alto already had a clock, <a href="https://ethw.org/Oral-History:Robert_Metcalfe">ticking every 170 nanoseconds</a>, so they borrowed it. At two ticks per bit, that came to 2.94 megabits per second, and Metcalfe refused for years to round it to 3, because the rounding error was bigger than an entire ARPANET trunk line.</p><div class="captioned-button-wrap" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/til-ethernet-ran-at-294-megabits?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="CaptionedButtonToDOM"><div class="preamble"><p class="cta-caption">Thanks for reading! This post is public so feel free to share it.</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/til-ethernet-ran-at-294-megabits?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://daily.semidoped.com/p/til-ethernet-ran-at-294-megabits?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p></div><p></p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 20th, 2026]]></title><description><![CDATA[TSMC and Samsung halt Chinese chip orders, SK hynix becomes Kioxia&#8217;s top shareholder, SpaceX sells $32 billion of GPU debt, and a short seller targets AMD]]></description><link>https://daily.semidoped.com/p/daily-update-november-20th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-20th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Fri, 20 Nov 2026 14:55:31 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/04e53f2c-082c-4fd3-9ba3-ffa3f35de9c2_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Friday, November 20th, 2026</strong></p><p>Washington&#8217;s new rule reached the fab door this week, and TSMC and Samsung started turning Chinese orders away. SK hynix quietly became Kioxia&#8217;s biggest shareholder, SpaceX paid over 8% to fund $47 billion of leased GPUs, and a short seller decided Nvidia shouldn&#8217;t be the only chipmaker with a China problem.</p><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>Google Built Two New TPUs Without New Silicon: The 8i Max and 8TX</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=Cq-JY_JiMfs">YouTube</a>, <a href="https://podcasts.apple.com/us/podcast/google-built-two-new-tpus-without-new-silicon-the-8i/id1866707196?i=1000802500123">Apple Podcasts</a>, or <a href="https://open.spotify.com/episode/7DPrPlp0zEW3jYAnMqNCkT">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-google-built-two-new">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p><mark data-color="#d9d2e9" style="background-color: rgb(217, 210, 233); color: rgb(0, 0, 0);">Fine-tune and serve open models without managing a single GPU.</mark></p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p><strong>Accounts are free, and get $5 in credit when you sign up.</strong></p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning &#8594;</a></p></div><h3>TSMC and Samsung stop taking Chinese chips on trust</h3><p>The remote-access rule that took effect on Thursday also closes the foundry route. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">TSMC has told affected Chinese designers to halt orders at 16nm and below unless they get U.S. licenses, generally limiting them to 28nm and 40nm parts without HBM or advanced packaging, Reuters reports.</mark> TSMC said only that it will comply with the new rules.</p><p>Until now, foundries and packagers could check a Chinese design&#8217;s specs and ship it without a license. That check is gone. The FT says TSMC and Samsung have told Chinese clients some shipments will wait for licenses, and Samsung&#8217;s customers there include ByteDance and Baidu&#8217;s Kunlunxin. Even phone chips for Xiaomi, Oppo and Vivo are caught, though those licenses are likely to come.</p><p>The overseas GPU detour is closing too. Moonshot AI scrapped a $150 million deal to rent Blackwell GPUs from an Indonesian neocloud, per The Information, and Bloomberg says lenders paused a $480 million loan for a ByteDance-serving Malaysian data center.</p><p><a href="https://www.investing.com/news/stock-market-news/tsmc-tightens-china-ai-chip-curbs-as-new-us-rules-close-manufacturing-loophole-sources-say-5038806">Source: Reuters via Investing.com</a> &#183; <a href="https://www.federalregister.gov/documents/2026/11/19/2026-23529/addressing-remote-access-to-advanced-semiconductors">Federal Register</a> &#183; <a href="https://www.digitimes.com/news/a20261120VL218/tsmc-taiwan-chips-communications-samsung.html">Digitimes</a> &#183; <a href="https://www.trendforce.com/news/2026/11/20/news-tsmc-samsung-reportedly-pause-china-consumer-chip-shipments-under-new-u-s-rules-16nm-orders-affected/">TrendForce</a> &#183; <a href="https://www.theinformation.com/articles/chinas-moonshot-ai-scraps-150-million-deal-access-nvidia-chips-indonesia">The Information</a> &#183; <a href="https://www.thejakartapost.com/business/2026/11/20/chinese-ai-firm-moonshot-reportedly-scraps-150m-chip-deal-in-indonesia">The Jakarta Post</a> &#183; <a href="https://www.trendforce.com/news/2026/11/20/news-u-s-ai-chip-curbs-hit-southeast-asia-gpu-financing-bytedance-linked-us480m-loan-reportedly-paused/">TrendForce on GPU loans</a></p><h3>SK hynix becomes Kioxia&#8217;s largest shareholder</h3><p>SK hynix said on Friday it will convert &#165;129 billion (about $810 million) of convertible bonds tied to its 2018 Toshiba Memory investment into shares, and buy Bain&#8217;s last sliver of the holding vehicle. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Once the conversion closes on Dec. 4, SK hynix will indirectly own 14.1% of Kioxia, ahead of Toshiba&#8217;s 10.46%.</mark></p><p>The new cash outlay is just &#165;100 million, and no Kioxia shares change hands. What changes is the voting: as a shareholder rather than a creditor, SK hynix can now vote the Kioxia stake through the holding vehicle.</p><p>SK hynix says it has no plans to buy more, sell, or seek a bigger role in management, and its agreement with Kioxia caps its voting rights at 15% until 2028 without Kioxia&#8217;s consent. Still, a NAND rival that already owns Solidigm now holds the biggest single stake in Kioxia.</p><p><a href="https://en.yna.co.kr/view/AEN20261120005051320">Source: Yonhap</a> &#183; <a href="https://www.mk.co.kr/en/business/12214783">Maeil Business Newspaper</a> &#183; <a href="https://en.edaily.co.kr/news/eda202611205519/">Edaily</a></p><h3>SpaceX sells $32 billion of GPU debt</h3><p>SpaceX sold $32 billion of bonds on Thursday to pay for Nvidia chips, part of a $40 billion-plus package that also includes a bank loan, Bloomberg reports. The money buys about $47.3 billion of Vera Rubin NVL72 equipment, with deliveries starting in January.</p><p>The structure keeps the debt off SpaceX&#8217;s balance sheet. An Apollo-backed trust buys the racks and leases them to SpaceX for about five years. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">SpaceX must keep paying even if the GPUs fail or a disaster strikes, and that promise is what the investment-grade ratings lean on.</mark></p><p>The price says investors aren&#8217;t sure how to value used GPUs. The largest tranche yielded 8.26%, against 5.95% for the average U.S. high-grade bond. Expect more deals built this way.</p><p><a href="https://theedgemalaysia.com/node/826632">Source: Bloomberg via The Edge Malaysia</a> &#183; <a href="https://www.kbra.com/publications/McYcSQYT/kbra-assigns-preliminary-a-ratings-with-stable-outlook-to-saturn-funding-trust-s-senior-and-subordinated-secured-notes">KBRA</a> &#183; <a href="https://www.zaobao.com.sg/en/news/world/spacex-issues-us32-billion-bonds-finance-nvidia-ai-chip-purchases">Lianhe Zaobao</a></p><h3>Hunterbrook points the China lens at AMD</h3><p>Short seller Hunterbrook said on Thursday that IEIT Systems, a Chinese server maker the U.S. blacklisted in March 2025, kept putting newer AMD firmware into its servers, at least six versions since then. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Three IEIT-linked vendors, including AMD&#8217;s Instinct partner Aivres, posted nearly identical firmware.</mark> The report doesn&#8217;t show who supplied the code, or whether a license covered it.</p><p>AMD said the claims &#8220;misrepresent AMD&#8217;s business practices and relationships.&#8221; The stock closed down 7.7% and fell further before Friday&#8217;s open.</p><p>The same day, House Foreign Affairs chair Brian Mast released license data showing Nvidia and AMD sought to sell 3.4 million AI chips to China between December and May, with more than 879,000 approved. Mast flagged two firms in Nvidia&#8217;s applications as suspected smuggling fronts, and one of them, Maginfra, also turns up in Hunterbrook&#8217;s report. Separately, California&#8217;s attorney general subpoenaed Nvidia over whether chips or confidential information reached China.</p><p><a href="https://hntrbrk.com/amd-china">Source: Hunterbrook</a> &#183; <a href="https://www.cnbc.com/2026/11/19/amd-falls-on-short-seller-report-of-alleged-china-exposure.html">CNBC</a> &#183; <a href="https://foreignaffairs.house.gov/news/press-releases/chairman-brian-mast-exercises-legislative-authority-to-publicly-release-ai-chip-license-information-for-china">House Foreign Affairs Committee</a> &#183; <a href="https://www.cnbc.com/2026/11/19/nvidia-amd-china-chip-sales-congress.html">CNBC on license data</a> &#183; <a href="https://oag.ca.gov/news/press-releases/attorney-general-bonta-serves-investigative-subpoena-nvidia">California DOJ</a></p><h3>Samsung hands memory to its HBM developer</h3><p>Samsung promoted Hwang Sang-joon, its head of memory development, to president and put him in charge of the memory business on Friday. He led DRAM and HBM development and will keep running DRAM development, The Korea Herald reports.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">All three new presidents come from the chip division.</mark> Cho Sang-yeun moves from running Samsung&#8217;s U.S. chip business to System LSI, where Samsung wants him to win new customers, and process-development chief Lee Jong-myeong becomes the chip division&#8217;s CTO.</p><p>Co-CEO Jun Young-hyun gives up the memory job to focus on the wider chip business, including foundry. Choi Won-joon takes over the mobile business.</p><p><a href="https://news.samsung.com/global/samsung-electronics-announces-new-leadership-5">Source: Samsung</a> &#183; <a href="https://www.koreaherald.com/article/10943786">The Korea Herald</a> &#183; <a href="https://www.digitimes.com/news/a20261120VL202/samsung-business-hbm-development-dram.html">Digitimes</a></p><div class="callout-block" data-callout="true"><h4>Supported by Meter</h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter-daily">See how Meter builds data center networks</a></p></div><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>Google TPU packaging</strong> &#8212; Google tested Intel&#8217;s EMIB-T for its next TPUs, found yields short of target, and asked TSMC for help again (<a href="https://www.moneydj.com/kmdj/news/newsviewer.aspx?a=04b30cff-be94-4191-a2a1-f28389d7cf68">MoneyDJ</a> &#183; <a href="https://www.newsquawk.com/headlines/intels-intc-emib-t-have-reportedly-failed-to-meet-googles-googl-expectations-which-has-led-google-to-ask-tsmc-tsm-if-it-can-switch-back-to-cowos-for-its-tpus-moneydj-reports">Newsquawk</a>).</p></li><li><p><strong>SK hynix hybrid bonders</strong> &#8212; SK hynix will buy its first batch of hybrid bonders for next-generation HBM from Besi, ahead of Korean rivals (<a href="https://www.digitimes.com/news/a20261120VL212/sk-hynix-hbm-equipment-dutch-production.html">Digitimes</a>).</p></li><li><p><strong>AMD Zen 6 desktop</strong> &#8212; AMD moved its Zen 6 desktop Ryzen chips to TSMC&#8217;s N2X process, pushing the launch past Computex 2027 (<a href="https://www.techpowerup.com/355030/amd-zen-6-for-desktop-reportedly-pushed-to-computex-2027">TechPowerUp</a> &#183; <a href="https://videocardz.com/newz/amd-zen-6-desktop-ryzen-cpus-reportedly-delayed-until-after-computex-2027">VideoCardz</a>).</p></li></ul><h2>Sector Watch</h2><p><strong>Earnings</strong></p><ul><li><p><strong>Keysight</strong> &#8212; Fourth-quarter orders rose 53% and revenue 41% to $1.99 billion, with commercial communications sales up 73%. (<a href="https://investor.keysight.com/investor-news-and-events/financial-press-releases/press-release-details/2026/Keysight-Technologies-Reports-Fourth-Quarter-and-Fiscal-Year-2026-Results/default.aspx">Source: Keysight</a> &#183; <a href="https://www.investing.com/news/transcripts/earnings-call-transcript-keysight-beats-q4-2026-estimates-shares-rise-after-hours-93CH-5040253">Investing.com</a>)</p></li></ul><p><strong>Memory and compute</strong></p><ul><li><p><strong>Micron</strong> &#8212; Micron moved the first tool into its $7 billion Singapore HBM packaging plant, with initial output expected in early 2027. (<a href="https://www.edb.gov.sg/en/news-and-insights/micron-marks-first-tool-move-in-at-singapore-hbm-advanced-packaging-facility">Source: Micron via Singapore EDB</a> &#183; <a href="https://www.trendforce.com/news/2026/11/20/news-micron-singapore-hbm-plant-receives-first-equipment-ahead-of-2027-output-as-taoyuan-labor-dispute-persists/">TrendForce</a>)</p></li><li><p><strong>Alibaba</strong> &#8212; Alibaba is training its next model on in-house T-Head chips, aiming to move all training to domestic chips by 2027. (<a href="https://www.trendforce.com/news/2026/11/20/news-alibaba-begins-ai-model-training-with-t-head-chips-targets-full-domestic-adoption-by-2027-amid-new-u-s-curbs/">Source: TrendForce</a> &#183; <a href="https://www.digitimes.com/news/a20261120PD234/alibaba-t-head-ai-chip-computing-power-capacity-2027.html">Digitimes</a>)</p></li><li><p><strong>Microsoft</strong> &#8212; Nvidia Vera Rubin NVL72 systems are now running in Azure data centers, with Azure virtual machines opening in preview. (<a href="https://blogs.nvidia.com/blog/microsoft-ignite-accelerated-computing-azure-windows-pcs/">Source: Nvidia</a> &#183; <a href="https://www.digitimes.com/news/a20261120VL206/nvidia-microsoft-azure-rubin-windows.html">Digitimes</a>)</p></li><li><p><strong>Google</strong> &#8212; Google picked Blue Origin&#8217;s New Glenn to launch its TPU-carrying Project Suncatcher satellites in early 2029. (<a href="https://blog.google/innovation-and-ai/models-and-research/google-research/blue-origin-project-suncatcher-partnership/">Source: Google</a> &#183; <a href="https://spacenews.com/new-glenn-to-launch-project-suncatcher-data-center-satellites-for-google/">SpaceNews</a>)</p></li></ul><p><strong>Optics and networking</strong></p><ul><li><p><strong>Dell&#8217;Oro</strong> &#8212; Optical vendors&#8217; revenue from big North American clouds rose 75% last quarter, and vendors are now taking orders for 2028 delivery. (<a href="https://www.prnewswire.com/news-releases/75-percent-increase-in-cloud-provider-spending-on-optical-transport-equipment-in-3q-2026-according-to-delloro-group-302946278.html">Source: Dell&#8217;Oro Group</a>)</p></li><li><p><strong>Coherent</strong> &#8212; The optical component maker&#8217;s board authorized a $2 billion share buyback with no expiration date. (<a href="https://www.sec.gov/Archives/edgar/data/820318/000119312526467737/d867801d8k.htm">Source: SEC filing</a>)</p></li><li><p><strong>Samsung</strong> &#8212; Samsung bought four silicon photonics patents from Rockley Photonics, covering optical modulators, multiplexers and filters for data center links. (<a href="https://www.trendforce.com/news/2026/11/20/news-samsung-electronics-expands-silicon-photonics-portfolio-through-patent-acquisitions/">Source: TrendForce</a>)</p></li></ul><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark></p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/intel-custom-asics">Intel Custom ASICs?</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/hbm-lower-die-count-stacks-and-custom-base-die">HBM: Lower Die-Count Stacks and Custom Base-Die Performance</a></p></li></ul><p><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better!</mark> <a href="http://semidoped.com/poll"><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Give us your feedback</mark></a></p><p>Want to become a sponsor? Send an email to <a href="mailto:sponsor@semidoped.com">sponsor@semidoped.com</a></p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 19th, 2026]]></title><description><![CDATA[Intel lands Microsoft and an 18A customer, China&#8217;s clouds feel the new rule, Micron sees a tighter 2027, and Anthropic raises $107 billion]]></description><link>https://daily.semidoped.com/p/daily-update-november-19th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-19th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Thu, 19 Nov 2026 14:02:33 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/69a2930b-2a02-431d-a471-4033cfcdf21d_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Thursday, November 19th, 2026</strong></p><p>Intel Foundry spent years promising outside customers, and on Wednesday it named one and signed another. Alibaba admitted Washington&#8217;s new rule will cut its access to overseas compute, Micron said the memory squeeze tightens from here, and Anthropic raised $107 billion to pay its chip bills. Meanwhile, OpenAI&#8217;s models are quietly editing AMD&#8217;s GPU blueprints.</p><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>Google Built Two New TPUs Without New Silicon: The 8i Max and 8TX</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=Cq-JY_JiMfs">YouTube</a>, <a href="https://podcasts.apple.com/us/podcast/google-built-two-new-tpus-without-new-silicon-the-8i/id1866707196?i=1000802500123">Apple Podcasts</a>, or <a href="https://open.spotify.com/episode/7DPrPlp0zEW3jYAnMqNCkT">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-google-built-two-new">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p>Fine-tune and serve open models without managing a single GPU.</p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p>Accounts are free, and get $5 in credit when you sign up.</p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning &#8594;</a></p></div><h3>Intel Foundry lands Microsoft and an 18A customer</h3><p>Microsoft will package its future Azure chips at Intel. The two said on Wednesday that Microsoft is the first company to announce use of Intel Foundry&#8217;s EMIB packaging for its first-party Azure silicon, under a long-term deal that also locks in multiyear Xeon supply, including next-generation Diamond Rapids.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Separately, Lip-Bu Tan told the TomorrowX Summit in Austin that Intel &#8220;just signed the first customer&#8221; for 18A, without naming it.</mark> CTO Pushkar Ranade said the customer is on 18A-P, the improved version that entered risk production in June.</p><p>EMIB gives hyperscalers a second route for packaging big AI chips while TSMC&#8217;s CoWoS capacity stays stretched. A leading-edge wafer customer is the harder prize, and Intel has chased one for five years. Socionext already announced 18A-P design work in August, and Intel didn&#8217;t say how the new deal differs.</p><p><a href="https://www.intel.com/content/www/us/en/newsroom/news/corporate/microsoft-azure-infrastructure-collaboration.html">Source: Intel</a> &#183; <a href="https://news.microsoft.com/source/2026/11/18/microsoft-and-intel-deepen-collaboration-to-deliver-differentiated-azure-infrastructure/">Microsoft</a> &#183; <a href="https://www.cnbc.com/2026/11/18/intel-first-18a-foundry-customer.html">CNBC</a> &#183; <a href="https://www.oregonlive.com/silicon-forest/2026/11/intel-has-its-first-customer-for-its-most-advanced-manufacturing-process.html">The Oregonian</a> &#183; <a href="https://thenextweb.com/news/intel-ceo-lip-bu-tan-says-intel-has-signed-its-first-outside-18a-customer">The Next Web</a></p><h3>China&#8217;s cloud giants count the cost of the remote-access rule</h3><p>Washington&#8217;s remote-access rule took effect on Thursday with its publication in the Federal Register, and China&#8217;s biggest clouds are already writing it into their results. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Alibaba said the rule &#8220;will impact our company&#8217;s ability to access compute from certain providers outside China,&#8221; and that it can&#8217;t yet estimate the financial hit.</mark></p><p>Alibaba is still spending. September-quarter capex rose 60% to 50.5 billion yuan, partly on higher chip-component prices, and AI cloud revenue grew 53%. Tencent said on Wednesday the curbs will likely slow its next model, though its inference-heavy cloud can run on domestic chips.</p><p>The pain is spreading. Japan&#8217;s Datasection, which press reports have tied to Tencent, cut its revenue forecast 89% this week after a large AI infrastructure order was cancelled. Chinese GPU designers including Hygon and Cambricon slid on Thursday on rumors that overseas foundries will review Chinese orders. Trump, for his part, told Fox News that Blackwell sales to China could come &#8220;in two years or three years.&#8221;</p><p><a href="https://www1.hkexnews.hk/listedco/listconews/sehk/2026/1119/2026111900873.pdf">Source: Alibaba results (HKEX)</a> &#183; <a href="https://www.federalregister.gov/documents/2026/11/19/2026-23529/addressing-remote-access-to-advanced-semiconductors">Federal Register</a> &#183; <a href="https://www.investing.com/news/stock-market-news/tencent-quarterly-revenue-growth-slows-cloud-unit-benefits-from-ai-demand-5035220">Reuters via Investing.com</a> &#183; <a href="https://www2.jpx.co.jp/disc/39050/140120261116581693.pdf">Datasection filing</a> &#183; <a href="https://www.investing.com/news/stock-market-news/japans-datasection-slumps-on-big-guidance-cut-amid-uschina-chip-access-row-5037539">Investing.com</a> &#183; <a href="https://en.infomaxai.com/news/articleView.html?idxno=149366">Yonhap Infomax</a> &#183; <a href="https://www.straitstimes.com/asia/east-asia/trump-suggests-us-can-sell-blackwell-chips-to-china-in-future">The Straits Times</a></p><h3>Micron says the memory squeeze tightens from here</h3><p>Micron CEO Sanjay Mehrotra told the TomorrowX Summit on Wednesday that 75% of the company&#8217;s 2027 production is already sold, and customers are now negotiating 2028. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">He expects 2027 and 2028 to be even tighter than 2026.</mark></p><p>The contracts are the story. Micron has signed 26 supply and collaboration agreements, many running five years or longer, backed by $32 billion of upfront cash and financial commitments, take-or-pay terms and pricing bands. Mehrotra argued customers won&#8217;t walk away because memory now sets the economics of AI factories.</p><p>Next comes stacked 3D DRAM. Micron says every major customer is asking to partner on it, with market penetration starting in 2028 to 2029.</p><p><a href="https://www.investing.com/news/transcripts/micron-at-tomorrowx-summit-ai-memory-demand-reshapes-outlook-93CH-5036868">Source: Micron at TomorrowX (Investing.com)</a> &#183; <a href="https://stockanalysis.com/stocks/mu/transcripts/801555-tomorrowx-summit/">Transcript (StockAnalysis)</a></p><h3>Anthropic raises $107 billion to pay for compute</h3><p>Anthropic priced its IPO at $83 a share on Wednesday, above its $70 to $78 range, in a 980 million-share offering worth $81.3 billion, the largest stock-market debut on record. A $26 billion private placement at the same price takes the total to $107.3 billion. Trading starts on Nasdaq on Thursday under ANT.</p><p>That money already has a destination. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Anthropic has committed to nearly $900 billion of compute spending with partners including Amazon and Google, Reuters reports.</mark> Compute costs rose another $5 billion in the third quarter, tipping it to a $567 million operating loss.</p><p>The filing also names the soft spot in the buildout. Its compute providers lean on third-party financing to buy chips, and if that dries up, contracted capacity &#8220;may be delayed, repriced, reduced, or not delivered.&#8221;</p><p><a href="https://www.anthropic.com/news/anthropic-prices-initial-public-offering">Source: Anthropic</a> &#183; <a href="https://www.investing.com/news/stock-market-news/anthropic-prices-blockbuster-ipo-at-83-a-share-source-says-5037110">Reuters via Investing.com</a> &#183; <a href="https://www.npr.org/2026/11/18/nx-s1-6046212/anthropic-record-ipo-price">NPR</a></p><h3>OpenAI&#8217;s models are redesigning AMD&#8217;s GPUs</h3><p>OpenAI said on Wednesday its models rewrote the SystemVerilog code behind AMD&#8217;s flagship GPU IP and found up to 14% gains in power, performance and area over heavily optimized baselines. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">AMD has already put some of the changes into a design it sent to manufacturing, and added others to production files for future chips.</mark></p><p>OpenAI calls it roughly a year of progress in under five months. AMD also verified single-digit PPA gains on key blocks of its next flagship CPU, and the same approach cut standard-cell area by about 13% in the vector unit of OpenAI&#8217;s own Jalape&#241;o chip.</p><p>Every rewrite was checked with Synopsys tools, and OpenAI is now training a model with Synopsys to drive EDA software like an expert designer. &#8220;Up to&#8221; is doing some work in that headline number, since OpenAI gave no block-by-block breakdown.</p><p><a href="https://openai.com/index/advancing-chip-design-with-our-models/">Source: OpenAI</a> &#183; <a href="https://thenextweb.com/news/openai-models-amd-gpu-chip-design-14-percent">The Next Web</a></p><div class="callout-block" data-callout="true"><h4>Supported by Meter</h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter-daily">See how Meter builds data center networks</a></p></div><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>Nvidia&#8217;s next CEO</strong> &#8212; Heidrick &amp; Struggles, the search firm hired by Nvidia&#8217;s board, has approached Arm CEO Rene Haas about the job (<a href="https://news.bloomberglaw.com/business-and-practice/nvidia-sizes-up-arm-ceo-haas-in-search-for-huang-successor">Bloomberg Law</a> &#183; <a href="https://www.business-standard.com/world-news/nvidia-sizes-up-arm-holdings-ceo-rene-haas-in-search-for-huang-successor-126111900173_1.html">Business Standard</a>).</p></li><li><p><strong>Intel 18A yields</strong> &#8212; Intel&#8217;s 18A yields have climbed to about 70%, a sharp turn from the &#8220;horrible&#8221; yields Lip-Bu Tan inherited (<a href="https://www.trendforce.com/news/2026/11/19/news-intel-scores-dual-foundry-wins-microsoft-named-first-emib-customer-tan-says-18a-secures-first-external-deal/">TrendForce</a>).</p></li><li><p><strong>SK hynix hybrid bonding</strong> &#8212; SK hynix has installed hybrid bonders from Hanwha Semitech and Besi&#8211;Applied Materials at Cheongju to test 20-high HBM5 stacks (<a href="https://www.munhwa.com/article/11632525">Munhwa Ilbo</a>).</p></li><li><p><strong>Terafab tools</strong> &#8212; Musk&#8217;s Terafab told Korean toolmaker GST it needs thousands of scrubbers and chillers and asked how many it can deliver (<a href="https://www.thelec.net/news/articleView.html?idxno=15735">The Elec</a>).</p></li><li><p><strong>Anthropic and SpaceX</strong> &#8212; Elon Musk says Anthropic&#8217;s most likely plan is to rent another 10 gigawatts of compute from SpaceX (<a href="https://www.benzinga.com/markets/tech/26/11/63419499/elon-musk-says-anthropic-extremely-likely-to-expand-spacex-compute-deal-could-seek-10-gigawatts-more-thats-a-big-number">Benzinga</a> &#183; <a href="https://investinglive.com/stocks/musk-says-anthropic-wants-another-10-gigawatts-of-spacex-compute-over-30-times-its-deal-alt-headlines/">investingLive</a>).</p></li></ul><h2>Sector Watch</h2><p><strong>Earnings</strong></p><ul><li><p><strong>Kulicke &amp; Soffa</strong> &#8212; Fiscal fourth-quarter revenue more than doubled to $410.3 million, and the bonder maker guided flat for the December quarter. (<a href="https://investor.kns.com/2026-11-18-Kulicke-Soffa-Reports-Fourth-Quarter-2026-Results">Source: Kulicke &amp; Soffa</a>)</p></li><li><p><strong>Soitec</strong> &#8212; The wafer maker signed about $1.2 billion of deposit-backed photonics-SOI capacity deals and expects that business to more than triple this year. (<a href="https://www.soitec.com/home/group/corporate/newsroom/press-releases/content/2026/11/18/soitec-reports-second-quarter-revenue-and-half-year-results-of-fiscal-year-2027">Source: Soitec</a> &#183; <a href="https://www.techtimes.com/articles/331033/20261118/soitec-earnings-ai-wafer-maker-books-75m-photonics-sales-customers-commit-12b.htm">Tech Times</a>)</p></li><li><p><strong>Lenovo</strong> &#8212; Revenue jumped 48% to a record $30.2 billion on AI demand, but non-cash charges left a $640 million net loss. (<a href="https://investor.lenovo.com/en/financial/results/press_2627_q2.pdf">Source: Lenovo</a> &#183; <a href="https://www.bloomberg.com/news/articles/2026-11-19/lenovo-sales-beat-estimate-due-to-solid-ai-hardware-demand">Bloomberg</a>)</p></li></ul><p><strong>Memory, packaging and fabs</strong></p><ul><li><p><strong>JEDEC</strong> &#8212; The standards body previewed DDR6, with server modules starting at 12.8 GT/s and over 50% more bandwidth than DDR5. (<a href="https://www.jedec.org/news/pressreleases/jedec%C2%AE-previews-next-generation-ddr6-dram-higher-performance-greater-capacity">Source: JEDEC</a> &#183; <a href="https://www.trendforce.com/news/2026/11/19/news-jedec-previews-preliminary-ddr6-specs-targeting-over-50-higher-bandwidth-and-12-8-gts-speeds/">TrendForce</a>)</p></li><li><p><strong>ASE</strong> &#8212; ASE&#8217;s Taiwan Foredge unit broke ground on a NT$248.9 billion AI chip test plant in Kaohsiung, due in 2028 and 2029. (<a href="https://www.digitimes.com/news/a20261119PD228/testing-taiwan-ase-plant-subsidiary.html">Source: Digitimes</a> &#183; <a href="https://ec.ltn.com.tw/article/breakingnews/5645968">Liberty Times</a>)</p></li><li><p><strong>Rapidus</strong> &#8212; Japan&#8217;s government is weighing up to &#165;54 billion of tax relief when it hands the Hokkaido fab over to Rapidus. (<a href="https://www.jiji.com/jc/article?g=eco&amp;k=2026111801092">Source: Jiji Press</a> &#183; <a href="https://www.japantimes.co.jp/business/2026/11/19/companies/japan-rapidus-tax-relief/">The Japan Times</a>)</p></li></ul><p><strong>Optics and AI cloud</strong></p><ul><li><p><strong>Eoptolink</strong> &#8212; The Chinese optical module maker denied order-cut rumors, saying its 1.6T modules are in volume production and its backlog is ample. (<a href="https://www.tradingview.com/news/reuters.com,2026:newsml_L3N46Z0VK:0-eoptolink-technology-holds-positive-expectations-for-order-demand-growth-in-q4-and-next-year/">Source: Reuters via TradingView</a> &#183; <a href="https://hk.investing.com/news/stock-market-news/article-1764879">Zhitong via Investing.com</a>)</p></li><li><p><strong>Nscale</strong> &#8212; The neocloud brought its first Vera Rubin NVL72 racks online on Dell systems, with a wider rollout planned over coming months. (<a href="https://www.nscale.com/press-releases/nscale-first-nvidia-vera-rubin-nvl72-brought-up">Source: Nscale</a> &#183; <a href="https://www.dell.com/en-us/blog/dell-technologies-delivers-the-industry-s-first-fully-integrated-liquid-cooled-networking-for-nvidia-vera-rubin-nvl72-to-nscale">Dell</a>)</p></li></ul><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark></p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/intel-custom-asics">Intel Custom ASICs?</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/hbm-lower-die-count-stacks-and-custom-base-die">HBM: Lower Die-Count Stacks and Custom Base-Die Performance</a></p></li></ul><p><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better!</mark> <a href="http://semidoped.com/poll"><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Give us your feedback</mark></a></p><p>Want to become a sponsor? Send an email to <a href="mailto:sponsor@semidoped.com">sponsor@semidoped.com</a></p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 18th, 2026]]></title><description><![CDATA[Nvidia beats and raises without Huang, books $390 billion of supply, stops H200 shipments to China, and TrendForce sees HBM prices doubling while NAND turns]]></description><link>https://daily.semidoped.com/p/daily-update-november-18th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-18th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Wed, 18 Nov 2026 13:32:53 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/304282f3-b969-4481-8c7d-f91aae39dcc4_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Wednesday, November 18th, 2026</strong></p><p>Nvidia&#8217;s first earnings call without Jensen Huang opened with its chair saying he isn&#8217;t coming back, then moved on to $116 billion of business. It also tied up $390 billion of supply and quietly stopped shipping H200s to China. TrendForce, meanwhile, told memory buyers to keep their wallets open through 2027.</p><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>Google Built Two New TPUs Without New Silicon: The 8i Max and 8TX</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=Cq-JY_JiMfs">YouTube</a>, <a href="https://podcasts.apple.com/us/podcast/google-built-two-new-tpus-without-new-silicon-the-8i/id1866707196?i=1000802500123">Apple Podcasts</a>, or <a href="https://open.spotify.com/episode/7DPrPlp0zEW3jYAnMqNCkT">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-google-built-two-new">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p>Fine-tune and serve open models without managing a single GPU.</p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p>Accounts are free, and get $5 in credit when you sign up.</p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning &#8594;</a></p></div><h3>Nvidia beats and raises in its first quarter without Huang</h3><p>Nvidia reported fiscal third-quarter revenue of $115.8 billion on Tuesday, up 103% from a year ago and well above the roughly $109 billion analysts expected. Data center revenue was $107.6 billion, and Vera Rubin brought in nearly $20 billion in its first quarter of production shipments. The fourth-quarter guide is $130 billion, plus or minus 3%, with an extra week and no China data center compute in it.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Interim CEO Colette Kress raised Nvidia&#8217;s fiscal 2028 growth outlook to about 75% from 70%, the forecast investors feared would be scrapped.</mark> Gross margin was 74.8%. Kress guided 71% to 72% next quarter and called it the bottom, because Nvidia is honoring old purchase orders before new pricing kicks in, with margins back to 72% to 73% in fiscal 2028.</p><p>Board chair Harvey Jones opened the call by saying the board does not intend to reinstate Huang and has hired Heidrick &amp; Struggles to find a permanent CEO. Huang said on Tuesday he is open to coming back. Shares rose about 4% after hours.</p><p><a href="https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-third-quarter-fiscal-2027">Source: Nvidia</a> &#183; <a href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000088/q3fy27cfocommentary.htm">CFO commentary (SEC)</a> &#183; <a href="https://www.cnbc.com/2026/11/17/nvidia-nvda-earnings-report-q3-2027.html">CNBC</a> &#183; <a href="https://www.investing.com/news/transcripts/earnings-call-transcript-nvidia-tops-q3-2026-estimates-and-lifts-outlook-93CH-5034300">Call transcript (Investing.com)</a></p><h3>Nvidia&#8217;s $390 billion shopping list</h3><p>Nvidia&#8217;s supply commitments jumped to $390 billion at the end of October from $279 billion a quarter earlier. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The CFO commentary says the increase is mainly multiyear memory procurement, so Nvidia is locking up DRAM and HBM years ahead.</mark> Inventory rose to $39.7 billion for the Rubin ramp.</p><p>Nvidia also widened its substrate and packaging deals. Kinsus will grow its ABF substrate capacity dedicated to Nvidia by more than 3.5x at its Taoyuan campus, with output from the second half of 2027, and says Nvidia is investing more than $1 billion. SPIL signed a multiyear deal backed by multibillion-dollar Nvidia funding to expand packaging capacity by mid-2029, and Samsung Electro-Mechanics amended its substrate deal, with the two expanding into MLCCs.</p><p>The bank of Nvidia is getting choosier, though. Kress said neocloud support will be &#8220;disciplined and selective,&#8221; and Nvidia is &#8220;rigorously evaluating&#8221; future investments and guarantees.</p><p><a href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000088/q3fy27cfocommentary.htm">Source: CFO commentary (SEC)</a> &#183; <a href="https://nvidianews.nvidia.com/news/nvidia-expands-commitments-with-supply-chain-partners-to-support-ai-infrastructure-buildout">Nvidia</a> &#183; <a href="https://www.kinsus.com.tw/en/news/newsroom/detail/news_category1/news_20261118">Kinsus</a> &#183; <a href="https://en.yna.co.kr/view/AEN20261118002600320">Yonhap</a> &#183; <a href="https://www.taiwannews.com.tw/news/6478093">Taiwan News</a></p><h3>Nvidia stops shipping H200s to China</h3><p>Nvidia sold $2.4 billion of licensed Hopper data center chips to China last quarter. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">On the call, Kress said Nvidia is &#8220;no longer shipping licensed Hopper data center products to China,&#8221; and the guide includes none.</mark> She did not say why or whether the halt is permanent.</p><p>The H200 deal was Huang&#8217;s project, and it was never smooth. Washington approved sales to about 10 Chinese buyers, but Beijing held back imports, and Reuters has reported some 600,000 H200s sitting in Hong Kong. Earlier on Tuesday, CNBC reported that Nvidia told buyers it won&#8217;t sell to data centers that rent more than half their GPUs to Chinese companies.</p><p>The remote-access rule is landing in Southeast Asia too. Singapore booked 22% of Nvidia&#8217;s chip revenue in 2025, yet only about 1% of those chips are used there, according to The Straits Times. Prime Minister Lawrence Wong expects &#8220;some impact&#8221; on transshipment and says Singapore can help U.S. officials check where chips go.</p><p><a href="https://www.investing.com/news/stock-market-news/nvidia-stops-shipping-usapproved-ai-chips-to-china-5035204">Source: Reuters via Investing.com</a> &#183; <a href="https://www.afp.com/en/nvidia-says-no-longer-shipping-hopper-ai-chips-china">AFP</a> &#183; <a href="https://www.cnbc.com/2026/11/17/nvidia-china-cloud-ai-chip-access.html">CNBC</a> &#183; <a href="https://www.straitstimes.com/asia/spore-can-facilitate-us-checks-on-ai-chip-use-wont-allow-circumvention-of-export-controls-pm-wong">The Straits Times</a></p><h3>TrendForce: HBM prices double, NAND turns</h3><p>TrendForce told its Seoul roadshow on Wednesday that HBM4 will cost $3 to $5.50 per gigabit in 2027, up from at least $2.05 this year. Average HBM prices should rise 122%, and HBM4 should be 51% of HBM shipments.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">TrendForce does not expect conventional DRAM prices to fall in any quarter of 2027, because HBM eats three to four times the wafers of regular DRAM.</mark> Increases slow from 10% to 15% this quarter to 0% to 5% by late 2027. NAND is the weak spot: contract prices start falling in the second quarter of 2027 as Chinese cloud providers cut enterprise SSD orders.</p><p>The risk is 2028, when Samsung&#8217;s P5, SK hynix&#8217;s Yongin site and CXMT&#8217;s expansion all add capacity. For now, Nvidia&#8217;s $390 billion of commitments says the buyers agree with TrendForce.</p><p><a href="https://biz.chosun.com/en/en-it/2026/11/18/YCDJTCE5CVFFND7AW5NJRISJDA/">Source: Chosunbiz</a> &#183; <a href="https://www.tokenpost.com/news/business/60184">TokenPost</a> &#183; <a href="https://www.trendforce.com/research/download/RP261116WI">TrendForce</a></p><div class="callout-block" data-callout="true"><h4>Supported by Meter</h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter-daily">See how Meter builds data center networks</a></p></div><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>Chinese optics ban</strong> &#8212; The White House is pressing ahead with an FCC ban on imports of new Chinese optical transceiver models despite lobbying from Amazon, Google, Meta, Microsoft, CoreWeave and Cisco (<a href="https://www.investing.com/news/stock-market-news/exclusivetrumps-china-data-center-curb-advances-despite-amazon-google-meta-lobbying-5032280">Reuters via Investing.com</a> &#183; <a href="https://news.bloomberglaw.com/artificial-intelligence/us-to-push-ahead-on-china-data-center-tech-import-ban-rtrs-1">Bloomberg Law</a>).</p></li><li><p><strong>TSMC credit limits</strong> &#8212; TSMC is setting per-customer credit limits and asking some custom AI chip designers to pay cash or prepay to secure capacity, per Kuai Keji (<a href="https://investinglive.com/news/tsmc-reportedly-demands-cash-up-front-from-ai-chip-clients-amid-circular-financing-fears/">investingLive</a>).</p></li><li><p><strong>Google HBM</strong> &#8212; Google is moving to buy HBM for its TPUs directly instead of through Broadcom&#8217;s bundled deal (<a href="https://www.trendforce.com/news/2026/11/18/news-google-reportedly-eyes-direct-hbm-purchases-for-tpus-reducing-broadcom-related-costs/">TrendForce, citing Chosun Biz</a>).</p></li><li><p><strong>Honam fab sites</strong> &#8212; Samsung and SK hynix are close to splitting the Gwangju airfield chip site, with Samsung taking about 1.2 million pyeong in the south and SK hynix 1.0 million in the north (<a href="https://biz.chosun.com/en/en-policy/2026/11/18/NKLBJAMUCRE6RL6XNFCPBPGG4Q/">Chosunbiz</a> &#183; <a href="https://news.tvchosun.com/site/data/html_dir/2026/11/18/2026111890078.html">TV Chosun</a>).</p></li><li><p><strong>SpaceX GPU debt</strong> &#8212; SpaceX could sell about $4 billion of bonds as soon as this week, the first piece of a roughly $40 billion Nvidia chip financing led by Goldman Sachs and Apollo (<a href="https://finance.yahoo.com/technology/ai/articles/spacex-debt-ai-chips-hit-145310043.html">Bloomberg via Yahoo Finance</a> &#183; <a href="https://www.zaobao.com.sg/en/finance/world/spacex-may-launch-part-us40-billion-ai-chip-debt-financing-early-week-sources">Lianhe Zaobao</a>).</p></li><li><p><strong>Nvidia&#8217;s next CEO</strong> &#8212; Semafor says a minority of Nvidia directors still want Huang back, and that Marvell CEO Matt Murphy has been considered as a successor (<a href="https://www.semafor.com/article/11/17/2026/some-nvidia-directors-push-to-bring-huang-back-despite-board-opposition">Semafor</a>).</p></li><li><p><strong>Bridge Data Centres</strong> &#8212; Citi and JPMorgan added a clause to the Bain-backed operator&#8217;s $6.6 billion loan that makes breaking U.S. chip export rules a default (<a href="https://thenextweb.com/news/citi-and-jpmorgan-reportedly-add-chip-curbs-clause-to-bain-data-centre-loan">The Next Web</a>).</p></li></ul><h2>Sector Watch</h2><p><strong>Compute and cloud</strong></p><ul><li><p><strong>AMD</strong> &#8212; AMD closed its roughly $2.9 billion purchase of inference-chip startup Taalas and will fold it into the Instinct roadmap. (<a href="https://newsroom.amd.com/news/amd-completes-acquisition-taalas-advance-ai-inference-innovation/">Source: AMD</a> &#183; <a href="https://www.tipranks.com/news/amd-stock-up-after-the-company-completed-a-2-9-billion-acquisition-of-ai-chipmaker-taalas">TipRanks</a>)</p></li><li><p><strong>Microsoft</strong> &#8212; Azure&#8217;s Arm-based Cobalt 200 VMs went generally available at Ignite, with up to 50% better performance than Cobalt 100. (<a href="https://newsroom.arm.com/news/arm-microsoft-azure-cobalt-200">Source: Arm</a> &#183; <a href="https://azure.microsoft.com/en-us/blog/azure-cobalt-200-vms-now-available-proven-on-real-workloads-and-built-for-the-agentic-era/">Microsoft Azure</a>)</p></li><li><p><strong>Cerebras</strong> &#8212; Cerebras put disaggregated inference into production, with AMD GPUs handling prefill and its wafer-scale CS-3 systems handling decode. (<a href="https://www.cerebras.ai/blog/cerebras-launches-disaggregated-inference">Source: Cerebras</a> &#183; <a href="https://www.unite.ai/cerebras-puts-disaggregated-inference-into-production-with-gpus/">Unite.AI</a>)</p></li></ul><p><strong>Memory</strong></p><ul><li><p><strong>CXMT</strong> &#8212; At its third-quarter briefing, CEO Mark Cao said China&#8217;s DRAM maker has lined up long-term supply deals with several tier-one customers. (<a href="https://www.digitimes.com/news/a20261118PD226/cxmt-dram-production-demand-president.html">Source: Digitimes</a> &#183; <a href="https://www.alphabiz.co.kr/news/articleView.html?idxno=198802">Alphabiz</a>)</p></li><li><p><strong>Samsung</strong> &#8212; Samsung has reportedly chosen three next-generation DRAM tracks, including wafer bonding for HBM5 and a conventional design for mobile. (<a href="https://www.digitimes.com/news/a20261118VL214/samsung-dram-roadmap-wafer.html">Source: Digitimes</a> &#183; <a href="https://www.thelec.net/news/articleView.html?idxno=15696">The Elec</a>)</p></li></ul><p><strong>Equipment and EDA</strong></p><ul><li><p><strong>Lam Research</strong> &#8212; Lam broke ground on a technology center in Zhunan, Taiwan, opening in 2028 and growing its Taiwan lab space ninefold. (<a href="https://www.digitimes.com.tw/tech/dt/n/shwnws.asp?id=775255">Source: Digitimes</a> &#183; <a href="https://www.ctee.com.tw/news/20261118701756-431401">Commercial Times</a>)</p></li><li><p><strong>ACM Research</strong> &#8212; The wet-clean and plating toolmaker opened an R&amp;D and manufacturing site in Hillsboro, Oregon, part of a $100 million-plus U.S. buildout. (<a href="https://ir.acmr.com/news-releases/news-release-details/acm-research-expands-us-operations-oregon-rd-and-manufacturing/">Source: ACM Research</a> &#183; <a href="https://www.bizjournals.com/portland/news/2026/11/17/acm-research-opens-hillsboro-r-and-d-manufacturing.html">Portland Business Journal</a>)</p></li><li><p><strong>Siemens</strong> &#8212; Siemens agreed to buy Breker Verification Systems to add system-level chip verification to its EDA portfolio, with closing expected in December. (<a href="https://news.siemens.com/en-us/siemens-to-acquire-breker-verification-systems/">Source: Siemens</a> &#183; <a href="https://www.engineering.com/siemens-to-acquire-breker-for-expanded-soc-verification/">Engineering.com</a>)</p></li></ul><p><strong>Networking and optics</strong></p><ul><li><p><strong>Credo</strong> &#8212; Credo settled its active electrical cable patent dispute with Marvell, ending Marvell&#8217;s patent office challenges on confidential terms. (<a href="https://investors.credosemi.com/news-events/news/news-details/2026/Credo-Reaches-Settlement-in-Patent-Dispute/default.aspx">Source: Credo</a> &#183; <a href="https://www.investing.com/news/stock-market-news/credo-marvell-settle-intellectual-property-battle-over-data-center-cable-tech-5034193">Investing.com</a>)</p></li><li><p><strong>Applied Optoelectronics</strong> &#8212; The transceiver maker set up a new at-the-market program to sell up to $550 million of stock. (<a href="https://www.sec.gov/Archives/edgar/data/1158114/000110465926129055/0001104659-26-129055.txt">Source: SEC filing</a> &#183; <a href="https://www.investing.com/news/stock-market-news/applied-optoelectronics-falls-8-on-550-million-stock-offering-plan-5034098">Investing.com</a>)</p></li></ul><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark></p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/intel-custom-asics">Intel Custom ASICs?</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/why-ai-data-center-constraints-make">Why AI Data Center Constraints Make Efficiency More Valuable</a></p></li></ul><p><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better!</mark> <a href="http://semidoped.com/poll"><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Give us your feedback</mark></a></p><p>Want to become a sponsor? Send an email to <a href="mailto:sponsor@semidoped.com">sponsor@semidoped.com</a></p>]]></content:encoded></item><item><title><![CDATA[🎙️ NEW EPISODE: Google Built Two New TPUs Without New Silicon: The 8i Max and 8TX]]></title><description><![CDATA[Watch now (54 mins) | Google's TPU 8TX and 8i Max, confidential computing with AMD EPYC CPUs, TPU Sync for memory management, Boardfly vs. 3D Torus topologies, and more]]></description><link>https://daily.semidoped.com/p/new-episode-google-built-two-new</link><guid isPermaLink="false">https://daily.semidoped.com/p/new-episode-google-built-two-new</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Tue, 17 Nov 2026 15:13:13 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/b634bf3a-a4af-4d8e-a7f9-b0cf72d5daed_2000x1116.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Austin and Vik break down Google&#8217;s two new TPU systems, the 8TX and 8i Max. These aren&#8217;t new silicon, but rather new system configurations built around the existing V8 chips. They explore why Google is creating a secure version with an AMD CPU, a memory-maxed version for inference, and what the new TPU Sync software layer means for the ecosystem.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p><strong>Fine-tune and serve open models without managing a single GPU.</strong></p><p>Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour.</p><p>Accounts are free, and <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">get $5 in credit</mark> when you sign up.</p><p><strong><a href="https://semidoped.com/crusoe">Try Serverless Fine-Tuning &#8594;</a></strong></p></div><p><strong>Things we cover:</strong></p><ul><li><p>The new TPU 8TX and 8i Max systems</p></li><li><p>Why the 8i Max doubles down on memory for inference</p></li><li><p>Confidential computing and the role of AMD&#8217;s EPYC CPU</p></li><li><p>The mismatch: training silicon (8T) with an inference network (Boardfly)</p></li><li><p>TPU Sync, Google&#8217;s answer to Nvidia&#8217;s Dynamo</p></li><li><p>The business strategy of selling TPUs to external customers</p></li></ul><div class="callout-block" data-callout="true"><p><strong>This episode is supported by Meter.</strong></p><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks? Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS. One call reaches the team that designed the hardware and wrote the software.</p><p><strong><a href="https://semidoped.com/meter">See how Meter builds datacenter networks &#8594;</a></strong></p></div><p><em>This podcast is lightly edited for clarity.</em></p><h2>Four Flavors of TPU V8</h2><p><strong>Austin:</strong> Google just announced two new TPUs. Except, they really aren&#8217;t new chips. They&#8217;re just derivatives of their V8 chips announced earlier this year. So Google is taking existing silicon and building two very different systems around it. We&#8217;ll also discuss TPU sync, which looks a lot like Nvidia Dynamo. What is really Google up to here? Let&#8217;s get into it.</p><p>Hello listeners and welcome to another Semi Doped episode. I&#8217;m Austin Lyons of Chipstrat. I write a newsletter for institutional investors and technologists.</p><p><strong>Vik:</strong> I&#8217;m Vik Sekar and I write a substack called Vik&#8217;s newsletter and I also am the founder of Semi Exponent, which is an independent research firm for institutional investors. And you should check out our daily newsletter on daily.semidoped.com. We cover all the daily news around the AI and semiconductor industry and that is an entirely free newsletter to read on Substack. So you&#8217;ll have that delivered every day to your inbox along with our takes on the news. So sign up for that at daily.semidoped.com.</p><p><strong>Austin:</strong> Yes, and which by the way, I haven&#8217;t published ours today. I think you put your takes in there and you&#8217;re waiting on me, but we&#8217;re recording this on Thursday, November 12th, and we were going to record this TPU thing, but then of course, there was some big news this morning about Nvidia and Jensen Huang retiring. So instead, we just recorded a hot take and then so people will have already heard this. And now, in the same day, we&#8217;re wearing the same clothes, and it&#8217;s like almost midnight for Vik, and we&#8217;re recording this TPU one as well, which means I didn&#8217;t get the Semi Doped daily out yet. So let&#8217;s record this and then if I remember, I&#8217;ll try to add my takes and hit send.</p><p><strong>Vik:</strong> Yes, the AI world never stops. Time zones do not matter. So we have to get to it. So, what&#8217;s with the whole Google thing? Let&#8217;s get into it actually because if you saw the announcement, Austin, it&#8217;s like suddenly there were two new TPU SKUs that landed. And the earlier ones announced this year for the version 8 TPU chips was the TPU V8T for training and the V8I for inference, which was a nice way to name it and it was very understandable. And we were very happy with it when we discussed the networking that went around these chips at the time, which was basically they introduced this new network topology called Boardfly, which we explained in a YouTube video, a podcast here. And it was seemingly one of our most watched podcasts. People really seem to like it. So anybody wants to check that out, should be there in our back catalog. But the recent announcement a few days ago is that we have two more versions of this eighth generation chip, which is the TPU 8TX and the TPU 8i Max. Okay, so we&#8217;re going to go and discuss what is happening with these four versions of the same chip in one generation, that is version 8. So what do you think is going on? What&#8217;s happening here?</p><p><strong>Austin:</strong> Yes, yes, so first of all, I barely saw the news because I think this came out at the start of the week and obviously there&#8217;s all the Nvidia stuff in the news. And so most people&#8217;s eyes and attention was on the Nvidia stuff and so this was kind of as Cloud would say, it was quietly announced that Google had some extra chips here. And I was surprised. I was like, wait a minute, there&#8217;s four. Two was kind of a big deal because it used to be back in the day that for most all the generations, the TPU just had one and they use the same one for training and inference. And then I wrote an article when the 8T and 8i came out around that time, I&#8217;d been talking about like, hey, it&#8217;s the right, you need the right silicon for the right workload these days. And we&#8217;ve talked a lot about different memory hierarchy decisions and different networking decisions, not only with the TPU, but also with other AI accelerator startups that are like, oh, they&#8217;re choosing more SRAM here and more HBM here or also using DDR and all these different decisions. And so it made a lot of sense when Google came out with the TPU for training and the TPU for inference because the training one was for the more compute bound workload, and the inference one was for the memory bound and so it had more HBM. And then to your point, they had different networking and so you could see topologies, whether it&#8217;s the 3D Torus or the Boardfly. And so you could see, oh, hey, this is pretty cool. Google is designing not just chips, but entire system for training, an entire system for inference. So then when I saw that news, and I saw it just kind of fly past on my screen, and it was like, oh, also they have the 8TX and the 8i Max or whatever, I was just like, wait, we already have a system for training and a system for inference. What are these two other systems for? So, with that, what are they for? Let&#8217;s start with the 8i Max. Should we start there?</p><h2>A Refresher on 8T and 8i</h2><p><strong>Austin:</strong> This episode of Semi Doped is presented by Crusoe, where you can fine tune and serve open models without managing a single GPU. When we sat down with Crusoe&#8217;s serverless fine tuning product team, we asked the question every AI user wonders, what will you do with my data? Crusoe has a simple answer. They don&#8217;t touch it or see it. Your data is yours to fine tune models with or simply store for future use. Their fine tuning is designed so that both your training data and the fine tune model itself remain your property. Your training data is encrypted, so even Crusoe&#8217;s cloud engineering team can&#8217;t see it. You can deploy it on Crusoe Cloud in one click or download the checkpoints and run it anywhere you&#8217;d like. The fine tuning process is fully managed, so you upload a data set, pick an open model, and Crusoe takes care of the GPUs, orchestration, and all the infrastructure management. You pay per token processed rather than per GPU hour, so run a few versions, compare them, and download the one that actually works best for your use case. Create a new Crusoe account for free and get $5 in credit to start testing today. This part is super important, guys. Support Semi Doped by going to semidoped.com/cruso, click on that link in the show notes, type it in yourself, and you&#8217;ll get directed to a sign-in page. Sign in with Google or create a new account, and Crusoe will immediately create that account for you with $5 in it. Boom. Again, just click that link in the show notes, sign in, $5 for you, and you support Semi Doped. So, thank you for your support. Go check out Crusoe, try to fine tune your own data. Again, the link to try serverless fine tuning is in the show notes. So, hit that now, click through, and make sure you claim that $5, give it a try, and support us. Back to the show.</p><p><strong>Vik:</strong> I want to go and just set the stage for what the original chips were so that we could discuss what the new ones are because I want to just remind listeners what the previously announced ones this year were. So let&#8217;s start with the 8T. The 8T was basically for large scale pre-training, right? It&#8217;s a training chip. And the 8i was for inference, like that much is clear. The 8i had more HBM than the 8T. So it had 288 GB for the inference chip and 216 GB for the training chip. And the HBM bandwidth per chip was also higher in the inference chip. What was significantly higher in the inference chip was on-chip SRAM. It was almost three times as much. It&#8217;s 384 MB of SRAM on the inference chip versus the training chip. And you can see the chip too, it&#8217;s a humongous chip. The inference chip is humongous compared to the training chip and at Hot Chips the speaker from Google put up a picture and had a quiz and said, which one is the training chip? And everybody feels like they want to say that the training chip is a bigger chip for whatever reason you think the training is a hard problem and therefore bigger chip. No, no, it&#8217;s completely the opposite. It&#8217;s actually the inference chip because of the massive SRAM capacity in it was actually the inference chip. So the training chip was smaller.</p><p>So what Nvidia&#8217;s design choice for the inference chip was that they put in a lot of memory. So whether it&#8217;s HBM or SRAM, it&#8217;s memory. And this is a good reason for that is because inference is a memory bound problem and we will need this piece of information to discuss the 8i Max. That&#8217;s why I&#8217;m setting the stage. So we need to have a lot of memory because inference has basically two stages, the pre-fill and the decode, and the decode is an entirely memory problem. So it makes sense for an inference chip to have a lot of memory. So that&#8217;s why they did it that way. And the other way is that the training chip actually has more compute than the inference chip. So that&#8217;s the internals of the chip. There&#8217;s a little bit, like you said, there&#8217;s a system involved here really. So there&#8217;s a little bit more detail and that is for the scale up networking of the training chip, they use a 3D Torus topology. So this is what they have used for the longest time really with all their optical circuit switches. It&#8217;s a very complicated 3D interconnection that involves both copper and optics. And then the 8i actually uses Boardfly. And this is something to remember, and we did the whole video on Boardfly, but it&#8217;s important to remember that the inference chip used Boardfly. And we&#8217;ll see why this is important when we get to the 8TX. Because it&#8217;s very, very confusing. And then finally, both of these 8T and 8i as a head node, they use the Google Axion CPU, which is an Arm-based CPU.</p><p><strong>Austin:</strong> Yes, yes, I&#8217;m with you so far. And yeah, I was just going to add in the 3D Torus, people can remember, I think of the Rubik&#8217;s Cube where it&#8217;s like, oh, I want to talk to my neighbors in different directions, and that makes a ton of sense for training because there&#8217;s very consistent communication patterns. And then the Boardfly is good on inference because these days we&#8217;ve got mixture of experts and you don&#8217;t know exactly who you&#8217;re going to talk to. It kind of depends on which expert&#8217;s getting fired for which token. So you do need a different topology that&#8217;s suited for that.</p><p>And then I&#8217;ll add in lastly, I&#8217;m glad you called out the Axion CPU and that it&#8217;s Arm-based, and it reminds me whenever I listen to Arm&#8217;s earnings calls, I listen to Intel&#8217;s earnings call and then AMD&#8217;s and then Arm&#8217;s, and in somewhat of that order, maybe Arm&#8217;s came first, they all run together. But, there&#8217;s always like an X86 one that&#8217;ll say, oh yeah, we sell, we still have the majority of CPU server sales or whatever. And then Arm comes in and they&#8217;re like, yeah, we sell, or 60% of the CPU server sales are Arm, and you&#8217;re always like, what? I&#8217;m so confused. These add up to more than 100%. But it&#8217;s usually like one is focused on just X86 and then Arm is focused on everything. And so head node, I bring this up because head node, because we tend to think like, oh yeah, Axion and it&#8217;s in Google Cloud, but you also forget they can count head nodes. So if you see, oh, a bazillion TPUs got made, then you know, I guess, well, there&#8217;s a bazillion head nodes to go with it or a half a bazillion or whatever. And that right now for this generation, it has mostly been their Arm-based Axion, but we will get into that as well in a little bit.</p><p><strong>Vik:</strong> Yes, it&#8217;s a very, very important fact that Axion is going to be replaced by something else. But yeah, so this is the difference and the Boardfly topology was used for inference fundamentally over the 3D Torus as a refresher from the other video if you haven&#8217;t seen it, is that it reduces the overall number of hops required to go from one point to another within the networked GPUs or TPUs. And reducing the number of hops is important because you can reduce the latency. The less the networking you use, the better. So that&#8217;s the reason Boardfly is good for inference because we want latency to be low. So that&#8217;s where we are now. Now is a good point to jump into like you said, the 8i Max.</p><h2>The 8i Max: Double the Memory</h2><p><strong>Vik:</strong> So the 8i Max is nothing different from the 8i. The silicon is exactly the same. They just took two 8i chips and put them together on a single package. That&#8217;s pretty much it, right?</p><p><strong>Austin:</strong> Yes, yes. And there&#8217;s an amazing graphic that they made that we will try to include here that shows, yes, I think one server node or one pizza box if you will. I believe it used to have four individual chips. And now it has those same four columns if you will, but two chips kind of per column. So there&#8217;s eight in a pizza box instead of four. At least that&#8217;s what their graphic illustrated.</p><p><strong>Vik:</strong> Yeah. So now the question is, why would somebody take the same chip and put two copies of it and then package it together?</p><p><strong>Austin:</strong> Exactly. Or why didn&#8217;t they do that in the first place?</p><p><strong>Vik:</strong> That was my thought. So there&#8217;s always an evolution because now what happens is normally people don&#8217;t want to jump generations like that because you want to make sure one chip works and before the other chip comes up because now if the package doesn&#8217;t work, two chips don&#8217;t work. By the way, I have to tell you this funny story. This is a true story from my personal experience in the semiconductor world. I worked at a company once that had to make two chips like this. I think it&#8217;s a great time to tell this story. And they decided that for whatever reason, because RF is very sensitive stuff, they decided they will make the mirror version of the same chip because the package required two chips. So they decided they&#8217;ll just mirror the chips so that you could create a symmetric package layout to put both the chips on. And they said, fine, this is a beautiful approach. Look at the symmetry and the beauty of this. RF guys love symmetry. And then what happened was the first iteration of the chip, something went wrong. Now the problem is the mirror image of the chip is not the same chip. They&#8217;re two different chips.</p><p><strong>Austin:</strong> Oh, what? Oh yeah, yeah.</p><p><strong>Vik:</strong> So for every iteration, yeah, because the complete the mask is a mirror mask. So even the transistors are mirrored. They are not like everything is not the same. So they had to run the same chip twice every time and pay for all the masks from beginning to end because they decided to use a mirror without determining if the first chip worked.</p><p><strong>Austin:</strong> Oh, oh, no. So they, okay, so the first chip if it&#8217;s broken and then you create a mirror of it in an expensive way. It wasn&#8217;t just a rotation of the first chip. It&#8217;s an actual mirror. So you have all new masks and then, oh crap, the first one didn&#8217;t work. And so now our mirror probably doesn&#8217;t work.</p><p><strong>Vik:</strong> No. It&#8217;s a mirror. So the development cost for this chip was double what it should have been because they mirrored the chip, and the chip didn&#8217;t work once. So every time you made a fix on one, you had to make the mirror chip fix and you had to double everything, mask costs and everything. It&#8217;s like, oh my God. So you know what? Make sure one chip works first before you make two of them. I have seen this in my own little world, and so Google obviously know what they&#8217;re doing. So, obviously the 8i is a good way to do this. But then the question is, why would you put two of them together? I think the advantage is that you double HBM, you double the SRAM. Remember it was 384 MB of SRAM per 8i chip. Now you have double that. You have 760 whatever MB of SRAM. Imagine what amount of stuff you can do with that. So now your bandwidth is enormous bandwidth, and your HBM capacity is so much more. It&#8217;s really good to hook up two chips. Everything in terms of hardware specs kind of doubles almost. And that&#8217;s it has its advantages.</p><p><strong>Austin:</strong> Yeah, let me ask you a question about that then. I guess you&#8217;re making the assumption that the program has access to all the chips that are on the board and therefore to the programmer, suddenly there&#8217;s just twice as much SRAM and twice as much HBM. Is that how you&#8217;re thinking about it? Because if there are individual programs running on each chip, then it&#8217;s like, oh, well, they&#8217;re just in proximity to each other. So clearly these have to be like the system, the kernel can see now double the HBM, double the SRAM.</p><p><strong>Vik:</strong> Yeah, I think that&#8217;s how it is because the way when you package things together in an advanced package or something like this, the idea is that they behave more like one chip working together, like it&#8217;s more like one huge chip.</p><p><strong>Austin:</strong> Yeah. Like one monolithic chip. It just happened to be disaggregated but then reconnected through advanced packaging.</p><p><strong>Vik:</strong> Yeah, when you use advanced packaging, the benefit of that fine grain nature of advanced packaging is that this 8i Max chip is literally like a chip that is two times bigger than the 8i. And you can&#8217;t make two times bigger chips because they&#8217;re already very reticle limited. Remember I was saying how big the 8i chip is to begin with, you can&#8217;t make it. You become reticle limited on the wafer anyway. So Nvidia does this too, with their Rubin packages, they have two Rubin die in a package and they were talking about putting four Rubin die in a package in an advanced package and things like that. So you&#8217;ll burn a lot more power per 8i Max, but you get so much more out of it. And this is good because this is all towards a trend of rack densification. How do you put more TPUs or GPUs into a rack? It&#8217;s very important to get as much compute as possible in a rack because networking to get the same kind of end result is a very difficult thing to do. And what I mean by that is you could just put two chips instead of advanced packaging them together. You can just put two separate chips and hook them up with even optics. Why do people not do that? Because there is no match to putting two chips in an advanced package. Optics can&#8217;t match that kind of latency. It&#8217;s always going to be slower.</p><p><strong>Austin:</strong> And it costs more power.</p><p><strong>Vik:</strong> Yes, exactly. If you&#8217;re moving bits over a longer distance, you&#8217;re going to burn more power. So nobody wants that. So you want to put everything into a single rack and densify the rack as much as possible because another benefit is that maybe you can do it with copper. The reach is limited. It&#8217;s the reach limitation means that you...</p><p><strong>Austin:</strong> Yes, I was going to get there too. I had had the exact same thought, which was essentially, yes, we haven&#8217;t talked about moratoriums and stuff like this, but that&#8217;s where my head went. It was like, well, I guess if it&#8217;s going to be harder, which time will tell, if you&#8217;ve already got space and you can take one rack out and put another rack in and get twice as many, of course, that is a good thing for continuing to scale the AI build out for sure.</p><p><strong>Vik:</strong> Exactly. So it&#8217;s a good thing. And the benefit is that the 8i Max now has so much more it can hold so much more of the model or it can hold so much more of the KV cache. And all that kind of stuff, you have to move data less, so you can save power. You can run long agentic tasks because you have so much storage lying around. You can run fast inferencing because you have double the SRAM. So it gives a lot of benefits in terms of what kind of performance you can extract out of a huge chip like 8i Max. So I think it&#8217;s definitely built for long context agentic workloads.</p><p><strong>Austin:</strong> Yes. So I guess, okay, two thoughts. One, I remember Nvidia was going to go to tying four reticle limited GPUs together, but then there was a question about warpage from CoWoS because it means you needed a bigger substrate. And then it made me think, oh, is this is just going from one to two, but I&#8217;m assuming that the TPU is also built using CoWoS. I don&#8217;t know if it&#8217;s CoWoS-S or CoWoS-L. And so then I&#8217;m thinking, what&#8217;s the TPU going to do in the ninth generation? Do you think that they will still have the one and two variants, like the 9i and 9i Max, or you think they&#8217;ll just go with this Max? Because don&#8217;t we always just want more memory for inference? And then of course, then I also wondered, are they going to scale it even bigger to four? Or is 9i going to have two and then 9i Max actually has four? What do you think?</p><p><strong>Vik:</strong> Yeah, I think they should go the 9i should be whatever the 8i Max is. That&#8217;s the logical way to go. And I don&#8217;t think they&#8217;ll... I mean, the 9i Max might be the four version. Total guess, by the way. I have no idea what the road map is. I have no inside information or any such thing. I&#8217;m just totally guessing here. 9i Max may have to go and solve those warpage issues and then we can put more chips. And I&#8217;m guessing Nvidia will also reach that point where they will put four Rubin die into a... I think Rubin Ultra, right? That was supposed to be for the Rubin Ultra. And so they may reach that point too where we can make this big enough. All these advanced packaging houses are working towards making bigger and bigger packages. We&#8217;re talking about glass interposers and so many things are coming down the line. Those are all going to be solved eventually and so we will go to four-die systems. But yeah, that&#8217;s the way to push the technology forward.</p><p><strong>Austin:</strong> Yes, come on, Intel, EMIB, get in the game here, work with Marvell and whoever, MediaTek, make these TPUs, try to package them. Oh, and I&#8217;ll also say the Max naming reminds me of the iPhone, and it feels like we&#8217;re talking about iPhone generations because it&#8217;s like, oh yeah, this year&#8217;s Max will be next year&#8217;s base model.</p><p><strong>Vik:</strong> I think they&#8217;re losing the naming game. The 8T and 8i was nice. But now it&#8217;s going to get out of hand. Do we have to cover anything else about the 8i Max or do we go to the other crazy name?</p><p><strong>Austin:</strong> I don&#8217;t think so. I think it makes sense. Let&#8217;s go to 8TX. Okay, 8TX. What does that mean, Vik?</p><h2>The 8TX: A Secure System</h2><p><strong>Vik:</strong> Okay, so what is the 8TX? So let&#8217;s get into that. It&#8217;s a weird name and I have no idea what the name means, but Austin, you will suggest a better name at the end, I&#8217;m sure. Because you have a flair for these kinds of things, but I&#8217;ll just explain what it is. So it takes the 8T silicon, which is basically the smaller chip. Remember, this is the one with less SRAM, less HBM. And then throws out the Google Axion host node CPU and instead puts a fifth gen AMD EPYC CPU as the host node CPU. So it uses the 8T, the training version of the silicon, and it uses the AMD chip instead of the Arm-based Axion chip. And funnily, it uses the Boardfly topology. So remember the 8T silicon that we spoke about used 3D Torus because it was a training chip. But this uses the 8T silicon with Boardfly topology and an AMD CPU. You see the confusion here, it&#8217;s a complete rejigging of everything.</p><p><strong>Austin:</strong> Right, right. It&#8217;s like the training chip with the inference network and then with a different CPU. I would not name it. I would not put a T in it. Now, well, I guess, okay, so T is use case to me, it&#8217;s training, right? So what did they say the use cases are for this, by the way?</p><p><strong>Vik:</strong> The use case seems to be what they call confidential inference and we&#8217;ll get to it, but that comes from the use of the AMD EPYC CPU. And they also said this used for private fine tuning of models.</p><p><strong>Austin:</strong> Okay. So I think the first bit was inference and then the second bit is training. But security sounds like the most important thing here. So maybe the best name would be 8S.</p><p><strong>Vik:</strong> Yes, good. Let&#8217;s go with that.</p><p><strong>Austin:</strong> Because then it doesn&#8217;t conflict the use case and the topology and stuff. So maybe, so Google, if you&#8217;re listening, maybe next time chat with us, maybe something like 8S, eight secure chip.</p><p><strong>Vik:</strong> Eight secure is so much nicer.</p><p><strong>Austin:</strong> System, system. It&#8217;s a system.</p><p><strong>Vik:</strong> Yes, it&#8217;s a secure system. See, you just have to use S. And 8TX, I don&#8217;t even know what the X stands for.</p><p><strong>Austin:</strong> Well, I mean, for an RF guy, TX probably means transmission.</p><p><strong>Vik:</strong> That&#8217;s what I thought. Yeah. But it&#8217;s not that. Okay, so confidential... so what is the use case for this? I think that essentially this chip is designed to make sure that you can use these chips to do AI in information sensitive environments. Like a trading firm might be training on their data to make stock purchase decisions. It&#8217;s very important for their business and it involves a lot of money that they&#8217;re training towards making maybe these trading decisions, maybe hundreds of millions, billions, requires security. Nobody can access this. I think of somebody like Jane Street or a big company like that doing something like quantitative trading, high frequency trading, things like that. Or the other use case I can think of is think about Nvidia, for example, or probably a bad example, but anyway, think of any company that has a super secret way to design a chip. This is their secret sauce and they want to now train AI on their secret sauce chip design. And they don&#8217;t want to run it on a server that&#8217;s shared by anyone else. The thing is, if you go to the cloud and you start running inference, your hardware is being shared with other people who are also running their inference on yours. So how do you guarantee that your tokens or your information or whatever is not going to mix in with somebody else&#8217;s or leak in any way? There&#8217;s no way to guarantee that. So I think this chip is based towards that.</p><p><strong>Austin:</strong> Correct. Yes. I think that is actually the key and primary use case, which is Neoclouds or clouds or whatever and having neighbors potentially sharing the same system that you don&#8217;t want to see their information. Because prior to this, if the first and best customer was Google, well, of course, when they&#8217;re training Gemini and all this stuff, wouldn&#8217;t they want it to be secure? Why didn&#8217;t they build this security in hardware security in mind in the first place? Well, yeah, of course they want to be secure, but they were the only customer. It was all just internally used compute. So now when they&#8217;re trying to give their compute to someone else who&#8217;s going to be a landlord and rent it out to other people, I think now is where you want to say, oh, yes, you can put your data on our accelerators and when it moves from the accelerator to HBM or from the accelerator to the CPU or whatever, it will move encrypted data is my understanding.</p><p><strong>Vik:</strong> Correct. So, the main purpose of this is to make sure that the data only reaches the target it is intended to reach. And by that, I mean the data is encrypted. And so there is a key. The key is held only by the correct party who can unlock the data. So, even if this data goes to some other place by whatever means, they can&#8217;t do anything with the data. There&#8217;s no key to unlock it. So, that&#8217;s what this whole confidential computing comes to. So, only the intended recipient can interpret that data and that kind of encryption is very important for trusted, multi-tenant situations.</p><p><strong>Austin:</strong> Yes. Yes. Absolutely. Totally. That makes sense. Okay, so why an AMD CPU then? Why not the Axion?</p><h2>Why AMD? And Why Boardfly?</h2><p><strong>Vik:</strong> So, Axion does not have the feature to basically encrypt this information to the level at least that AMD can do it. I don&#8217;t know if that feature exists or not, maybe it does. But I don&#8217;t think that it&#8217;s as good as what AMD can do with their EPYC systems.</p><p><strong>Austin:</strong> Right, right.</p><p><strong>Vik:</strong> So, the AMD chip has something called a secure encrypted virtualization or SEV. So, essentially, you can have a virtual machine that is completely isolated from other virtual machines and from the underlying operating system that the virtual machine sits on. That underlying operating system is often called a hypervisor. So, the VM is completely insulated from the hypervisor underneath. Hypervisors are shared by many VMs. So, you don&#8217;t want any information leaking from VM through the hypervisor to the other VM. So, this capability allows not just any kind of VMs hosted on these CPUs to be secure and encrypted, but also it allows information that is coming in and out of the CPU or devices that are connected to it, like the TPU in this case, the 8T training TPU. All of them inherit this secure encrypted virtualization feature, which is what allows secure transmission of data between these various parts of the system.</p><p><strong>Austin:</strong> Gotcha. So, the 8TX is using the 8T silicon and nothing has changed. So, all of this is done in software, managed by the CPU, but the TPU must have some sort of software to be able to handle this as well.</p><p><strong>Vik:</strong> I&#8217;m not sure of that software layer within the CPU, but there&#8217;s nothing new in the hardware because the silicon is exactly the same as what came up earlier. So, the CPU orchestrates the confidential computing. That&#8217;s the key thing to take away here. So, all that encryption process, all that is handled by the CPU. And that&#8217;s why they had to go and use this Arm CPU instead. And actually what is funny is that during the announcement, AMD actually put out a press release as well talking about their SEV feature, which I don&#8217;t know, never made it into the news, but if you see it&#8217;s there. They announced that it&#8217;s there. This SEV is why it is inside the Google 8TX.</p><p><strong>Austin:</strong> So, yes, the natural question by the way to that is, okay, is this just a temporary lead? Like AMD got it because they have the SEV and will Arm end up saying, okay, that&#8217;s on our road map now. So that the next Google Axion can have it or is this like a sustainable advantage for AMD?</p><p><strong>Vik:</strong> No, no. So, I think that one of the reasons is that Arm does have client confidential computing. I think they call it client confidential computing architecture or something. If you see their Arm AGI, this phrase is in there. So, Arm already has a CPU that can do this. I don&#8217;t think it has that kind of flight history that AMD has, because AMD has had this feature for a while. So, I think it&#8217;s not the right time. I don&#8217;t think it locks somebody out forever. Other people have it too. Intel has a version of this as well. And I wrote it down here and it is actually called trusted domain extension or TDX. And I was looking up like why did Google actually choose AMD over Intel for this chip. Why don&#8217;t you get an Intel chip? I just saw that there was some security concerns and stuff with the trusted domain extension. So, I figured maybe they just didn&#8217;t want to deal with it and they went to the AMD one because it&#8217;s well established and nice to go.</p><p><strong>Austin:</strong> Or yes, yes. Or maybe they weren&#8217;t sure there would be enough supply from Intel. Hope that&#8217;s not true. But you know, maybe they just said, hey, AMD seems to have the supply stuff figured out and we&#8217;re going to need a bunch of these. Now, yes, to your point, I agree. Clearly, the Axion was made at a specific point in time with certain design decisions that made sense for the Google Cloud. And so in this case, when they&#8217;re like, hey, we need this security stuff, why not just pick one up off the shelf from a merchant silicon vendor who already has a portfolio of CPUs. So, obviously, AMD already had silicon that can do this, so does Intel. Maybe in the future, we&#8217;ll see a portfolio of Axion, you know, this is our head node one for the secure environment and stuff like that. But point in time, easiest to just pick something off the shelf, plop it in.</p><p><strong>Vik:</strong> Yeah, what is so funny about this announcement is that at this time, I got an X invitation. Somebody followed me. I&#8217;m like, oh, who is this person who followed me? Oh, it&#8217;s Arm. Oh wait, Arm is following me. That&#8217;s cool. And then I&#8217;m like, oh wait, Arm DM&#8217;d me. Like what did they say? They&#8217;re like, oh wait, Arm DM&#8217;d me to say that the 8i Max actually uses the Arm-based Google Axion. Because it wasn&#8217;t clear from the press release what that was using. The 8TX uses the AMD. Yes, we know. But what was the 8i Max using? So, they did it to a bunch of people on X.</p><p><strong>Austin:</strong> Okay, so they were just proactively telling people on X. Had you been... okay, okay. I was like, did you talk about it and then they DM&#8217;d you, but no, they&#8217;re just proactively.</p><p><strong>Vik:</strong> They&#8217;re proactively telling people on X.</p><p><strong>Austin:</strong> Good job, Arm&#8217;s comms social media team for proactively saying like, just in case it wasn&#8217;t clear, we&#8217;re also in this slot here.</p><p><strong>Vik:</strong> Yes, please send it out on your channels just so that we know. So, I mean, I&#8217;m doing my part. Semi Doped is where I talk about this stuff. And yeah, so 8i Max uses the Google Axion, like everything else. And the same reason is simple. There&#8217;s no encryption-based confidentiality computing here. It&#8217;s only a 8TX feature. Can we just call it 8S? So, that&#8217;s the whole point. So, that&#8217;s where we are right now with these chips. I think it&#8217;s a good addition. I think it unlocks... the 8i Max is honestly a bit more exciting to me because of the sheer size of that chip.</p><p><strong>Austin:</strong> Yes, but you can fine tune on it because it is just an accelerator at the end of the day and like a GPU, you can run several workloads. And I think the interesting point too is that really one way to think about the 3D Torus is it&#8217;s also very good for large scale training when you&#8217;re like pod level training, you know, like I need to talk to the chips that are closest to me on this board and then the chips that are closest to me in this rack, but then I also need to be able to talk to maybe other pods, using RDMA. And of course, yes, Boardfly is more about minimizing the hops because you don&#8217;t know who you&#8217;re going to talk to yet, you don&#8217;t know who you&#8217;re going to talk to next, but if you&#8217;re doing small fine tuning, you&#8217;ve got an open like a some medium-sized open model and you&#8217;re just taking your own internal data that you want to be secure, and you are doing some training. It will still work with Boardfly. It&#8217;s not like it won&#8217;t work without the 3D Torus. It might not be the most efficient thing in the world, but at the end of the day, you&#8217;re not the kind of person that&#8217;s like, oh, I&#8217;m putting 100,000 TPUs together and if this takes me an extra day, it costs an extra couple million bucks. No, you&#8217;re just fine tuning small stuff that probably lives all within one rack or maybe even within one server node or pizza box, so to speak. So it can it&#8217;s inference first, but it can be used for fine tuning.</p><h2>TPU Sync: Google&#8217;s Software Layer</h2><p><strong>Vik:</strong> Yeah, yeah. Part of this slew of blog posts that came on the Google Cloud blog, I mean, if anybody wants to look it up, it&#8217;s very interesting. They had so many other blog posts that was relatively lost in the noise. One of them was called the Google Kubernetes engine. It&#8217;s called the GKE and it does exactly what you&#8217;re talking about. It&#8217;s called dynamic slicing. And the whole point is very nice actually. And there are really good animations. We&#8217;ll leave links in the show notes. But essentially, you can connect these groups of TPUs together to form a bigger pod. You can take many smaller pods of GPUs and you can hook up to them to become what they call a super slice. And then you hook up all the external connections via optical circuit switches, OCS. So you can connect smaller pods into huge big pods with OCS and that&#8217;s called super slicing. But you can do the reverse too. You can go into a single pod and just carve out a few TPUs. So basically, you have complete flexibility, you just throw a bunch of TPUs into a data center. You can slice them any way you want. You want to put them together to make a bigger thing, you can do it. You want to take a single one and break into it to sub slicing, you can do it. It&#8217;s fascinating the kind of things that you can do with these things.</p><p>But yeah, all that is really detailed. The blog posts as usual from Google have beautiful animations and explanations. They&#8217;re not even very long, so I highly recommend people read the links in the description. But I think one other point we should touch on now that you mentioned that other people like Oracle can go and use Google and they have some announced some plans around this too, is that what makes this happen? How does Oracle use Google TPUs without... CUDA was the biggest thing for Nvidia, like everybody could program on CUDA and use GPUs. I think Google is making it a lot more easier now to do those kinds of things with TPUs. And I&#8217;m no software expert, so I can&#8217;t go into all the details, but I think one thing stood out to me nicely that straddles the world between software and hardware and that&#8217;s called TPU sync. So TPU sync is now an open source project that essentially, I feel like it&#8217;s Google&#8217;s version of Nvidia&#8217;s Dynamo. Maybe I&#8217;m wrong, but somebody can correct us. But basically what it helps you do is that it helps you manage the memory hierarchy. It allows you to handle HBM, DRAM, SSD storage, and even network storage. I think Google has a special version of network storage that they use. So all of that, this software layer, which is open source, allows the inference engine to really distribute work through the right places at the right times. It doesn&#8217;t recompute anything. There is always a place in the memory hierarchy to store something. And so that&#8217;s what TPU sync is. And we&#8217;ll get into a little bit more detail, but what do you think of this software layer?</p><p><strong>Austin:</strong> Yeah, yeah. No, I think it&#8217;s very important. Obviously, if you&#8217;re trying to sell your Silicon, as AMD knows, if you want to sell accelerators, but people aren&#8217;t happy with your software, you&#8217;re not going to sell accelerators. That was the ROCm thing that they had to deal with like two years ago. They&#8217;re pretty much past it now. No one talks about ROCm anymore because it just works. So clearly, if you want to use TPUs, you want to sell them to external customers, the software has to be very easy to use. And to your point, a big part of LLM workloads at scale on accelerators today is disaggregation and then it&#8217;s also memory management. And actually, I think this probably fits in nicely too with having the 8i Max is like, oh, some of the chips have a ton of HBM and a ton of SRAM and how can you make sure that you put the hot KV cache stuff where you need it and offload some of the colder stuff to still have it something that you can access, but not as close. And it actually it reminds me of the conversation you had with Val Berkovici of Weka. You guys talked a lot about this like LLMs needing to think very carefully about having as much essentially having a memory hierarchy and can you also do tricks in software to make it feel like some of this stuff like flash or something is lower latency than you might expect because you&#8217;re thoughtful about what you put where and how you fetch it in time.</p><p><strong>Vik:</strong> Yeah, exactly. It feels like TPU sync helps with that. It helps with that. And the blog post I was reading says that the amount of the KV cache transfer overhead drops by like 90%. There are some time to first token improvements that span anywhere from 10% improvement all the way to 90% improvement depending on various configurations and model sizes and things like that. And also the time between inter token latency, it&#8217;s called ITL. It&#8217;s just the time that is taken between tokens, that comes down as well with the use of TPU sync. And there&#8217;s a nice analogy in the article, which is basically like think about a very fast executive chef at a restaurant. Now, the guy is cutting all the vegetables and cutting all the meats and cooking and but he has a very small countertop. So he can&#8217;t literally keep all the meat and vegetables and the food or whatever that he&#8217;s cut. So what he has to do is he has to cook it and then when the next batch comes, he has to cut everything from scratch and store it in the small countertop. He needs to keep doing this. It&#8217;s very slow. So the TPU sync idea is that now you don&#8217;t have this small cutting board where he is operating. Now you roll up a cart next to him where he can store all the food that he&#8217;s cut. And if there&#8217;s meats, then you have a cold storage there where you can store cut meats even if you wanted to. So everything is in its right place when you need it. So the chef only needs to get stuff from the right place. Sometimes, yes, it takes time to go to cold storage to get the meat, but it&#8217;s far better than cutting the meat from scratch.</p><p><strong>Austin:</strong> I love the analogy. And it just goes to show, when you&#8217;re operating these workloads at such big scale, it&#8217;s very important to have the right stuff at the right place at the right time. And co-design and co-optimization from hardware into software can make this possible. So, what I just thought when you were explaining the analogy even, it just made me think like, oh, what a cool job for some people out there, hardware and software engineers working on this to think about like, oh, how do we make sure that the KV cache is in the right place at the right time and how can we write software to make that even easier? Because obviously, it feels very technical and kind of like a niche problem, but even like inter-token latency, that kind of stuff could impact millions or billions of users, like they they would legitimately feel that. Of course, that&#8217;s a whole another conversation of if Google can get people to use Gemini and, where&#8217;s Google&#8217;s agent and stuff like that. We won&#8217;t go there. But very cool, very cool technical problems to work on.</p><h2>The Business of Selling TPUs</h2><p><strong>Austin:</strong> Today&#8217;s sponsor is Meter. You&#8217;ve got the GPUs and the power. Now, who builds the network and who do you call when it breaks? With Meter, it&#8217;s one call. Tell them your GPU or server count and they design, deploy, and run your front-end storage and management networks. Meter builds its own networking hardware, switches and all, and writes its own network OS. So, that call goes to the team who built it. Running a Neocloud or clusters in a colo, go to semidoped.com/meter. Links in the show notes.</p><p><strong>Vik:</strong> Yeah, I&#8217;ll just explain maybe two use cases that are very interesting to me. One is, think about enterprise use cases. Before a prompt is sent by any employee, companies say they want to put a lot of guard rails and rules and everything into place. So, what they usually do is they prepend the entire prompt that you send with all the guard rails first. So, it&#8217;s like you send an instruction manual of what you can and cannot do before Austin can send a request.</p><p><strong>Austin:</strong> Yes. It&#8217;s like attached to every email that I send.</p><p><strong>Vik:</strong> Yes, yes, that&#8217;s attached to every email that you send. So, this kind of stuff can be stored with TPU sync in the right locations. So that you don&#8217;t have to recompute or redo this. You have the instructions already always and you can always process requests quickly. That&#8217;s one use case.</p><p>The other use case is that I think one way to drop inference cost is to make sure that KV cache is not recomputed. You want to have KV cache hits as much as possible. You want to use tokens that are already existing from the past or things like that. So, if you have all this storage hierarchy, you will have a much higher KV cache hit rate, which means you will drop your inference costs by an order of magnitude in this environment. So, it&#8217;s a nice, nice thing, I think. They also have a nice thing called host-to-host streaming, which means if you&#8217;re doing prefill on a separate rack, and you&#8217;re doing decode on a separate rack, they have these high-speed interconnect, it&#8217;s not interconnect, it&#8217;s like a protocol that they call host-to-host streaming, H2H, that makes sure that the entire weights from the prefill goes to decode quickly. So this kind of stuff, they&#8217;ve thought through every part of the memory hierarchy nicely. But I feel like this is not unique to Google. I feel like Google, Nvidia has thought about a lot of this as well. So, but it&#8217;s good to see the progress, right?</p><p><strong>Austin:</strong> Totally. Totally. Agree, agree, agree. Yeah, and I&#8217;m sure AMD, for example, thinks about this too, but good to see Google doing it and to know, if you&#8217;re asking yourself, can they compete in selling these to people? Clearly this shows that yes, they can build this, people can be confident that they can run on top of this. And obviously, I would like to think that Google&#8217;s using this all internally too, so, you know, dogfooding it, so to speak. Which would be a different, and that actually could be an unfair advantage that they have compared to a merchant silicon vendor like an AMD or Nvidia where AMD and Nvidia don&#8217;t necessarily, yes, they do their own training, this could be a reason, a good reason why AMD bought World Labs, pushing into the having your own internal team to really understand the 3D world model type use case. But Nvidia obviously with Nemotron, they train their own model and it&#8217;s not because they want to undercut their buyers and say, yes, we&#8217;re going to try to beat Anthropic and beat OpenAI. No, no, no. They train their own model so that they can make these intelligent decisions about how to architect their system. But just saying, if Google&#8217;s not so big and hierarchical that they don&#8217;t talk, but hopefully their the Gemini team, the DeepMind team, etcetera, is talking to the TPU team a lot and so that they&#8217;re using this software and they&#8217;re talking and sharing all that knowledge as they&#8217;re building the systems that they&#8217;re consuming this all right away too.</p><p>Yeah, it&#8217;s exciting when people build exciting hardware, we are excited to talk about it. So, if you&#8217;re listening to this podcast and work in one of these companies doing all this cool stuff, you know, keep at it.</p><p><strong>Vik:</strong> Indeed.</p><p><strong>Austin:</strong> Indeed. Yeah, let us come by. Show it to us sometime. Have Semi Doped, have a Semi Doped tour sometime.</p><p><strong>Vik:</strong> Absolutely.</p><p><strong>Austin:</strong> Cool. All right. Well, Vik, anything else on the TPU front?</p><p><strong>Vik:</strong> No, that&#8217;s it for now. We&#8217;ll leave all the blog articles on the show notes, but yeah, you can also go just find them. They&#8217;re really nice. I love reading Google Cloud blog articles because they have like the best pictures and the neatest explanations and they&#8217;re kind of short. So, it&#8217;s very much fun to read it. I always have fun.</p><p><strong>Austin:</strong> Awesome. Yes, I love those too. Good job, whoever makes them. All right, with that, everyone, thank you for listening. Send this to your friends. We appreciate recommendations and like, comment, subscribe, whatever. Vik and I read it all, so please, you know, send us your thoughts.</p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 17th, 2026]]></title><description><![CDATA[Washington shuts the remote-access door, Nvidia&#8217;s board builds a compliance wall, forecasters trim the AI boom, OpenAI books 457 MW of AMD, and Korea&#8217;s memory makers keep spending]]></description><link>https://daily.semidoped.com/p/daily-update-november-17th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-17th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Tue, 17 Nov 2026 12:38:08 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/9865ec52-a3be-4897-98dd-4c181ec57cba_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Tuesday, November 17th, 2026</strong></p><p>Washington just made renting a GPU as hard as buying one. Nvidia&#8217;s board answered with a compliance committee, not its founder, hours before earnings. SemiAnalysis and Sam Altman both took a red pen to the AI demand math, and OpenAI booked 457 MW of AMD anyway.</p><blockquote><p><em><strong>Austin and Vik are both at TomorrowX this week. If you see them, please say hello!</strong></em></p></blockquote><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>Jensen Huang Resigns</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=dWrrSg4XlZA">YouTube</a>, <a href="https://podcasts.apple.com/us/podcast/jensen-huang-resigns/id1866707196?i=1000801597305">Apple Podcasts</a>, or <a href="https://open.spotify.com/episode/5egxsEWge1dO3VOEvFwMLX">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-jensen-huang-resigns">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p><mark data-color="#d9d2e9" style="background-color: rgb(217, 210, 233); color: rgb(0, 0, 0);">Fine-tune and serve open models without managing a single GPU.</mark></p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p><strong>Accounts are free, and get $5 in credit when you sign up.</strong></p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning &#8594;</a></p></div><h3>Washington closes the remote-access loophole</h3><p>The Commerce Department&#8217;s BIS released an interim final rule on Monday that treats renting AI compute like shipping it. Any data center, anywhere, now needs a license for advanced U.S. chips if it gives more than 12 million &#8220;aggregate TPP&#8221; of compute to companies headquartered in China, Macau or other D:5 countries. Data centers already serving those customers get 90 days to get licensed.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Advanced chips headed to Singapore, Malaysia, Indonesia, Thailand and Brazil now need a license too, unless the buyer is a U.S.-headquartered company that reports to Commerce.</mark> The rule also narrows the paths foundries and packagers can use to make advanced chips for designers in restricted countries, and formally scraps the Biden-era AI Diffusion Rule. It takes effect when it is published in the Federal Register, scheduled for November 19.</p><p>Markets read it as a sales hit. Nvidia fell about 4% on Monday. Samsung and TSMC slipped in Asia on Tuesday, with CLSA flagging extra customer checks for Samsung Foundry, and Southeast Asian governments and data center operators are scrambling to assess the damage.</p><p><a href="https://www.bis.gov/press-release/bis-issues-rule-addressing-remote-access-american-super-intelligence-technology">Source: BIS</a> &#183; <a href="https://www.bis.gov/media/documents/rin-0694-ak50.pdf">Interim final rule (BIS)</a> &#183; <a href="https://www.nixonpeabody.com/insights/alerts/2026/11/16/bis-targets-remote-access-to-advanced-chips-through-overseas-data-centers">Nixon Peabody</a> &#183; <a href="https://www.cnbc.com/2026/11/16/nvidia-commerce-china-lutnick.html">CNBC</a> &#183; <a href="https://www.benzinga.com/markets/equities/26/11/63320949/tech-stocks-960-billion-china-ai-chips-remote-access-rule">Benzinga</a> &#183; <a href="https://www.investing.com/news/stock-market-news/samsung-elec-tsmc-shares-fall-after-us-announces-curbs-on-chinas-chip-access-5031711">Reuters via Investing.com</a> &#183; <a href="https://asia.nikkei.com/business/technology/singapore-malaysia-grapple-with-us-chip-curbs-targeting-southeast-asia">Nikkei Asia</a></p><h3>Nvidia&#8217;s board builds a compliance wall</h3><p>Nvidia&#8217;s board created a standing National Security and Export Control Oversight Committee on Tuesday, chaired by director Suzanne Nora Johnson. Nvidia&#8217;s national security and trade compliance teams will report to it, as well as to interim CEO Colette Kress, and the committee will bring in outside legal and trade experts.</p><p>Washington filled in its side of the Megaspeed story. In a joint statement on Monday, Commerce and the Justice Department said Commerce opened its probe in 2024 and asked Nvidia for Megaspeed sales records in May 2025. Nvidia put Megaspeed on its &#8220;no sell&#8221; list the next month.</p><p>Inside Nvidia, the pressure runs the other way. The Wall Street Journal reports that Nader Khalil, a director of developer technology, wrote to the board and Huang&#8217;s roughly 60 direct reports warning that &#8220;irreplaceable talent&#8221; is thinking about leaving. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The board is not considering bringing Huang back, the Journal reports, so its answer to Washington is more oversight rather than its founder.</mark> Results land after Tuesday&#8217;s U.S. close, early Wednesday in India.</p><p><a href="https://nvidianews.nvidia.com/news/nvidia-board-of-directors-establishes-national-security-and-export-control-oversight-committee">Source: Nvidia</a> &#183; <a href="https://www.businesstimes.com.sg/international/us-agents-singapore-officials-conducted-megaspeed-check-june-2025-alice-huang-interviewed-us">Reuters via The Business Times</a> &#183; <a href="https://www.hindustantimes.com/world-news/nvidia-board-warned-of-irreplaceable-talent-loss-in-employee-letter-101794885991870.html">Wall Street Journal via Hindustan Times</a> &#183; <a href="https://www.yna.co.kr/view/AKR20261117130000009">Yonhap</a></p><h3>The forecasters blink</h3><p>SemiAnalysis cut its AI accelerator spending forecast to $758 billion for 2028, down 28%, and $923 billion for 2030, down 40%. It points to slower growth in new revenue at frontier labs and tighter financing for projects without a hyperscaler behind them. Nvidia takes most of the cut, partly because it sells more to smaller neoclouds.</p><p>Sam Altman explained how the demand got inflated. At a Semafor forum on Monday, he said several new chip companies are booking capacity at the same foundry, and each one assumes it will win most of the orders from OpenAI and Anthropic. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Altman&#8217;s point is that suppliers are double counting the same few buyers, and OpenAI neither can nor needs to buy from all of them.</mark></p><p>He still called data center demand solid and said long-term compute demand is intact. The trouble is that a forecast built on the same handful of customers can shrink fast.</p><p><a href="https://www.tokenpost.com/news/technology/59159">Source: TokenPost</a> &#183; <a href="https://en.sedaily.com/world/2026/11/17/altman-warns-ai-chip-demand-is-being-double-counted">Seoul Economic Daily</a> &#183; <a href="https://en.infomaxai.com/news/articleView.html?idxno=148832">Yonhap Infomax</a></p><h3>OpenAI books 457 MW of AMD through TensorWave</h3><p>TensorWave signed a six-year deal on Monday to supply OpenAI with 457 MW of IT capacity across several U.S. sites, running AMD MI455X GPUs in Helios racks. The capacity is dedicated to OpenAI alone, starts coming online in the first half of 2027 and should be fully deployed by early 2028.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">TensorWave is the first neocloud publicly named as a supplier under OpenAI&#8217;s AMD partnership.</mark> AMD Ventures is an investor in TensorWave. Neither side disclosed the value of the deal or named the sites.</p><p><a href="https://www.businesswire.com/news/home/20261116130015/en/TensorWave-Announces-Agreement-with-OpenAI-to-Deliver-More-Than-450MW-of-AMD-Powered-AI-Infrastructure">Source: TensorWave via Business Wire</a> &#183; <a href="https://www.datacenterdynamics.com/en/news/openai-signs-on-for-457mw-ai-capacity-with-tensorwave/">Data Center Dynamics</a> &#183; <a href="https://thenextweb.com/news/tensorwave-openai-457mw-amd-helios">The Next Web</a></p><h3>Korea&#8217;s memory makers keep spending</h3><p>Samsung and SK hynix spent a record combined 88.16 trillion won ($65.7 billion) on capex in the first nine months of the year, according to filings released Monday. Samsung&#8217;s spending rose 76.6%, almost all of it on chips. SK hynix has already spent more than it did in all of 2025.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The cash is piling up faster than they can spend it: the pair held 329.2 trillion won in cash and liquid assets at the end of September, up nearly 193 trillion won in a year.</mark> SK hynix alone has net cash of 98.66 trillion won.</p><p>That is the supply Altman warns may be planned against the same few buyers.</p><p><a href="https://www.koreaherald.com/article/10939605">Source: The Korea Herald</a> &#183; <a href="https://www.asiae.co.kr/en/article/2026111619495885604">The Asia Business Daily</a></p><div class="callout-block" data-callout="true"><h4>Supported by Meter</h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter-daily">See how Meter builds data center networks</a></p></div><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>Samsung and Apple</strong> &#8212; Samsung and Apple are in final talks on a four-to-five-year DRAM supply deal starting in the first half of 2027, with guaranteed minimum volumes and market-based pricing (<a href="https://www.mk.co.kr/news/business/12211342">Maeil Business Newspaper</a> &#183; <a href="https://www.macrumors.com/2026/11/17/apple-rare-long-term-memory-deal-samsung/">MacRumors</a>).</p></li><li><p><strong>Foundry pricing</strong> &#8212; Foundries including TSMC have sent Taiwanese chip designers notices of 2027 wafer price increases ranging from 10% to more than 20% (<a href="https://www.trendforce.com/news/2026/11/17/news-foundry-price-hikes-of-10-to-over-20-reportedly-loom-in-2027-asic-firms-face-margin-pressure/">Commercial Times via TrendForce</a>).</p></li><li><p><strong>Chips in China</strong> &#8212; SemiAnalysis says its accelerator model tracks more than 1.5 GW of export-restricted Nvidia GPUs that ended up in China (<a href="https://newsletter.semianalysis.com/p/nvidia-and-the-world-without-jensen">SemiAnalysis</a> &#183; <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/mass-smuggling-of-nvidia-ai-gpus-could-be-the-underlying-cause-of-jensen-huangs-departure-analysts-claim-more-than-1-5-gigawatts-of-banned-chips-tracked-to-china-despite-export-controls">Tom&#8217;s Hardware</a>).</p></li><li><p><strong>DeepSeek</strong> &#8212; DeepSeek is capping free users&#8217; daily usage and input length because of &#8220;insufficient computing power,&#8221; per user notices (<a href="https://www.yicaiglobal.com/news/deepseek-sets-new-limits-for-free-users-due-to-insufficient-computing-power">Yicai Global</a>).</p></li></ul><h2>Sector Watch</h2><p><strong>Memory</strong></p><ul><li><p><strong>TrendForce</strong> &#8212; The tracker raised its 2027 server DRAM demand forecast and trimmed mobile and PC demand as high memory prices bite. (<a href="https://www.hankyung.com/article/202611172043i">Source: Korea Economic Daily</a>)</p></li><li><p><strong>NAND makers</strong> &#8212; NAND spot prices fell about 20% in two weeks on weak consumer demand and rising YMTC supply. (<a href="https://www.trendforce.com/news/2026/11/17/news-nand-spot-prices-reverse-course-sliding-20-in-2-weeks-amid-weak-consumer-demand-ymtc-competition/">Source: TrendForce</a> &#183; <a href="https://en.sedaily.com/business/2026/11/17/why-nand-spot-prices-fell-20-percent-in-a-month-amid-ai-boom">Seoul Economic Daily</a>)</p></li></ul><p><strong>Compute and systems</strong></p><ul><li><p><strong>Nvidia</strong> &#8212; At SC26, Nvidia unveiled Rubin HPC, a simulation GPU due in 2028, as Vera Rubin debuted atop MLPerf Training. (<a href="https://blogs.nvidia.com/blog/rubin-hpc/">Source: Nvidia</a> &#183; <a href="https://blogs.nvidia.com/blog/mlperf-training-endpoints-vera-rubin/">Nvidia MLPerf blog</a>)</p></li><li><p><strong>CoreWeave</strong> &#8212; The neocloud was the only cloud to submit multi-rack Vera Rubin MLPerf results, training DeepSeek-V3 in 17.65 minutes on 256 GPUs. (<a href="https://wf.coreweave.com/blog/coreweave-becomes-the-first-cloud-provider-to-deliver-multi-rack-nvidia-vera-rubin-results-in-mlperf-r-training-v6-1">Source: CoreWeave</a>)</p></li><li><p><strong>AMD</strong> &#8212; AMD closed its roughly $8.2 billion purchase of World Labs and named founder Fei-Fei Li its chief scientist. (<a href="https://ir.amd.com/news-events/press-releases/detail/1306/amd-completes-acquisition-of-world-labs">Source: AMD</a> &#183; <a href="https://www.thelec.net/news/articleView.html?idxno=15689">The Elec</a>)</p></li></ul><p><strong>Networking and memory pooling</strong></p><ul><li><p><strong>Lambda</strong> &#8212; The neocloud linked over 10,000 Blackwell Ultra GPUs with Nvidia co-packaged optics switches, calling it the first large-scale CPO deployment. (<a href="https://lambda.com/blog/worlds-first-large-scale-deployment-of-nvidia-photonics-switches-with-cpo">Source: Lambda</a>)</p></li><li><p><strong>Marvell</strong> &#8212; Marvell and Penguin Solutions will pair CXL memory expansion with Photonic Fabric optical memory pooling, with photonic samples due in late 2027. (<a href="https://investor.marvell.com/news-events/press-releases/detail/1044/marvell-and-penguin-solutions-collaborate-to-address-data-center-ai-memory-and-accelerate-inference-at-scale">Source: Marvell</a> &#183; <a href="https://www.thelec.net/news/articleView.html?idxno=15690">The Elec</a>)</p></li></ul><p><strong>Data centers</strong></p><ul><li><p><strong>Amazon</strong> &#8212; AWS is selling 75% of 13 data center shells to Blackstone&#8217;s REIT in a $5 billion deal and leasing them back. (<a href="https://www.blackstone.com/news/press/blackstone-digital-infrastructure-trust-announces-5-billion-acquisition-of-stabilized-data-center-portfolio/">Source: Blackstone</a> &#183; <a href="https://www.datacenterdynamics.com/en/news/blackstone-digital-infrastructure-trust-acquires-majority-stake-in-aws-joint-venture-for-5bn/">Data Center Dynamics</a>)</p></li><li><p><strong>Core Scientific</strong> &#8212; The miner-turned-AI-host filed plans for a $1.3 billion, 285MW Texas building on the campus where AMD leases 110MW. (<a href="https://www.datacenterdynamics.com/en/news/core-scientific-files-for-285mw-data-center-outside-dallas-texas/">Source: Data Center Dynamics</a> &#183; <a href="https://mlq.ai/news/core-scientific-files-for-a-13-billion-data-center-building-near-dallas-at-the-campus-where-amd-plans-to-lease-110-megawatts/">MLQ.ai</a>)</p></li></ul><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark></p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/intel-custom-asics">Intel Custom ASICs?</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/why-ai-data-center-constraints-make">Why AI Data Center Constraints Make Efficiency More Valuable</a></p></li></ul><p><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better!</mark> <a href="http://semidoped.com/poll"><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Give us your feedback</mark></a></p><p>Want to become a sponsor? Send an email to <a href="mailto:sponsor@semidoped.com">sponsor@semidoped.com</a></p><p></p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 16th, 2026]]></title><description><![CDATA[Huang eyes a comeback, Nvidia shuts its checkbook, TSMC farms out CoWoS, GlobalFoundries breaks ground, and Washington picks AI champions]]></description><link>https://daily.semidoped.com/p/daily-update-november-16th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-16th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Mon, 16 Nov 2026 13:09:50 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/d1413674-5502-461f-b56c-cc3ff75ccd86_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Monday, November 16th, 2026</strong></p><p>Jensen Huang&#8217;s allies want him back at Nvidia, the Treasury Secretary wants an outsider, and Taiwan&#8217;s stock market just wanted out, falling 6% on Monday. The Wall Street Journal says Nvidia was already closing its checkbook to the neoclouds before Huang left. The buildout itself kept moving, with TSMC handing CoWoS work to Xintec and GlobalFoundries breaking ground on a Japan-funded fab in New York.</p><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>Jensen Huang Resigns</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=dWrrSg4XlZA">YouTube</a>, <a href="https://podcasts.apple.com/us/podcast/jensen-huang-resigns/id1866707196?i=1000801597305">Apple Podcasts</a>, or <a href="https://open.spotify.com/episode/5egxsEWge1dO3VOEvFwMLX">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-jensen-huang-resigns">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p>Fine-tune and serve open models without managing a single GPU.</p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p>Accounts are free, and get $5 in credit when you sign up.</p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning &#8594;</a></p></div><h3>Huang wants back in, Washington wants an outsider</h3><p>Jensen Huang is considering a return as Nvidia CEO and has discussed it with confidants, the Financial Times reported Sunday. One ally is assembling a &#8220;war room&#8221; of Silicon Valley executives to lobby the board. Elon Musk has signed a public letter to &#8220;bring Jensen back,&#8221; and shareholders Altimeter and Atreides asked the board on Friday to keep &#8220;all options on the table.&#8221; Huang does not hold a controlling stake and cannot overrule the board.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Treasury Secretary Scott Bessent said Nvidia needs a &#8220;dramatic shakeup&#8221; with a CEO from outside the company, and wants answers on export compliance within 10 days.</mark> He asked whether Nvidia pressed &#8220;seemingly illicit sales&#8221; to lift revenue, pointing to its Singapore billings. House Republicans have asked Huang to sit for a closed-door interview on December 1, and the Senate Banking Committee wants him to testify publicly on December 3.</p><p>Investors did not wait. Taiwan&#8217;s Taiex fell 6.03% on Monday, its second-largest point drop ever, with TSMC down 5.74%. South Korea&#8217;s Kospi lost 3.59%, and SK hynix fell 7.89%. Nvidia reports on Tuesday, with analysts expecting about $109 billion in revenue on Colette Kress&#8217;s first call as interim CEO.</p><p><a href="https://www.investing.com/news/stock-market-news/jensen-huang-discusses-comeback-as-nvidia-ceo-with-allies-ft-reports-5028065">Source: Reuters via Investing.com (FT report)</a> &#183; <a href="https://www.cnbc.com/2026/11/15/bessent-nvidia-ceo-lutnick-megaspeed.html">CNBC</a> &#183; <a href="https://abcnews.com/Politics/house-senate-republicans-call-howard-lutnick-jensen-huang/story?id=138111390">ABC News</a> &#183; <a href="https://focustaiwan.tw/business/202611160014">Focus Taiwan</a> &#183; <a href="https://www.yna.co.kr/view/AKR20261116113751008">Yonhap</a></p><h3>Nvidia closes the open checkbook?</h3><p>Interim CEO Colette Kress is expected to rein in Nvidia&#8217;s financing of the AI industry, and Huang had already started pulling back before he left, the Wall Street Journal reported Sunday. At GTC Berlin in October, he told neocloud CEOs to secure upfront payments from their own customers. &#8220;I could just give all this capacity to Elon!&#8221; he said, according to people who heard him.</p><p>Nvidia joined a record 129 venture deals over the past year, versus 49 in 2023, the Journal says. Its last quarterly filing listed $29 billion of cloud service agreements and $279 billion of supply and capacity commitments. UBS&#8217;s Tim Arcuri estimates a pullback could take $50 billion or more out of the market.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The neocloud model now depends on customers paying before the GPUs arrive, not on Nvidia standing behind the deal.</mark> Lenders are already pickier: Blackstone&#8217;s QTS cut a planned private debt sale by about 60% to roughly $300 million after investors pushed back on pricing.</p><p><a href="https://www.wsj.com/tech/ai/nvidias-days-as-an-open-checkbook-for-the-ai-economy-are-numbered-90c7b388">Source: Wall Street Journal</a> &#183; <a href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm">Nvidia 10-Q (SEC)</a> &#183; <a href="https://thenextweb.com/news/nvidia-financing-kress-neoclouds-upfront-payments">The Next Web</a> &#183; <a href="https://www.businesstimes.com.sg/property/blackstones-qts-cutting-ai-debt-sale-more-half">Bloomberg via The Business Times</a></p><h3>Taiwan&#8217;s packaging buildout spreads beyond TSMC</h3><p>TSMC signed a seven-year deal with Xintec, its 41%-owned affiliate, to build and run the &#8220;on-substrate&#8221; step of CoWoS inside TSMC&#8217;s plant in the Central Taiwan Science Park. The agreement runs through 2033, and the first phase reaches volume production in 2029. Xintec approved NT$15 billion (about $456 million) of convertible bonds to fund the equipment.</p><p>The same day, ASE opened the first phase of its Lujhu plant in Kaohsiung and broke ground on the second. The two phases total about NT$100 billion of investment and should bring in about NT$80 billion of revenue a year at full capacity. The first phase is a former AUO flat-panel plant that ASE bought for NT$6.5 billion.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">TSMC is pulling more partners into CoWoS, and this one will run its line inside a TSMC fab.</mark> ASE has already raised its 2026 capex budget three times, to $11.5 billion, because it expects demand to outrun capacity next year.</p><p><a href="https://www.aseglobal.com/press-room-ch/lujhu-plant-phase-1-inauguration-and-phase-2-groundbreaking/">Source: ASE</a> &#183; <a href="https://www.taiwannews.com.tw/news/6476137">Taiwan News</a> &#183; <a href="https://www.taipeitimes.com/News/biz/archives/2026/11/14/2003867696">Taipei Times</a></p><h3>GlobalFoundries breaks ground on a Japan-funded fab</h3><p>GlobalFoundries started construction on a new 300mm fab in Malta, New York, the first of two planned under a U.S.-Japan initiative worth up to $21 billion. The first fab costs about $10 billion, with $570 million in New York State incentives, and should come online in 2029. Each fab adds about 900,000 wafers a year of capacity for automotive, industrial and infrastructure chips.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Japan supplies the money and GlobalFoundries supplies the technology and operations, a new way to pay for U.S. fabs.</mark> The two fabs should create more than 1,000 manufacturing jobs and about 27,000 construction and ecosystem jobs.</p><p><a href="https://gf.com/news-and-events/news/globalfoundries-breaks-ground-on-first-new-semiconductor-fab-under-landmark-us-japan-initiative/">Source: GlobalFoundries</a> &#183; <a href="https://www.esd.ny.gov/esd-media-center/press-releases/governor-hochul-announces-groundbreaking-globalfoundries-new-10-billion-semiconductor">Empire State Development</a></p><h3>Washington signs up allies and picks champions</h3><p>The U.S. and nine partners signed the San Francisco Declaration on Sunday, pledging to protect the technologies and infrastructure behind AI supply chains. The signers are Australia, India, Israel, Japan, South Korea, Qatar, Singapore, the UAE and the UK. Working groups will cover supply chains, security hubs and resilience, and the Netherlands, Germany, Sweden and the European Commission are set to sign next week, per the FT.</p><p>Separately, the Wall Street Journal reports the administration will name more than two dozen AI &#8220;national champions,&#8221; including Nvidia, Amazon, Microsoft, OpenAI and Anthropic, to help them win business abroad against Chinese rivals. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Nvidia makes Washington&#8217;s champion list in the same week Treasury is questioning its export compliance.</mark></p><p><a href="https://www.state.gov/releases/office-of-the-spokesman/2026/11/the-san-francisco-declaration-on-the-secure-foundations-of-enduring-prosperity">Source: U.S. State Department</a> &#183; <a href="https://www.streetinsider.com/Reuters/US%2C+allies+team+up+to+protect+AI+tech+and+infrastructure+amid+China+rivalry/27346833.html">Reuters via StreetInsider</a> &#183; <a href="https://www.wsj.com/tech/ai/trump-administration-to-name-nvidia-among-ai-national-champions-c9284344">Wall Street Journal</a> &#183; <a href="https://www.investing.com/news/stock-market-news/trump-administration-to-name-nvidia-other-ai-firms-as-us-national-champions-wsj-reports-5028349">Reuters via Investing.com</a></p><div class="callout-block" data-callout="true"><h4>Supported by Meter</h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter-daily">See how Meter builds data center networks</a></p></div><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>Vera Rubin timing</strong> &#8212; Commercial Times says HBM4 bottlenecks push Vera Rubin&#8217;s volume ramp into the first half of 2027, with GB300 filling the gap, and Taiwan&#8217;s substrate and PCB stocks fell hard on Monday (<a href="https://abmedia.io/nvidia-vera-rubin-delay-hbm4-taiwan-supply-chain">ABMedia</a> &#183; <a href="https://en.infomaxai.com/news/articleView.html?idxno=148794">Yonhap Infomax</a>).</p></li><li><p><strong>Google TPU</strong> &#8212; Ming-Chi Kuo says OpenAI and Anthropic orders extend Google&#8217;s TPU visibility to 2029, and that TPU v9 onward uses Intel&#8217;s EMIB-T packaging (<a href="https://en.kocpc.com.tw/archives/27686">King of Computer Media</a> &#183; <a href="https://x.com/jukan05/status/2122135556784328777">post on X</a>).</p></li><li><p><strong>Taiwan server makers</strong> &#8212; Taiwan&#8217;s AI server makers are vetting customers and shipping destinations and pausing orders from unfamiliar small clients (<a href="https://money.udn.com/money/story/5612/9877864">Economic Daily News</a>).</p></li><li><p><strong>Korean memory visits</strong> &#8212; Korean memory makers&#8217; planned November visits to Nvidia have been canceled or postponed (<a href="https://x.com/damnang2/status/2121352502595747846">Damnang on X</a>).</p></li><li><p><strong>TSMC glass substrates</strong> &#8212; TSMC is working with AUO on glass substrates for CoPoS packaging in the two plants it bought from AUO last month (<a href="https://x.com/dnystedt/status/2122122553586307408">Dan Nystedt on X</a>).</p></li><li><p><strong>Nvidia&#8217;s customers</strong> &#8212; SemiAnalysis notes that six of Nvidia&#8217;s seven biggest customers, all but SpaceX, are building their own chips (<a href="https://x.com/SemiAnalysis_/status/2121725048402681946">SemiAnalysis on X</a>).</p></li></ul><h2>Sector Watch</h2><p><strong>Memory</strong></p><ul><li><p><strong>SK hynix</strong> &#8212; Electronic Times says it is ordering more HBM4 back-end tools for Cheongju P&amp;T6, adding over 20,000 wafers a month. (<a href="https://en.etnews.com/20261116200004">Source: Electronic Times</a> &#183; <a href="https://www.trendforce.com/news/2026/11/16/news-sk-hynix-reportedly-plans-new-hbm4-equipment-orders-with-pt6-in-focus-despite-nvidia-uncertainty/">TrendForce</a>)</p></li><li><p><strong>Nanya and Winbond</strong> &#8212; Taiwan&#8217;s DRAM makers are adopting ASML EUV, with Nanya&#8217;s new Fab 5A taking wafers from the second half of 2027. (<a href="https://www.digitimes.com/news/a20261116PD210/dram-samsung-euv-winbond-sk-hynix.html">Source: Digitimes</a> &#183; <a href="https://biz.chosun.com/en/en-it/2026/11/13/R246PGDEGRGZROXRGNCGU5FTQY/">Chosunbiz</a>)</p></li><li><p><strong>Wonik QnC</strong> &#8212; The quartz and cleaning supplier&#8217;s third-quarter operating profit jumped 455% to 44.4 billion won on strong chip demand. (<a href="https://heyholder.com/disclosure/20261116000665">Source: Wonik QnC filing via Heyholder</a> &#183; <a href="https://www.tradingview.com/news/urn:summary_document_report:quartr.com:4634101:0-wonik-qnc-9m-earnings-rose-as-semiconductor-demand-lifted-quartz-and-cleaning-sales/">TradingView</a>)</p></li></ul><p><strong>Foundry and fabs</strong></p><ul><li><p><strong>Samsung</strong> &#8212; Seoul Economic Daily says a $1.9 billion Austin expansion will make STMicro&#8217;s Starlink microcontroller, with production targeted by 2030. (<a href="https://en.sedaily.com/business/2026/11/15/samsung-to-make-spacex-satellite-chips-in-texas">Source: Seoul Economic Daily</a> &#183; <a href="https://www.digitimes.com/news/a20261116VL215/samsung-austin-chips-starlink-plant.html">Digitimes</a>)</p></li><li><p><strong>TSMC</strong> &#8212; The foundry booked a record NT$167.6 billion of government subsidies in nine months, and TSMC Arizona earned NT$54 billion. (<a href="https://www.taipeitimes.com/News/biz/archives/2026/11/16/2003867791">Source: Taipei Times</a> &#183; <a href="https://money.udn.com/money/story/5612/9877322">Economic Daily News</a>)</p></li><li><p><strong>Rapidus</strong> &#8212; President Atsuyoshi Koike reaffirmed 2nm mass production for late 2027 or early 2028, despite reports of SRAM delays. (<a href="https://www.digitimes.com/news/a20261116PD216/rapidus-2nm-production-2027-chips.html">Source: Digitimes</a> &#183; <a href="https://www.asahi.com/articles/ASVCF2QK7VCFULFA020M.html">Asahi Shimbun</a>)</p></li></ul><p><strong>AI infrastructure</strong></p><ul><li><p><strong>Nokia</strong> &#8212; Nokia, Nscale, Telia, Verda and Skeleton formed AI Gigafactory Northern to bid for an EU-backed Nordic and Baltic AI build. (<a href="https://www.nokia.com/newsroom/european-technology-leaders-form-ai-gigafactory-northern-consortium-to-build-trusted-sustainable-ai-infrastructure/">Source: Nokia</a> &#183; <a href="https://www.telecompaper.com/news/telia-and-nokia-tie-up-with-nscale-verda-and-skeleton-to-form-ai-gigafactory-northern-consortium--1588945">Telecompaper</a>)</p></li><li><p><strong>BizLink</strong> &#8212; The AI cable maker posted a seventh straight quarterly revenue record, with high-performance computing reaching 46% of sales. (<a href="https://www.investing.com/news/company-news/bizlink-q3-2026-slides-margins-hit-records-as-hpc-surges-61-yoy-93CH-5024360">Source: Investing.com</a> &#183; <a href="https://www.tradingview.com/news/urn:summary_document_slides:quartr.com:4277626:0-bizlink-hpc-and-spe-growth-lifted-3q26-revenue-to-nt-27-88b/">TradingView</a>)</p></li><li><p><strong>LS Electric</strong> &#8212; The Korean power equipment maker won a $93.8 million switchgear order for a U.S. big tech company&#8217;s AI data center. (<a href="https://en.yna.co.kr/view/AEN20261116002700320">Source: Yonhap</a> &#183; <a href="https://www.koreajoongangdaily.com/business/ls-electric-lands-938-million-us-ai-deal/12974394">Korea JoongAng Daily</a>)</p></li></ul><p><strong>Materials and legal</strong></p><ul><li><p><strong>ADEKA</strong> &#8212; The Japanese chemical maker is teaming with Entegris to produce advanced materials for 3D chips in Japan from October 2027. (<a href="https://prtimes.jp/main/html/rd/p/000000044.000072203.html">Source: ADEKA via PR Times</a> &#183; <a href="https://www.marketscreener.com/news/adeka-and-entegris-collaborate-on-advanced-materials-for-3d-semiconductor-devices-powering-ai-ce785edede88f021">MarketScreener</a>)</p></li><li><p><strong>Intel</strong> &#8212; Shareholders sued Intel, Lip-Bu Tan and David Zinsner over the outlook that sent the stock down 29% in a day. (<a href="https://www.investing.com/news/stock-market-news/intel-sued-by-shareholders-following-29-stock-plunge-5028171">Source: Reuters via Investing.com</a> &#183; <a href="https://www.globenewswire.com/news-release/2026/11/14/3403029/0/en/intc-investor-alert-robbins-geller-rudman-dowd-llp-files-class-action-lawsuit-against-intel-corporation-and-announces-opportunity-for-investors-with-substantial-losses-to-lead-the-.html">Robbins Geller via GlobeNewswire</a>)</p></li></ul><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark></p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/intel-custom-asics">Intel Custom ASICs?</a> and <a href="https://newsletter.chipstrat.com/p/qualcomm-custom-asics">Qualcomm Custom ASICs?</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/why-ai-data-center-constraints-make">Why AI Data Center Constraints Make Efficiency More Valuable</a></p></li></ul><p><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better!</mark> <a href="http://semidoped.com/poll"><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Give us your feedback</mark></a></p><p>Want to become a sponsor? Send an email to <a href="mailto:sponsor@semidoped.com">sponsor@semidoped.com</a></p>]]></content:encoded></item><item><title><![CDATA[Sunday Pub Quiz]]></title><description><![CDATA[A magnetic stabilizing ring, a power cell, energy gate, primary and focussing crystals, cycling field energizer and energy modulation circuits are all the components of what?Thanks for reading!]]></description><link>https://daily.semidoped.com/p/sunday-pub-quiz-758</link><guid isPermaLink="false">https://daily.semidoped.com/p/sunday-pub-quiz-758</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Sun, 15 Nov 2026 12:03:34 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!SmUY!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>A magnetic stabilizing ring, a power cell, energy gate, primary and focussing crystals, cycling field energizer and energy modulation circuits are all the components of what?</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Thanks for reading! Subscribe for free to receive new posts and support our work.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div><h3>CLUES</h3><blockquote><p>Recognizable by its distinct <em>snap-hiss</em> ignition, its blade is held in a compact loop by a magnetic stabilizing ring and cycling field energizer.</p></blockquote><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!SmUY!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!SmUY!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png 424w, https://substackcdn.com/image/fetch/$s_!SmUY!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png 848w, https://substackcdn.com/image/fetch/$s_!SmUY!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png 1272w, https://substackcdn.com/image/fetch/$s_!SmUY!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!SmUY!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png" width="1188" height="496" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/d8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:496,&quot;width&quot;:1188,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:587798,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://daily.semidoped.com/i/223651606?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!SmUY!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png 424w, https://substackcdn.com/image/fetch/$s_!SmUY!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png 848w, https://substackcdn.com/image/fetch/$s_!SmUY!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png 1272w, https://substackcdn.com/image/fetch/$s_!SmUY!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8b808b7-8793-469d-a0f9-ec4ca1237de7_1188x496.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3>SPOILER ALERT: ANSWER BELOW</h3><div class="pullquote"><h4>Answer: <strong>A</strong> <strong><a href="https://en.wikipedia.org/wiki/Lightsaber">lightsaber</a></strong> (from the <em><a href="https://en.wikipedia.org/wiki/Star_Wars">Star Wars</a></em> universe), which matches the exact internal component breakdown listed.</h4></div><div class="captioned-button-wrap" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/sunday-pub-quiz-758?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="CaptionedButtonToDOM"><div class="preamble"><p class="cta-caption">Thanks for reading! This post is public so feel free to share it.</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/sunday-pub-quiz-758?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://daily.semidoped.com/p/sunday-pub-quiz-758?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p></div><p></p><p></p>]]></content:encoded></item><item><title><![CDATA[TIL: The Radio With a Suction Pump]]></title><description><![CDATA[In 1942 Hedy Lamarr and George Antheil patented a radio that changed frequency with a suction pump and a piano roll. The same job is now done on a chip in about 220 microseconds.]]></description><link>https://daily.semidoped.com/p/til-the-radio-with-a-suction-pump</link><guid isPermaLink="false">https://daily.semidoped.com/p/til-the-radio-with-a-suction-pump</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Sat, 14 Nov 2026 16:03:29 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/da7a2ef7-a20a-474d-969e-6e5ecdf86b90_250x313.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><a href="https://patents.google.com/patent/US2292387A/en">US Patent 2292387</a> describes a radio that changes channels with a suction pump. A paper strip, punched like a player-piano roll, is pulled by a clock motor past a row of small pistons that the pump holds up against their springs. When a hole passes over a piston, air rushes in, the piston drops, and it closes a switch that connects one of the transmitter&#8217;s tuning condensers. A second strip on a second motor does the same thing in the receiver, so the two radios hop from frequency to frequency in a pattern only the strips know, and a jammer parked on one channel can&#8217;t follow. A standard piano roll, the patent notes, has 88 rows of holes, enough for 88 frequencies.</p><p>The receiver was meant to ride inside a torpedo. The patent was granted in August 1942 to <a href="https://en.wikipedia.org/wiki/Hedy_Lamarr">Hedwig Eva Maria Kiesler</a> and <a href="https://en.wikipedia.org/wiki/George_Antheil">George Antheil</a> &#8212; a Hollywood actress (whom MGM rebranded as Hedy Lamarr) and a composer. According to Hedy, the idea was hers, but the implementation was George&#8217;s.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!k3IK!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg 424w, https://substackcdn.com/image/fetch/$s_!k3IK!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg 848w, https://substackcdn.com/image/fetch/$s_!k3IK!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!k3IK!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!k3IK!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg" width="396" height="289.08" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:219,&quot;width&quot;:300,&quot;resizeWidth&quot;:396,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;File:Lamarrpatent.jpg&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="File:Lamarrpatent.jpg" title="File:Lamarrpatent.jpg" srcset="https://substackcdn.com/image/fetch/$s_!k3IK!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg 424w, https://substackcdn.com/image/fetch/$s_!k3IK!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg 848w, https://substackcdn.com/image/fetch/$s_!k3IK!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!k3IK!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bf1bcd5-5c12-468a-97bd-6269040f31a1_300x219.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div><figcaption class="image-caption"><a href="https://patents.google.com/patent/US2292387A/en">US Patent 2292387</a></figcaption></figure></div><p>In 1933, Hedy married <a href="https://en.wikipedia.org/wiki/Fritz_Mandl">Fritz Mandl</a>, an Austrian arms manufacturer, who kept her beside him through hundreds of dinners and meetings with weapons developers and buyers. Radio-controlled torpedoes reportedly came up at those tables, and so did the trouble with them: find the control frequency, flood it with noise, and the torpedo is lost.</p><p>She left Mandl in 1937 and signed with MGM, and between films she kept an inventing table at home. <a href="https://en.wikipedia.org/wiki/Howard_Hughes">Howard Hughes</a>, whom she was dating, gave her a smaller set of equipment for her trailer on set. Incidentally, he was an aviator, aerospace engineer, film producer, and an investor.</p><p>In the summer of 1940 she met George Antheil at a dinner party. His best-known piece, <em><a href="https://en.wikipedia.org/wiki/Ballet_M%C3%A9canique">Ballet M&#233;canique</a></em>, was scored for 16 synchronized player pianos, airplane propellers, sirens and electric bells. He could not get the pianolas to stay in sync in the 1920s, so he already knew the problem of making machines keep time with each other. He had also written Esquire columns and books claiming you could read a person&#8217;s character from their glands, a pseudoscience that was fashionable at the time, and, as he told it, that&#8217;s why Hedy came over: she wanted to ask him how to enlarge her bust.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!SUwy!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!SUwy!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg 424w, https://substackcdn.com/image/fetch/$s_!SUwy!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg 848w, https://substackcdn.com/image/fetch/$s_!SUwy!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!SUwy!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!SUwy!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg" width="404" height="505.808" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:313,&quot;width&quot;:250,&quot;resizeWidth&quot;:404,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!SUwy!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg 424w, https://substackcdn.com/image/fetch/$s_!SUwy!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg 848w, https://substackcdn.com/image/fetch/$s_!SUwy!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!SUwy!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c304aa7-fb5b-4312-a01e-8294914d80aa_250x313.jpeg 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Lamarr, c.&#8201;1944. By Unknown author - eBay, Public Domain, https://commons.wikimedia.org/w/index.php?curid=47176716</figcaption></figure></div><p>Eventually, the conversation turned to the war. Antheil&#8217;s younger brother Henry, a diplomatic courier, had been killed that June when his plane was shot down over the Baltic, and the loss had left him determined to help. Their torpedo work began in September.</p><p>They worked on it for several months, and in December 1940 they sent a description to the <a href="https://en.wikipedia.org/wiki/National_Inventors_Council">National Inventors Council</a>, a government body that was looking for inventions it could use in the war. Accounts differ on how the council took it: in one, its chairman, <a href="https://en.wikipedia.org/wiki/Charles_F._Kettering">Charles Kettering</a> of General Motors, told them to develop the idea far enough to patent; in another, the council showed little interest. Either way, the application was filed on June 10, 1941. </p><p>The Navy didn&#8217;t take the patent up. The officers reportedly thought the pair wanted to put a player piano in a torpedo, and the patent examiner had already noted that hopping itself was nothing new, so what they owned was only the mechanism. That mechanism is a big part of why the idea stayed on paper. Every hop is air moving through a passage, a piston falling and a contact closing, and the only things holding the ship and the torpedo together are the accuracy of two clock motors and an optional timing pulse sent from the transmitter. The first American hopping radio known to go into service, which <a href="https://en.wikipedia.org/wiki/Sylvania_Electric_Products">Sylvania</a> started building in 1955, replaced the paper with an electronic code generator, and its patents don&#8217;t cite Lamarr and Antheil.</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Thanks for reading! Subscribe for free to receive new posts and support our work.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div><p><a href="https://en.wikipedia.org/wiki/Wi-Fi">Wi-Fi,</a> which usually gets credited to Lamarr, went the other way. The first standard in 1997 allowed hopping as one option, but the version that took off in 1999 used direct sequence, which stays on its channel, and the spin doctors in charge wanted to name it something catchier than &#8216;<a href="https://en.wikipedia.org/wiki/IEEE_802.11b-1999">IEEE 802.11b Direct Sequence</a>.&#8217;</p><p>The same job now sits on a chip. In 1994 an Ericsson engineer named <a href="https://en.wikipedia.org/wiki/Jaap_Haartsen">Jaap Haartsen </a>was designing a short-range radio for the unlicensed 2.4 GHz band, where FCC rules required a transmitter to spread its signal, and he chose hopping partly because a hopping radio occupies only a narrow slice of spectrum at any instant. Classic Bluetooth, the result, splits time into 625-microsecond slots and hops to a new channel up to 1,600 times a second. The pistons work is done by a phase-locked loop, a frequency synthesizer that has to retune the radio and settle on the new channel, typically within about 220 microseconds, and the two paper strips have been replaced by a hop sequence both ends compute from the lead device&#8217;s clock and address.</p><p>In 1997 the <a href="https://en.wikipedia.org/wiki/Electronic_Frontier_Foundation">Electronic Frontier Foundation</a> gave Lamarr and Antheil its Pioneer Award. Antheil had died in 1959. Lamarr, who by then had her own star on the Hollywood Walk of Fame, was in her eighties and living in Florida.</p><p>Her radio needed a suction pump to change channels. If you&#8217;re wearing Bluetooth earbuds, yours have done it thousands of times while you read this sentence, and nobody had to wind them.</p><div class="captioned-button-wrap" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/til-the-radio-with-a-suction-pump?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="CaptionedButtonToDOM"><div class="preamble"><p class="cta-caption">Thanks for reading! This post is public so feel free to share it.</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/til-the-radio-with-a-suction-pump?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://daily.semidoped.com/p/til-the-radio-with-a-suction-pump?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p></div><p></p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 13th, 2026]]></title><description><![CDATA[Washington piles on Nvidia, Applied Materials sees a record 2027, Cisco doubles its AI target, neoclouds wait on buildings, and Samsung eyes Modine]]></description><link>https://daily.semidoped.com/p/daily-update-november-13th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-13th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Fri, 13 Nov 2026 13:33:50 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/dbf3a6fd-36bc-4a42-a3c8-f9d11900eb7f_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Friday, November 13th, 2026</strong></p><p>Nvidia&#8217;s first full day without Jensen Huang brought a Justice Department antitrust probe and a Congress that suddenly agrees on chip exports. The tool and switch makers had a better Thursday: Applied Materials expects wafer fab spending to rise more than 30% next year, and Cisco lifted its hyperscaler AI target to $9 billion. Nebius and Cerebras have the customers and are waiting on the buildings, while Samsung went shopping for chillers.</p><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>Jensen Huang Resigns</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=dWrrSg4XlZA">YouTube</a>, <a href="https://podcasts.apple.com/us/podcast/jensen-huang-resigns/id1866707196?i=1000801597305">Apple Podcasts</a>, or <a href="https://open.spotify.com/episode/5egxsEWge1dO3VOEvFwMLX">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-jensen-huang-resigns">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p>Fine-tune and serve open models without managing a single GPU.</p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p>Accounts are free, and get $5 in credit when you sign up.</p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning &#8594;</a></p></div><h3>Washington piles on Nvidia</h3><p>Multiple Justice Department probes now threaten to sweep up Nvidia, Semafor reported, citing two administration officials. One is an antitrust probe aimed at Nvidia itself. Two others involve Megaspeed and Supermicro and focus on Alice Huang, the former Megaspeed executive at the center of the scandal. Bloomberg adds that Manhattan prosecutors are examining whether entities linked to her helped smuggle Supermicro servers into China. There is no sign of criminal charges against Nvidia or Jensen Huang.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Congress is moving faster now that Nvidia&#8217;s loudest opponent of tighter China controls is gone.</mark> Speaker Mike Johnson said there is a &#8220;developing consensus&#8221; on limiting chip exports, and lawmakers are pushing at least four bills, including one to stop Chinese firms from renting Nvidia GPUs in overseas data centers and one to pressure allies on chipmaking tools. Republicans want to attach curbs to the must-pass defense bill, and Huang is expected to testify before Thanksgiving.</p><p>The stakes are large. SemiAnalysis estimates that more than 1.5GW of restricted Nvidia chips ended up in China anyway. Chinese AI and optics stocks fell on Friday, with Innolight and Eoptolink down more than 8%. Nvidia reports earnings on November 17 with Colette Kress as interim CEO.</p><p><a href="https://www.semafor.com/article/11/12/2026/us-doj-probes-nvidia">Source: Semafor</a> &#183; <a href="https://www.investing.com/news/stock-market-news/speaker-johnson-says-congress-developing-consensus-on-limiting-chip-exports-to-china-5022288">Reuters via Investing.com</a> &#183; <a href="https://www.moneycontrol.com/news/business/jensen-huang-scandal-widens-us-probes-alice-huang-links-to-suspected-super-micro-smuggling-into-china-14078913.html">Bloomberg via Moneycontrol</a> &#183; <a href="https://newsletter.semianalysis.com/p/nvidia-and-the-world-without-jensen">SemiAnalysis</a></p><blockquote><p><strong>Vik: </strong><em>The U.S.-China restrictions on both sides are about to get a lot more restrictive. Washington will impose chip export restrictions, and DeepSeek will withhold its open-source models and restrict the frontier to domestic users.</em></p><p><strong>Austin: </strong><em>Nvidia must continue to invest in American open models; they will continue to be unable to sell to China and have more eyes on any southeast Asia transactions, but there are plenty of enterprises who will buy compute so long as they have frontier open models to run on it. </em></p></blockquote><h3>Applied Materials calls another record year</h3><p>Applied Materials posted record fiscal fourth-quarter revenue of $10.41 billion, up 53% from a year ago, as its chip-equipment business grew about 64% to $8 billion. It guided the next quarter to about $11.1 billion, above the $10.89 billion analysts expected.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">CEO Gary Dickerson expects wafer fab equipment spending to rise more than 30% in 2027, and says customers are asking for tools sooner, not later.</mark> China fell to 25% of revenue from 29% a year ago, even as sales there grew. Applied also bought back $2 billion of stock in the quarter.</p><p><a href="https://www.appliedmaterials.com/us/en/newsroom/press-releases/111226-applied-materials-announces-fourth-qtr-fiscal-year-end-2026-results.html">Source: Applied Materials</a> &#183; <a href="https://www.straitstimes.com/business/applied-materials-forecasts-quarterly-revenue-above-estimates-on-strong-ai-chip-demand">Reuters via The Straits Times</a></p><blockquote><p><strong>Vik: </strong><em>As long as WFE demand is rising, chip demand is still going strong. AMAT posted great results this quarter and gave good guidance for the next. Things are looking up in WFE.</em></p><p><strong>Austin: </strong><em>Or Vik it could be the other way around? WFE is the tail end of the bullwhip, much further downstream from actual customer demand for AI. WFE&#8217;s customers are fabs, and fabs are trying to build in advance of future demand. If that future demand were to diminish, it&#8217;d take a quarter or two to make it to WFE orders. That said, I believe demand is real and sustainable. </em></p></blockquote><h3>Cisco doubles down on AI networking</h3><p>Cisco&#8217;s fiscal first-quarter revenue rose 24% to a record $18.5 billion, its fastest growth in more than 15 years. Networking revenue jumped 43%, and the company raised its full-year guidance to $74.8 billion to $75.8 billion.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Cisco now expects $9 billion of AI infrastructure revenue from hyperscalers this fiscal year, up from $7.5 billion just three months ago and more than double last year.</mark> It won three new hyperscaler design wins in the quarter, two for its Silicon One P200 systems that link data centers and one for 1.6T optics. Non-GAAP product gross margin fell 3.5 points from a year ago.</p><p><a href="https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2026/m11/cisco-reports-first-quarter-earnings.html">Source: Cisco</a> &#183; <a href="https://www.cnbc.com/2026/11/12/cisco-csco-q1-2027-earnings.html">CNBC</a> &#183; <a href="https://www.investing.com/news/company-news/cisco-q1-fy27-slides-ai-revenue-target-raised-to-9b-record-margins-93CH-5023714">Investing.com</a></p><blockquote><p><strong>Vik: </strong><em>Cisco&#8217;s main AI networking story is scale-out, and demand is looking up for its Silicon One P200 systems. I haven&#8217;t been looking at Cisco too closely, but after a great earnings call like this, it&#8217;s time to go deeper.</em></p><p><strong>Austin: </strong><em>Silicon One marketing website mentions scale up, but as Vik pointed out Silicon One is just scale out and scale across today. That said, Cisco has 18 AI design wins over five quarters, two new P200 wins this quarter for scale-across and data center interconnect, a 1.6T client-side optics (inside the datacenter) win for scale-out, and the G300 now shipping into hyperscaler labs. AI revenue guided to $9B for FY27, more than double FY26. </em></p></blockquote><h3>Neoclouds have the demand, not the buildings</h3><p>Nebius grew third-quarter revenue 571% to $981 million, but trimmed its 2026 revenue range to $3.0 billion to $3.2 billion because third-party data center capacity slipped into 2027. It raised 2026 capex to about $27 billion and plans $40 billion to $45 billion next year for $12 billion to $13 billion of revenue. The stock fell about 17%.</p><p>Cerebras beat with core revenue of $231 million, up 70%, and added Microsoft as its third hyperscaler customer after AWS and Oracle. It also admitted it is behind on OpenAI&#8217;s first capacity tranches and expects to get back on schedule in the second half of 2027.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Both companies blamed the same bottleneck: land, power and construction, not chip supply or customer demand.</mark> Nebius says prepayments and financing against existing contracts should cover 75% to 85% of next year&#8217;s capex.</p><p><a href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926126008/tm2629571d1_ex99-2.htm">Source: Nebius shareholder letter (SEC)</a> &#183; <a href="https://www.sec.gov/Archives/edgar/data/2021728/000162828026073366/cbrsannouncesfinancialresu.htm">Cerebras results (SEC)</a> &#183; <a href="https://www.streetinsider.com/Reuters/Nebius+beats+quarterly+revenue+estimates+on+strong+AI+demand/27330371.html">Reuters via StreetInsider</a> &#183; <a href="https://www.cnbc.com/2026/11/12/cerebras-systems-cbrs-q3-earnings-report-2026.html">CNBC</a></p><blockquote><p><strong>Austin: </strong><em>The obvious question in light of the news is whether Nebius will see any impacts from Jensen resigning. In theory no, the supply of GPUs should be fine&#8230;</em></p><p><strong>Vik: </strong><em>Land and power constraints are closely intertwined, and with the new sentiment against data center buildouts, they become increasingly valuable. That means every gigawatt of approved and built-out power becomes extremely valuable. I wrote about this in a free post on Substack: <a href="https://www.viksnewsletter.com/p/why-ai-data-center-constraints-make">Why AI Data Center Constraints Make Efficiency More Valuable</a>.</em></p></blockquote><h3>Samsung eyes Modine to cool AI data centers</h3><p>Samsung Electronics is in talks to buy Modine Manufacturing, a Wisconsin maker of data center cooling systems, Bloomberg reported, citing people familiar with the matter. A deal could come together within weeks, though talks could still fall apart. Modine was worth about $5.6 billion before the report, and its shares jumped 28% after hours.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">A deal would give Samsung a North American base in data center cooling, after it bought Europe&#8217;s Fl&#228;ktGroup for &#8364;1.5 billion last year.</mark> Modine&#8217;s data center revenue rose 95% last quarter to about $440 million, and Seoul Economic Daily says a full takeover could be Samsung&#8217;s largest acquisition ever.</p><p><a href="https://news.bloomberglaw.com/ip-law/samsung-said-in-talks-to-acquire-data-center-cooling-firm-modine">Source: Bloomberg</a> &#183; <a href="https://www.channelnewsasia.com/business/samsung-in-talks-buy-data-center-cooling-firm-modine-bloomberg-news-reports-6547876">Reuters via CNA</a> &#183; <a href="https://en.sedaily.com/business/2026/11/13/samsung-in-talks-to-buy-us-cooling-maker-modine-for-ai-data">Seoul Economic Daily</a></p><blockquote><p><strong>Vik: </strong><em>Downstream of higher power density in data centers is cooling. Samsung sees the trend and is going after cooling systems. With a Modine acquisition, the cooling space will be very interesting to watch over the next two to three years.</em></p></blockquote><div class="callout-block" data-callout="true"><h4>Supported by Meter</h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter-daily">See how Meter builds data center networks</a></p></div><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>Intel Crescent Island</strong> &#8212; Lip-Bu Tan told the SGLang Summit that Intel&#8217;s Crescent Island inference GPU already has about &#8220;2 billion&#8221; in customer purchases, though the currency and order details were not spelled out (<a href="https://x.com/labubu_trader/status/2121037307122012613">post on X</a>).</p></li><li><p><strong>Intel EMIB-T</strong> &#8212; Sell-side notes say Intel is lining up seven customer engagements for its EMIB-T advanced packaging platform (<a href="https://x.com/OptionKing666/status/2121089550659588590">post on X</a>).</p></li><li><p><strong>HBM4 pricing</strong> &#8212; SemiAnalysis estimates Nvidia pays roughly 19% to 20% less for HBM4 than Google, Amazon, AMD and Broadcom, a pricing edge the supply chain is now wondering whether Nvidia keeps after Huang (<a href="https://x.com/pequityresearch/status/2120952627911663968">post on X</a>).</p></li><li><p><strong>Samsung memory orders</strong> &#8212; Samsung told BofA it has seen no order cuts tied to worries about a US AI data center slowdown after the midterms (<a href="https://x.com/pequityresearch/status/2121040870325817790">post on X</a>).</p></li><li><p><strong>Broadcom supply</strong> &#8212; Damnang&#8217;s sources say Broadcom has already secured its supply for 2027 and 2028 as custom silicon looks to gain from Nvidia&#8217;s leadership vacuum (<a href="https://x.com/damnang2/status/2121025368329207833">Damnang on X</a>).</p></li><li><p><strong>Nvidia staff</strong> &#8212; Wired reports Nvidia employees are pushing for Huang&#8217;s return, with an internal suggestion-box post drawing more than 1,600 comments (<a href="https://x.com/SVTrivo/status/2121129927902990575">post on X</a>).</p></li></ul><h2>Sector Watch</h2><p><strong>Foundry</strong></p><ul><li><p><strong>SMIC</strong> &#8212; Record third-quarter revenue rose 34% to $3.19 billion, and management says AI chip supply stays tight through 2027. (<a href="https://www1.hkexnews.hk/listedco/listconews/sehk/2026/1112/2026111200682.pdf">Source: SMIC filing (HKEX)</a> &#183; <a href="https://hk.marketscreener.com/news/smic-says-ai-chip-supply-chain-will-remain-tight-through-2027-ce785ed9db89fe2d">Reuters via MarketScreener</a>)</p></li><li><p><strong>TSMC</strong> &#8212; Leaked employee data traced to an outside cloud provider, with no breach of TSMC&#8217;s own systems. (<a href="https://www.sec.gov/Archives/edgar/data/1046179/000104617926000800/a20261112statementregardin.htm">Source: TSMC filing (SEC)</a> &#183; <a href="https://www.nikkei.com/article/DGXZQOGM126FF0S6A111C2000000/">Nikkei</a>)</p></li></ul><p><strong>AI servers and cloud</strong></p><ul><li><p><strong>Foxconn</strong> &#8212; Third-quarter profit jumped 34% to a record, with AI cloud and networking now 54% of sales. (<a href="https://www.foxconn.com/en-us/press-center/press-releases/latest-news/2154">Source: Foxconn</a> &#183; <a href="https://focustaiwan.tw/business/202611120018">Focus Taiwan</a>)</p></li><li><p><strong>Compal</strong> &#8212; Revenue rose 30%, and the board approved $486 million to expand AI server capacity in the US. (<a href="https://www.compal.com/en-us/media/360/ren-bao-dian-nao-gong-bu-min-guo-115nian-di-san-ji">Source: Compal</a> &#183; <a href="https://www.compal.com/en-us/media/361/ren-bao-dian-nao-fen-ge-ji-chu-she-shi-jie-jue-an-">Compal (spin-off)</a>)</p></li><li><p><strong>Nscale</strong> &#8212; The Nvidia-backed neocloud&#8217;s amended IPO filing shows revenue up about 14-fold alongside a widening loss. (<a href="https://www.sec.gov/Archives/edgar/data/2110365/000119312526458722/ck0002110365-20261112.htm">Source: Nscale S-1/A (SEC)</a> &#183; <a href="https://www.cnbc.com/2026/11/12/nscale-updates-ipo-prospectus-that-shows-widening-loss.html">CNBC</a>)</p></li></ul><p><strong>Equipment and components</strong></p><ul><li><p><strong>Advantest</strong> &#8212; The tester maker pulled its 10,000-testers-a-year target a year forward to end-2027 on AI demand. (<a href="https://www.nikkei.com/article/DGXZQOUB127IY0S6A111C2000000/">Source: Nikkei</a> &#183; <a href="https://www.digitimes.com.tw/tech/dt/n/shwnws.asp?id=774680">Digitimes</a>)</p></li><li><p><strong>ams OSRAM</strong> &#8212; Third-quarter revenue hit &#8364;829 million on about &#8364;1.3 billion of new semiconductor design wins. (<a href="https://ams-osram.com/news/press-releases/q3-2026-results">Source: ams OSRAM</a>)</p></li><li><p><strong>JNTC</strong> &#8212; The Korean glass maker will supply HDD platters to Seagate from 2027, challenging Hoya&#8217;s monopoly. (<a href="https://www.thelec.net/news/articleView.html?idxno=15553">Source: The Elec</a>)</p></li><li><p><strong>Penguin Solutions</strong> &#8212; A new CXL card lets AI servers reuse DDR4 as extra memory amid the shortage. (<a href="https://ir.penguinsolutions.com/news/news-details/2026/Penguin-Solutions-Introduces-CXL-Enabled-DDR4-Module-to-Help-Address-Memory-Shortages-in-AI-Data-Centers/default.aspx">Source: Penguin Solutions</a>)</p></li></ul><p><strong>Leadership and investments</strong></p><ul><li><p><strong>Marvell</strong> &#8212; Three new independent directors join, including Lam Research CFO Doug Bettinger. (<a href="https://investor.marvell.com/news-events/press-releases/detail/1042/marvell-appoints-three-new-independent-directors-to-its-board">Source: Marvell</a>)</p></li><li><p><strong>AMD</strong> &#8212; AMD will take an equity stake in PsiQuantum to co-design utility-scale quantum computing infrastructure. (<a href="https://www.businesswire.com/news/home/20261112481657/en/PsiQuantum-and-AMD-Collaborate-to-Build-the-Computing-Foundation-for-Utility-Scale-Quantum-Computing">Source: PsiQuantum and AMD</a> &#183; <a href="https://money.usnews.com/investing/news/articles/2026-11-12/amd-to-invest-in-psiquantum-to-advance-quantum-computing-infrastructure">Reuters via U.S. News</a>)</p></li></ul><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark></p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/amd-q3-fy26-annotated">AMD Q3 FY26, Annotated</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/why-ai-data-center-constraints-make">Why AI Data Center Constraints Make Efficiency More Valuable</a></p></li></ul><p><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better!</mark> <a href="http://semidoped.com/poll"><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Give us your feedback</mark></a></p><p>Want to become a sponsor? Send an email to <a href="mailto:sponsor@semidoped.com">sponsor@semidoped.com</a></p>]]></content:encoded></item><item><title><![CDATA[🎙️ NEW EPISODE: Jensen Huang Resigns]]></title><description><![CDATA[Jensen Huang resigned as Nvidia CEO this morning, and Austin Lyons and Vik Sekar recorded two hours after the news broke.]]></description><link>https://daily.semidoped.com/p/new-episode-jensen-huang-resigns</link><guid isPermaLink="false">https://daily.semidoped.com/p/new-episode-jensen-huang-resigns</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Thu, 12 Nov 2026 22:50:04 GMT</pubDate><enclosure url="https://substack-video.s3.amazonaws.com/video_upload/post/224673178/df4d7f6a-ef2d-4151-9f87-e79cf97c4ad2/transcoded-00001.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Jensen Huang resigned as Nvidia CEO this morning, and Austin Lyons and Vik Sekar recorded two hours after the news broke. They argue that a board must prioritize corporate integrity over short-term shareholder value, even for a legendary founder. The discussion then turns to Nvidia&#8217;s critical lack of a succession plan and the long-term risk of the company&#8217;s unique, founder-led culture decaying into bureaucracy.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p><strong>Fine-tune and serve open models without managing a single GPU.</strong></p><p>Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour.</p><p>Accounts are free, and <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">get $5 in credit</mark> when you sign up.</p><p><strong><a href="https://semidoped.com/crusoe">Try Serverless Fine-Tuning &#8594;</a></strong></p></div><p><strong>Key takeaways:</strong></p><ul><li><p>A board&#8217;s first duty is governance, not shareholder value. Even a founder CEO must be forced out for breaking company rules, because keeping them sends a message that lying and rule-breaking are acceptable.</p></li><li><p>The hardest job demands the highest standard. The immense pressure of being Nvidia&#8217;s CEO is the reason for strict conduct rules, not an excuse for leniency.</p></li><li><p>Nvidia has no succession plan &#8212; Ian Buck is the only plausible internal CEO candidate. Unlike competitors, Nvidia&#8217;s &#8216;all Jensen, all the time&#8217; media strategy has left its leadership bench visibly thin.</p></li><li><p>Jensen&#8217;s unique management style is the biggest succession risk. His flat structure with ~60 direct reports is impossible for a successor to replicate, creating a vacuum that invites bureaucracy.</p></li><li><p>The real long-term risk is Nvidia decaying into a spreadsheet-CEO culture. The fear is a slide into the managerial, process-driven leadership that hampered competitors like Intel.</p></li><li><p>A leadership crisis at Nvidia is AMD&#8217;s opportunity to attack. For a competitor like Lisa Su, Nvidia&#8217;s moment of weakness is the time to &#8216;put my foot on the gas,&#8217; not switch teams.</p></li><li><p>Arm CEO Rene Haas is a credible external successor. As an Nvidia alumnus, he understands the culture but also brings proven CEO experience from a different silicon leader.</p></li></ul><div class="callout-block" data-callout="true"><p><strong>This episode is supported by Meter.</strong></p><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks? Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS. One call reaches the team that designed the hardware and wrote the software.</p><p><strong><a href="https://semidoped.com/meter">See how Meter builds datacenter networks &#8594;</a></strong></p></div><p><em>This podcast is lightly edited for clarity.</em></p><h2>The News Breaks: Jensen Huang Resigns</h2><p><strong>Austin:</strong> Hello listeners and welcome to another Semi Doped episode. I&#8217;m Austin Lyons of Chipstrat. I write a newsletter for institutional investors and tech executives.</p><p><strong>Vik:</strong> Hey, I&#8217;m Vik Sekar. I write a Substack newsletter called Vik&#8217;s Newsletter. And I run an advisory firm for financial institutions called Semi Exponent. And today, we are here because there is very big news in the chip world and we thought we&#8217;ll have a quick news flash about this and talk about it. Jensen Huang is no longer the CEO of Nvidia. That&#8217;s it. That&#8217;s the big news.</p><p><strong>Austin:</strong> That&#8217;s the news. Let&#8217;s break it down.</p><p><strong>Vik:</strong> We don&#8217;t normally talk about this type of content. Normally we focus on technology and product and business strategy and stuff. And obviously today, the topic is going to, we&#8217;re going to stick with questions like, should the board if the board forced him to resign, he resigned. That&#8217;s the news. He resigned. If the board forced him to resign, was that the right decision? We&#8217;ll let&#8217;s debate that. And then we&#8217;ll also talk, maybe we can debate kind of lighthearted, who might fill him in next. Or like who might be the next Nvidia CEO. We know that Colette Kress, the CFO is the interim CEO, but let&#8217;s talk through who it could be.</p><h2>The Backstory: National Security and Corporate Policy</h2><p><strong>Austin:</strong> We should set the stage for anybody who&#8217;s listening to this and not following this stuff as closely as some other people like listening to this podcast or us included follow this kind of stuff very closely. But essentially, we have to set the stage on why we are in this situation and I say we because Nvidia is now the face of AI and what happens to Nvidia and its leader Jensen Huang affects all of us in some way, shape or form. Also, Nvidia is a large part of the S&amp;P 500 and if something is going to move, the entire world wants to know.</p><p>So, we are in this situation and we will cover this very briefly because a lot of personal lives are involved and we don&#8217;t want to make fun of it. It&#8217;s a hard time for a lot of people. So, we want to be respectful of that and then go to talking about the technology and what we think is going to happen in the future.</p><p>So, essentially, sometime back, reports surfaced that Jensen had extramarital affairs with an executive from this company called Megaspeed, which is Singapore-based. The whole problem with that was Megaspeed supposedly was had relations to China and therefore this whole thing became a national security concern. Whether this company is using Jensen&#8217;s personal ties to supply leading edge Nvidia chips to China. So, that became the first big concern at the national level. And while that investigation is still underway to figure out what exactly is the impact and there is some movement regarding what policy changes will come towards China on that front. It is unclear still, but we might expect more tighter restrictions on China going forward because of this.</p><p>But then I think a couple of days ago, another report surfaced that there was another extramarital affair with an internal Nvidia employee. And then essentially that became a problem and for anybody who&#8217;s worked in professional environments and companies and such, companies generally frown upon this kind of stuff very severely, especially you don&#8217;t want to have relationships with somebody who&#8217;s like reporting up to you. And the reason for that is simple because workplaces is supposed to be fair and equitable and it&#8217;s based on merit. And of course, when these kinds of stuff happens, people don&#8217;t like it and companies have policies against these things.</p><p>And so, when Jensen was faced with this, in my opinion, I think the company and the board was left with no choice to but to stick to corporate policy because even the CEO is not an exception. Jensen Huang is not above the law, nobody is. And whether it&#8217;s like company policy, they had to do what they had to do. Of course, none of this is what Nvidia said is the reason why Jensen Huang resigned, but I think everybody kind of sees this the second incident as basically something he was not going to recover from.</p><p>So, we have lost the leader of a big company that is well respected. And Jensen Huang was the brainchild of Nvidia for 35 years and is probably his baby that he&#8217;s seen grow up to a $4 trillion company. And so, that accomplishment, regardless of his shortcomings or his mistakes or whatever you want to call it, that accomplishment of growing a company for multiple decades to become one of the biggest companies of the world today and one of the most important companies is something that we can&#8217;t forget and something we have to respect as a technologist and as a visionary. So, we always have Jensen as a leader of the AI movement and we will remember him proudly on that as his leadership that took us to where we are today. But now that we have an interim CEO and Nvidia is now looking for an alternative CEO to take his place, whether it&#8217;s internal or external, we don&#8217;t know. But, that&#8217;s where we are today.</p><p><strong>Austin:</strong> This episode of Semi Doped is presented by Crusoe, where you can fine-tune and serve open models without managing a single GPU. When we sat down with Crusoe&#8217;s serverless fine-tuning product team, we asked the question every AI user wonders, what will you do with my data? Crusoe has a simple answer. They don&#8217;t touch it or see it. Your data is yours to fine-tune models with or simply store for future use. Their fine-tuning is designed so that both your training data and the fine-tune model itself remain your property. Your training data is encrypted, so even Crusoe&#8217;s cloud engineering team can&#8217;t see it. You can deploy it on Crusoe Cloud in one click or download the checkpoints and run it anywhere you&#8217;d like. The fine-tuning process is fully managed, so you upload a dataset, pick an open model, and Crusoe takes care of the GPUs, orchestration, and all the infrastructure management. You pay per token processed rather than per GPU hour, so run a few versions, compare them, and download the one that actually works best for your use case. Create a new Crusoe account for free and get $5 in credit to start testing today. This part is super important, guys. Support Semi Doped by going to semidoped.com/crusoe, click on that link in the show notes, type it in yourself, and you&#8217;ll get directed to a sign-in page. Sign in with Google or create a new account, and Crusoe will immediately create that account for you with $5 in it. Boom. Again, just click that link in the show notes, sign in, $5 for you, and you support Semi Doped. So, thank you for your support. Go check out Crusoe, try to fine-tune your own data. Again, the link to try serverless fine-tuning is in the show notes. So, hit that now, click through, and make sure you claim that $5, give it a try, and support us. Back to the show.</p><h2>What Should the Board Do?</h2><p><strong>Vik:</strong> What&#8217;s your thought when this came out? Which is this news even dropped only like a couple of hours ago and we got to recording immediately. So, I haven&#8217;t had a chance to talk to you about what you think about it. So, what do you think?</p><p><strong>Austin:</strong> Yes, this will be fresh and raw and honest. So, yes, right now we&#8217;re recording. It&#8217;s Thursday, November 12th, it&#8217;s 10:00 a.m. Central Time. So, this news came out about two hours ago and we have not talked about it. So, when I read, okay, and by the way, the news about the direct report extramarital affair thing came out just yesterday, Wednesday. And so, this has all gone down very quickly. When I saw the announcement this morning that Jensen was resigning, I thought to myself, the question I had was, if I was on the board, what would I do?</p><p>I think everyone&#8217;s chiming in on X and you can tell everyone&#8217;s biases, at least the people talking on X where I think there&#8217;s a lot of Nvidia bag holders, you could call them, people with financial interest in investment, a lot of people, right? Because by the way, not only to your point is Nvidia, whatever, 8% of the S&amp;P 500. Also there&#8217;s a lot of other companies in the S&amp;P 500 that also benefit from Nvidia, whether it&#8217;s HBM makers or, CPU makers, other other people in the industry. So, there will be, it&#8217;s not just one company, but everyone&#8217;s kind of intertwined.</p><p>So anyway, so the question is, what should the board do? And, there&#8217;s one argument that&#8217;s like, hey, the best thing for shareholder value is to let Jensen keep running the company because he&#8217;s built it, he has the most context, he&#8217;s gotten it to where it is. Yes, he made some mistakes, but that&#8217;s the best thing in shareholder interest. Of course, to that I would say, sure, in the short run, you can make the argument, hey, the stock&#8217;s down, whatever percent and it&#8217;d probably be down less if he was not resigning. I sort of have a different zoomed out, sort of longer term perspective, which is, and you kind of alluded to this a little bit, but what message does it send if Jensen stays? I think it sends the message that it&#8217;s okay for the leader of the company to lie, which is what happened with the Alice Huang situation where he said, hey, there&#8217;s nothing going on here and then it comes out later.</p><p><strong>Vik:</strong> The Megaspeed employee, the first one.</p><p><strong>Austin:</strong> The Megaspeed employee.</p><p><strong>Vik:</strong> I have a counter take to what you say actually because I recently saw this interview of Jensen Huang with Bloomberg original, I believe. And it was a very well-produced interview and he was explaining basically what it takes to be a CEO of a company like Nvidia. The interviewer asked him, would you put your kids up to it? He was hesitant. He was like, I think they should be very careful in the decisions they make because the kind of pressure, the stress and the responsibility that you have to run the ship is enormous. And whoever the successor is, and in that interview, which was actually now come to think of it is kind of sad, he says, in 20 years, there&#8217;s going to be a successor and I hope that person can handle this job, you know, and then the moment I saw that and two days later, we are looking for a successor. It&#8217;s kind of a sad situation to be in.</p><p>But to the point that he was saying is, I think that it&#8217;s a very difficult position, a very hard job to actually hold. And you and I run a small outfit, right? Like we run a podcast, we run our substacks. You know, how many decisions we have to make on our day-to-day thing? Our listeners probably don&#8217;t really see all the naly details of running a tiny business like ours. Now, you know, multiply that and make it a $4 trillion company. It is insane, right? So I feel like, okay, why do all these rich people somehow get into this kind of scandals, right? So I feel like at some level, which I don&#8217;t really know and I don&#8217;t think I ever will understand truly, is that they have some need for an emotional connection somewhere. Their family is there for them. Yes, but I don&#8217;t know why it seems like they need something more. Maybe it&#8217;s lonely at the top. I don&#8217;t know. I&#8217;ll never find out. I&#8217;m never going to be a CEO of like Nvidia or anybody of that caliber level, really. I don&#8217;t think so. But yeah, that&#8217;s my counter argument. I don&#8217;t think it&#8217;s so much as confidence as people need some kind of emotional connection to deal with the stress they have, maybe.</p><p><strong>Austin:</strong> Hmm.</p><p><strong>Vik:</strong> But yeah, who knows what&#8217;s going on in his head? I don&#8217;t know. It&#8217;s all speculation.</p><p><strong>Austin:</strong> Right, right, right. I do acknowledge it&#8217;s got to be one of the hardest jobs in the world for sure. There&#8217;s still like rules, for example, that you can and can&#8217;t break. And so regardless of how hard your job is, I want the leader that I report to or the leader that I look up to, the leader of my company to be held to the highest standard, not the lowest standard. And I think if it&#8217;s like, yes, he&#8217;s in the highest pressure job and therefore he can break company rules or he can lie. That&#8217;s like because it&#8217;s a hard job, we&#8217;re going to lower the standard for him. And so I think it&#8217;s like Jensen, this is what you signed up for. You&#8217;re taking the hardest job and you&#8217;re held to the highest standard. It&#8217;s not easy.</p><p><strong>Vik:</strong> Yes, and you know, the benefits that come with it as a CEO, the respect, when you walk through this Computex, we were at Computex earlier this year when he walked on the floor, we couldn&#8217;t even stand. Like we had to leave the show floor because so many people piled in to see Jensen, right? Like that kind of thing, that kind of a fame and fortune comes at a price, I suppose. But anyway, what do you think is going to happen to the AI trade? Let&#8217;s get to the brass tax of what everybody&#8217;s thinking about.</p><h2>The Future of the AI Trade</h2><p><strong>Austin:</strong> Well, I&#8217;m still using Claude today. I&#8217;m still using OpenAI today. It&#8217;s still very productive for me. I&#8217;m still believing that Nvidia&#8217;s roadmap will happen with Rubin Ultra and with Feynman. I think I&#8217;ll still use an x86 laptop and I think AMD will still compete with GPUs. So on the one hand, again, as we&#8217;ve been saying, this is the story of the sort of AI trade, which is always just like the value&#8217;s real, the demand is real. The stock market moves up and down on various things because they&#8217;re pricing in a future, a future value of a company.</p><p>So, obviously, what&#8217;s going to happen to Nvidia&#8217;s future? Well, all we know in the short term is, right, Jensen Huang is not leading it. In the interim, Colette is leading it. You would like to think that it&#8217;s just like, okay, run in maintain mode. Let&#8217;s not make any drastic changes. Just execute on the roadmap for the next year. In the meantime, we&#8217;re going to go figure out who the CEO, the next CEO will be and we&#8217;ll figure that out. Maybe it&#8217;s someone internal, maybe it&#8217;s someone external, we can talk through that. But you would like to think that the ship in the near term, maybe for the next year or so, could continue to keep going.</p><p>As long as because it, right, it&#8217;s obviously Jensen&#8217;s the face of the company and he&#8217;s got the 60 direct reports and clearly he has a say in everything, sort of like a Steve Jobs. On the other hand, there are lots of really smart people that work in Nvidia. And so even if the captain is not on the ship anymore, clearly everyone rowing knows the direction they&#8217;re heading and can still row and the ship will continue to go. So on the one hand, I&#8217;d like to believe that regardless of the stock price and what the weighing machine of the stock market thinks that that Nvidia can put their heads down and keep rowing and keep going. But yeah, what do you think?</p><p><strong>Vik:</strong> So, yeah, short term, I agree that the demand hasn&#8217;t gone away, the technology hasn&#8217;t fundamentally changed, nothing has changed. All the projects internally will progress as planned, I hope. But, from having worked at big companies and I&#8217;m sure you have the same perception of this kind of stuff is that when there is a vacancy at the helm, especially a vacancy like this one, there are going to be a lot of contenders. And there&#8217;s going to be a lot of in fighting, politics, all of this takes away from what the job is at hand, which is to continue to maintain Nvidia at the leading edge of AI hardware. People lose track of that when they lose a leader like him.</p><p>You know, he had a relatively flat organization. I was reading that he has like 60 direct reports, right? He&#8217;s not like, I have three direct reports who then have 20 each or something like that, right? He&#8217;s like, he had 60 direct reports. So he was a very hands-on, culture, the culture of Nvidia was driven by his founder approach to doing things. He&#8217;s just, go directly and get it done kind of an environment. My fear, okay, in the short term, I don&#8217;t have any fear. Like, I think everything&#8217;s going to be fine. A new CEO will come and do this and that. All the sales pipelines are set up. Good positions are there, long-term agreements are in place. The scenario I fear is that somehow Nvidia sinks into&#8212;</p><p><strong>Austin:</strong> Right, right, right. So, so you make the very good point here, which is just the difference between a founder CEO and anyone else, any other CEO that comes in. Obviously, you want to have like a founder mindset, high ownership, and I think, for example, Lit Bhutan, he came in and I think people thought like, oh yeah, he&#8217;s, he like, he&#8217;s, he&#8217;s crying because he cares so much about Intel and he has that same type of mindset. But at the end of the day, it&#8217;s just different than the Elon, Zuck, Jensen type people who started the company. They can just make decisions differently and they can think very long-term and take different risks than anyone else who comes in and you&#8217;re now the new CEO, right?</p><p><strong>Vik:</strong> Correct. I love the example of Elon you brought up because I feel like Elon and Jensen have very similar traits. In fact, they are like extreme executors of a vision, right? Like they have a vision and they execute it to the last breath they have. Like Jensen was one such guy. Like I&#8217;ve seen his earlier interviews and he says, I was on vacation all the time. I never stopped working. My family was on vacation. I never stopped working. I kept, there was some note about how many emails he could answer in in an hour and it was something like hundreds or something like that. I&#8217;m like, how do you even read hundreds? And this was before AI, right? Like, how do you even read hundreds? So, he&#8217;s that kind of like a crazy execution guy. And I like the Elon example because supposing Elon was removed from his companies, like do you think any of that stuff will run the way it does? Right? Like, do you think somebody else could run Elon&#8217;s business the way he does?</p><p><strong>Austin:</strong> It would definitely, definitely, definitely be different. Cuz to your point, right? Like the 60 direct reports, when Jensen&#8217;s not there to just make the executive decision, now how are they going to decide who gets more budget and who gets flat budget next year, right?</p><p><strong>Vik:</strong> So it will fall into a bureaucracy, a red tape like situation where you&#8217;re like, oh, you need to get approval from this VP, or you need to get that approval. And then, those all flow upwards to another decision maker, moves to another decision maker. And by the time, the whole execution and the beauty of Nvidia is lost in this in this mess.</p><p><strong>Austin:</strong> Totally, totally. So no Jensen to just say, yes, yes, no, no, no in his emails all every Sunday. And same thing with the Elons, when you think about him sleeping at the SpaceX or the Tesla or the X or whatever, just you&#8217;re going to lose that. Part of why some of these companies can move so fast too is because they have that founder who has the long-term vision and can just make those decisions quickly. And of course, you can bring in CEOs that have a vision, you know, Satya Nadella has done a nice job at Microsoft. But it&#8217;s still there&#8217;s going to be a risk, which is like, when they bring someone in, what is their vision and can they execute like that?</p><h2>Who Could Be the Next CEO?</h2><p><strong>Vik:</strong> Yeah, I like your counter example of Satya Nadella. I think he really turned around Microsoft like really well. Like he adopted GitHub and the open source mindset and he really took it to a new level. Yeah, that&#8217;s a great example. I really hope somebody like that comes over and takes the helm here at some point. Who do you think that&#8217;ll be? Give me your top pick.</p><p><strong>Austin:</strong> Okay, let yeah, let&#8217;s get into it. All right, we we can just close with this too. Okay, some names that have been floating around. Someone online said Lisa Su. Now, I&#8217;ll tell you what, Lisa Su, so when I was thinking through the example of leaders who bring out the best in people, you can hold them to the highest standard, cuz you can make the argument of like, oh yeah, Austin, you&#8217;re being very idealistic. You want so much from a CEO, but like look at all these other CEOs, they have other ones have moral failings too. But then I&#8217;m like, no, I think I can find counter arguments, and I think of like Lisa Su was the first one who popped into my mind of just like super respect, I have tons of respect for her as an engineer, as a business strategist, as a leader. And so, I think very highly of her. Now, would why would she come over to Nvidia and run Nvidia? She&#8217;s doing, she&#8217;s running AMD and they&#8217;re they&#8217;re doing awesome. So, to the people that put that on X, no.</p><p><strong>Vik:</strong> What? No, no, no, dude, come on. Why should Lisa Su has always been trying to play catchup. Now, what if you suddenly given the the race car that wins the pole position, like as an F1 driver, would you not take that car?</p><p><strong>Austin:</strong> Man, she&#8217;s, yeah, yeah, I think she&#8217;s always had a chip on her shoulder and she&#8217;s David trying to take down Goliath. Now does she just want to flip sides? No, I think this is the time for Lisa to dig in and say, hey, Intel just had a terrible quarter and we&#8217;re crushing on the x86 side. Nvidia&#8217;s having a hard time. This is a huge opportunity for us. If I&#8217;m her, this is where I like put my foot on the gas, not switch teams.</p><p><strong>Vik:</strong> I totally agree. Yeah. I totally agree. Look at Nvidia, look at AMD&#8217;s position now. Like Intel is all over the place. We spoke about this on a previous podcast, like what&#8217;s going on here, right? And AMD comes right after that in their earnings call. You wrote a Substack post on this, like and they crushed it. They&#8217;re like, what happened like, do you not have sub server CPU problems? Like, we just repurposed our client, server CPU. We&#8217;re good. And that was like, whoa, okay. So you&#8217;re like really good on the CPU front and AMD is like now their biggest competitor has lost their leader. And I&#8217;m like, this is the time to go all in, right?</p><p><strong>Austin:</strong> Yes, yes, totally. So I don&#8217;t think it&#8217;ll be Lisa Su, but what do you think? What are some other names you&#8217;ve heard floated?</p><p><strong>Vik:</strong> I heard this is totally not going to happen, but I&#8217;m going to say it because it&#8217;s really funny. Somebody thought like Tim Cook might be the CEO.</p><p><strong>Austin:</strong> What what was the rationale?</p><p><strong>Vik:</strong> I don&#8217;t know.</p><p><strong>Austin:</strong> Just cuz he came in behind Steve Jobs and Steve Jobs was the founder.</p><p><strong>Vik:</strong> There&#8217;s no rationale to these kinds of things. It&#8217;s just yeah. Okay. Maybe, I don&#8217;t know. He&#8217;s a leader of a very big company and he&#8217;s a guy who sat why would you get a guy who sat out of the sidelines the whole time for the AI boom? He didn&#8217;t really invest in like CapEx or anything and then now you&#8217;re like, oh, just come and run this company over here that does all AI. Maybe I don&#8217;t know. It doesn&#8217;t seem likely, but</p><p><strong>Austin:</strong> Yes, it&#8217;s a very different business. I agree that he can come in after a founder leaves and he can operationalize and make a ship work as the company grows up from startup founder stuff to like, okay, we&#8217;re like grandparents now. We&#8217;re just a company that&#8217;s been around for a long time. But yes, it&#8217;s a totally different business. I also think he&#8217;s got to be at the point in his life where he doesn&#8217;t want that. I wouldn&#8217;t know that if I was him. I&#8217;d be like, dude, I just retired.</p><p><strong>Vik:</strong> What about Lit Bhutan?</p><p><strong>Austin:</strong> Lit Bhutan, Intel is Intel is not doing too well.</p><p><strong>Vik:</strong> No.</p><p><strong>Austin:</strong> No. Yep. Pass. Uh, great guy. He needs to stay with Intel.</p><h2>The Internal Candidates</h2><p><strong>Austin:</strong> Who else? I think I&#8217;ve seen floated, Madison Huang, Jensen&#8217;s daughter and she&#8217;s a VP, maybe in like, I can&#8217;t remember exactly what org. I maybe what it&#8217;s either she or his son who is like in autonomy. I don&#8217;t remember what the other like automotive autonomy. I don&#8217;t know about who does that.</p><p><strong>Vik:</strong> Okay. Yeah, I don&#8217;t know what their roles are. Yeah.</p><p><strong>Austin:</strong> What&#8217;s the son&#8217;s name?</p><p><strong>Vik:</strong> But you know, we shouldn&#8217;t miss, I&#8217;m going to save the most likely one for the last. But I think we should consider Renee Haas of Arm.</p><p><strong>Austin:</strong> Yeah, so that&#8217;s an interesting one because Renee Haas used to work at Nvidia. And so he has ties to the company, he understands the company. Also, Arm, although yes, it&#8217;s an IP business traditionally, they are getting into the Silicon game. So clearly, Renee has not lost his chops for Silicon and also those are big strategic business model changes. And so clearly, like the guy has a lot of credibility as a CEO. I think he could probably do it.</p><p><strong>Vik:</strong> Today&#8217;s sponsor is Meter. You&#8217;ve got the GPUs and the power. Now, who builds the network and who do you call when it breaks? With Meter, it&#8217;s one call. Tell them your GPU or server count and they design, deploy and run your front-end, storage and management networks. Meter builds its own networking hardware, switches and all, and writes its own network OS. So that call goes to the team who built it. Running a new cloud or clusters in a colo, go to semidoped.com/meter. Links in the show notes.</p><p>You know, I wanted to save what I thought was a probable candidate for last. And if Nvidia promotes anybody internally, it&#8217;ll probably be Ian Buck.</p><p><strong>Austin:</strong> Mhm. Totally.</p><p><strong>Vik:</strong> Agreed. What do you think?</p><p><strong>Austin:</strong> I think two things. One, yes, and&#8212;</p><p><strong>Vik:</strong> Which is a failure of itself, right? Like how can you not have succession planning? Yeah.</p><p><strong>Austin:</strong> 100%. And I&#8217;m sure like, I&#8217;m sure the board&#8217;s like, no, no, we have this on paper, but you can tell like, who else gets to talk at big events? Like it&#8217;s always Jensen. Yes, sometimes it&#8217;s Ian for a little bit here or various executives at these other little things over here. But at the end of the day, it&#8217;s kind of all Jensen all the time. And you don&#8217;t see it quite the same at the other companies. You can tell, even AMD is like grooming other people and they get on stage. And so literally the only person is that it could be in my estimation is Ian Buck. Can you even name any other executives that it could be? I mean, I can maybe think of like some names for like their networking business, but I can&#8217;t name any.</p><p><strong>Vik:</strong> Yeah, I don&#8217;t know. Yeah, Ian Buck is the only name that comes to mind. Right. Right. Yeah, so maybe that&#8217;s the best option. I don&#8217;t know. Let&#8217;s wait and see. They have to announce something. And next earnings call is coming up soon. Next week, I think.</p><p><strong>Austin:</strong> Yeah, next Tuesday. Yeah, we&#8217;ll be at Tomorrow/x. So maybe they&#8217;ll just they could just live stream it because, Jensen was supposed to talk at give a talk, like be one of the keynote speakers at Tomorrow/x. I&#8217;m guessing he won&#8217;t be there now. Maybe they can shuffle things around and just like live stream the earnings call. I think everyone would legitimately sit in and watch it together.</p><p><strong>Vik:</strong> Yeah, or we can just ask Nvidia, hey, send in the new guy or girl.</p><p><strong>Austin:</strong> Yeah, yeah. Right, right. Yeah, totally. Like they won&#8217;t have that figured out. So,</p><p><strong>Vik:</strong> No.</p><p><strong>Austin:</strong> If now the question, and we&#8217;ll just answer this later, we&#8217;ll see what happens. The question is, should it be an internal hire? And I think there&#8217;s lots of good reasons. You know, Ian Buck is known for being kind of the father of CUDA. He came out of, I think Stanford PhD and came over and built all that. There could be arguments for bringing in an external person, but both of them have big time pros and cons. So we&#8217;ll just have to discuss that later. Let&#8217;s see what happens. We&#8217;re not this will be a conversation that goes for a long time. I&#8217;m sure.</p><p><strong>Vik:</strong> It&#8217;s true. It&#8217;s true. And we shouldn&#8217;t dedicate more than half an hour to this. That&#8217;s all we really have to say about this. Let&#8217;s see what happens. We just wanted to bring this to you quickly because it&#8217;s such a hot topic in the AI industry and we want to cover it on Semi Doped just so that we have something to talk about because it&#8217;s such a big event.</p><p><strong>Austin:</strong> Exactly, exactly. And we said we&#8217;d keep it to 15 or 20 minutes and naturally it&#8217;s been about 28 or so. So we&#8217;re going to end it here. Thanks everyone for watching us on YouTube, for subscribing, listening on Spotify. Check out our Semi Doped daily newsletter that we write. Check us out on X. We see it all, so please give us your comments, send us ideas. We want to hear your thoughts. And we&#8217;ll catch you next time.</p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 11th, 2026]]></title><description><![CDATA[Nvidia&#8217;s board brings in a prosecutor, CoreWeave beats and raises, Anthropic sets IPO terms, and TSMC buys in]]></description><link>https://daily.semidoped.com/p/daily-update-november-11th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-11th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Wed, 11 Nov 2026 15:36:51 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/bbb1a083-34b7-4bc1-b9ad-d6437df7aeae_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Wednesday, November 11th, 2026</strong></p><p>Nvidia&#8217;s board hired a former Manhattan US attorney to investigate its own CEO, and a Republican senator asked Jensen Huang to step down. CoreWeave had a much better Tuesday, beating estimates and raising guidance. Anthropic put a price range on its IPO, TSMC agreed to buy shares in it and in OpenAI, and AMD&#8217;s 6th Gen EPYC is now shipping.</p><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="#cfe2f3" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>Intel&#8217;s Earnings Disaster: Bad Luck, Bad Planning, or Both?</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=1eEthO0oDXc">YouTube</a>, <a href="https://podcasts.apple.com/us/podcast/intels-earnings-disaster-bad-luck-bad-planning-or-both/id1866707196?i=1000800862636">Apple Podcasts</a>, or <a href="https://open.spotify.com/episode/1i5fXXh5SqQ0Rg0T7Wla29">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-intels-earnings-disaster">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3><strong>Presented by Crusoe</strong></h3><p><mark data-color="#d9d2e9" style="background-color: rgb(217, 210, 233); color: rgb(0, 0, 0);">Fine-tune and serve open models without managing a single GPU.</mark></p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p><strong>Accounts are free, and get $5 in credit when you sign up.</strong></p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning &#8594;</a></p></div><h3>Nvidia&#8217;s board hires a prosecutor to investigate Huang</h3><p>Nvidia&#8217;s independent directors hired Joon Kim, a former acting US attorney for the Southern District of New York, and his firm Wachtell Lipton to run the investigation into Jensen Huang that the board announced on Sunday. The statement was one sentence long. It said nothing about scope or timing.</p><p>The political pressure grew the same day. Senator Kevin Cramer, a senior Republican on the Senate Banking Committee, said Huang should resign because of the risk of extortion. Separately, Reuters reported that Northwestern Polytechnical University, a military-linked school on the US Entity List, bought two H200 chips in October through a local reseller and rented H200 servers through Chinese computing contracts.</p><p><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Choosing an outside prosecutor and a top Wall Street law firm suggests the board does not want this to look like a review run by Huang&#8217;s longtime friends.</mark> Nvidia reports earnings on November 17, and House committee chairs have asked Huang for a closed-door briefing.</p><p><a href="https://nvidianews.nvidia.com/news/statement-from-the-independent-members-of-the-nvidia-board">Source: Nvidia</a> &#183; <a href="https://www.cnbc.com/2026/11/10/nvidia-wachtell-lipton-huang-affair.html">CNBC</a> &#183; <a href="https://www.politico.com/news/2026/11/10/senate-banking-republican-nvidia-ceo-resign-01160543">Politico</a> &#183; <a href="https://www.straitstimes.com/asia/pla-linked-chinese-university-bought-nvidia-h200-chips-despite-us-blacklist-records-show">Reuters via Straits Times (H200)</a></p><blockquote><p><strong>Vik: </strong><em>The Jensen saga continues. People seem to think this will eventually blow over for Jensen, and we will go back to focusing on Nvidia chips instead of their CEO. We&#8217;ll continue to report on the ongoings till then because the CEO of a company which is 8% of the S&amp;P500 holds an important position in the world.</em></p><p><strong>Austin: </strong><em>I&#8217;ve always said Nvidia needed a better succession plan. Even though I don&#8217;t foresee succession anytime soon, I&#8217;m sure the plan is getting much more attention from the board. News stuff aside, looking forward to the earnings call next week. </em><strong>   </strong></p></blockquote><h3>CoreWeave beats, raises, and gets a reprieve</h3><p>CoreWeave&#8217;s third-quarter revenue rose 168% to $3.66 billion, ahead of the $3.55 billion analysts expected. It raised 2026 revenue guidance to $13.0 billion to $13.3 billion and lifted adjusted operating income guidance 26% at the midpoint. Backlog reached $132.5 billion, up from $104 billion in June.</p><p><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">CoreWeave says it expects about $9 billion in customer prepayments and deposits in the next three months, which helps answer the biggest question about the company: how it pays for the buildout.</mark> It still lost $709 million in the quarter, and net interest expense was $846 million. The stock rose about 13% in premarket trading and lifted other neoclouds with it.</p><p>Capex guidance is now $37 billion to $39 billion for 2026. The CFO said that if CoreWeave only builds for contracts already signed, 2027 capex would stay about flat while revenue grows a lot.</p><p><a href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Third-Quarter-2026-Results/default.aspx">Source: CoreWeave</a> &#183; <a href="https://www.cnbc.com/2026/11/10/coreweave-crwv-q3-earnings-report-2026.html">CNBC</a></p><blockquote><p><strong>Vik: </strong><em>CoreWeave&#8217;s beat and raise has quelled investor concerns about their costs, and stock price rose. But CoreWeave bond yields have climbed above 16%, putting them still firmly in the junk bond category with high default risk.</em></p></blockquote><h3>Anthropic sets its IPO terms</h3><p>Anthropic&#8217;s amended S-1 sets an IPO price range of $70 to $78 a share for 825 million shares, which would raise up to about $64 billion. Reuters says the range values the company at up to $1.4 trillion. The stock will trade on Nasdaq as ANT.</p><p>The filing is also a compute ledger. Anthropic now lists about $900.9 billion of committed spending, almost all of it compute and infrastructure. Its Microsoft commitment rose to at least $47.2 billion through mid-2033. <mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Its largest Lambda deal, about $34.9 billion for Nvidia GPU capacity, can now be cancelled by either side because Lambda missed a deadline to secure the facility lease.</mark> Anthropic says it does not plan to cancel.</p><p><a href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026072484/anthropicpbc-sx1ax2.htm">Source: Anthropic S-1/A</a> &#183; <a href="https://www.anthropic.com/news/roadshow-launch">Anthropic</a> &#183; <a href="https://mlq.ai/news/anthropics-amended-s-1-says-its-35-billion-lambda-compute-deal-can-now-be-cancelled-and-raises-its-microsoft-commitment/">MLQ.ai</a></p><blockquote><p><strong>Vik: </strong><em>$900B of committed spending on infrastructure but the increasing pressure against building datacenters might pressure actually securing the leases. What happens if large amounts of the contracted capacity gets cancelled by either party due to unavailability of facilites?</em></p></blockquote><h3>TSMC buys into Anthropic and OpenAI</h3><p>TSMC&#8217;s board approved up to $500 million for Anthropic shares at the IPO price and up to $500 million for OpenAI preferred stock. It also set up a separate investment company with up to $500 million to back semiconductor and related ecosystem companies.</p><p>Neither AI lab buys chips from TSMC directly. Their demand reaches TSMC through Nvidia, AMD, Google and Broadcom. <mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The stakes give TSMC a piece of the companies at the top of its demand chain, with no supply terms attached.</mark></p><p>The same meeting approved about $40.9 billion in new capital appropriations for advanced nodes, advanced packaging and fab construction, and raised the quarterly dividend to NT$9 from NT$7. TSMC co-COO Kevin Zhang told Nikkei Asia that demand for its leading-node silicon is &#8220;insatiable.&#8221;</p><p><a href="https://pr.tsmc.com/english/news/3352">Source: TSMC</a> &#183; <a href="https://www.investing.com/news/stock-market-news/tsmc-to-invest-up-to-1-billion-in-anthropic-openai-5015908">Reuters</a> &#183; <a href="https://focustaiwan.tw/business/202611110009">Focus Taiwan</a></p><blockquote><p><strong>Vik: </strong><em>This is a rare move for TSMC. Will they buy into more AI stonks through their investment company? Buying into the model layer makes sense for them, but I wonder if they will buy in the supply chain.</em></p><p><strong>Austin: </strong><em>Large investors get time with management. Could be a savvy way to get facetime to pressure test actual end demand. I know they already talk, but it&#8217;s different when it&#8217;s a large shareholder.</em><strong>   </strong></p></blockquote><h3>AMD&#8217;s 6th Gen EPYC is shipping</h3><p>AMD&#8217;s 6th Gen EPYC processors, built on Zen 6, are now shipping in systems from server makers. Microsoft Azure and Oracle Cloud instances are coming soon, and Meta is the lead customer as it prepares to deploy them at scale. The top chip, the EPYC 9996, has 256 cores at 600 watts and lists at $14,904.</p><p><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">AMD is pitching the CPU as its own AI product, sold separately for GPU host nodes, agent sandboxes and general-purpose servers.</mark> AMD claims up to 3.7 times the performance of an Nvidia Vera system and 2.2 times an Intel Xeon 6980P system on a standard integer benchmark, though the Vera number is AMD&#8217;s own estimate. The chips are built on TSMC&#8217;s 2nm process, which AMD started ramping in May.</p><p><a href="https://ir.amd.com/news-events/press-releases/detail/1305/amd-showcases-6th-gen-amd-epyc-processors-to-lead-next-era-of-ai-and-enterprise-computing">Source: AMD</a> &#183; <a href="https://ir.amd.com/news-events/press-releases/detail/1287/amd-announces-production-ramp-of-next-generation-amd-epyc-processor-venice-on-tsmc-2nm-process-technology">AMD (2nm ramp)</a> &#183; <a href="https://www.servethehome.com/amd-epyc-9006-results-are-here-and-amd-delivers-a-knockout-part-deux/">ServeTheHome</a> &#183; <a href="https://www.trendforce.com/news/2026/11/11/news-amd-epyc-venice-cpus-begin-shipping-on-tsmc-2nm-verano-reportedly-to-adopt-n2p-with-a14-planned-for-2028/">TrendForce</a></p><blockquote><p><strong>Vik: </strong><em>I&#8217;ve held the view for a long time that AMD&#8217;s CPU SKUs are brilliantly suited to various parts of agentic AI requirements. The fungibility of chiplet supply between Ryzen and Epyc gives them the chance to actually supply into the CPU demand. High hopes for AMD!</em></p></blockquote><div class="callout-block" data-callout="true"><h4><strong>Supported by Meter</strong></h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter-daily">See how Meter builds data center networks</a></p></div><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>SpaceX Kessel rack</strong> <strong>(a mini version of Kyber)</strong> &#8212; SemiAnalysis says SpaceX is designing Kessel, a custom Vera Rubin NVL72 rack that swaps NVLink cable cartridges for a passive 84-layer PCB midplane. SpaceX would buy chip packages from Nvidia and source the rest itself, skipping Nvidia&#8217;s compute boards (<a href="https://x.com/SemiAnalysis_/status/2120351564720136389">SemiAnalysis on X</a>).</p></li><li><p><strong>Nscale IPO</strong> &#8212; The Telegraph reports that Nvidia-backed Nscale is considering pushing its planned $35 billion New York listing into 2027 after the AI infrastructure selloff (<a href="https://www.investing.com/news/stock-market-news/ai-cloud-firm-nscale-considers-delaying-35-billion-listing-the-telegraph-reports-5016636">Reuters via Investing.com</a>).</p></li></ul><h2>Sector Watch</h2><p><strong>Packaging and servers</strong></p><ul><li><p><strong>ASE</strong> &#8212; The world&#8217;s largest chip packager posted its fourth straight monthly revenue record in October at NT$88.6 billion, up 47% from a year ago, with packaging and test up 58% (<a href="https://www.sec.gov/Archives/edgar/data/1122411/000095010326017186/dp256592_6k.htm">Source: ASE 6-K</a>).</p></li><li><p><strong>Inventec</strong> &#8212; The server maker expects GPU and ASIC server shipments to grow at a high double-digit rate in 2027, but says shortages of memory, CPUs, hard drives and passives could hold back fourth-quarter shipments (<a href="https://www.taipeitimes.com/News/biz/archives/2026/11/11/2003867507">Source: Taipei Times</a>).</p></li></ul><p><strong>China chips</strong></p><ul><li><p><strong>Alibaba</strong> &#8212; Alibaba said its chip design unit T-Head has filed for a Hong Kong IPO after the exchange cleared the spinoff; size and timing are not set (<a href="https://data.alibabagroup.com/ir_filings/HKEX/09988/en/202611102026111001061/2026111001061.pdf">Source: Alibaba filing</a>).</p></li></ul><p><strong>Demand signals</strong></p><ul><li><p><strong>OpenAI</strong> &#8212; OpenAI&#8217;s third-quarter revenue rose about 50% to roughly $10 billion while its operating loss widened to $10.9 billion, according to investor documents seen by the Wall Street Journal (<a href="https://www.investing.com/news/stock-market-news/openai-posts-steep-losses-even-as-quarterly-revenue-rises-to-10-billion-wsj-reports-5017353">Source: Reuters</a>).</p></li><li><p><strong>South Korea</strong> &#8212; Korean chip exports more than quadrupled from a year ago in the first ten days of November to $16.6 billion, half of all exports (<a href="https://en.yna.co.kr/view/AEN20261111002651320">Source: Yonhap</a>).</p></li></ul><p><mark data-color="#cfe2f3" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark>:</p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/amd-q3-fy26-annotated">AMD Q3 FY26, Annotated</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/why-ai-data-center-constraints-make">Why AI Data Center Constraints Make Efficiency More Valuable</a></p></li></ul><p><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better! </mark><a href="http://semidoped.com/poll"><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Give us your feedback</mark></a></p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 10th, 2026]]></title><description><![CDATA[Nvidia&#8217;s Megaspeed problem deepens, Google adds two TPUs with AMD inside, and TSMC posts another record month]]></description><link>https://daily.semidoped.com/p/daily-update-november-10th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-10th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Tue, 10 Nov 2026 15:20:35 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/3886edea-3cf1-4621-8fea-8aa22e93e40e_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Tuesday, November 10th, 2026</strong></p><p>Nvidia&#8217;s week got worse, with reports that its top lawyer contacted the officials investigating Megaspeed and a shareholder suit filed five minutes after Monday&#8217;s close. Google picked the same day to launch two more TPUs, one of them running on AMD host CPUs. TSMC posted another record month, Infineon expects its data center power sales to nearly double, and Rapidus is reportedly behind on 2nm.</p><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>Intel&#8217;s Earnings Disaster: Bad Luck, Bad Planning, or Both?</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=1eEthO0oDXc">YouTube</a>, <a href="https://podcasts.apple.com/us/podcast/intels-earnings-disaster-bad-luck-bad-planning-or-both/id1866707196?i=1000800862636">Apple Podcasts</a>, or <a href="https://open.spotify.com/episode/1i5fXXh5SqQ0Rg0T7Wla29">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-intels-earnings-disaster">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p><mark data-color="#d9d2e9" style="background-color: rgb(217, 210, 233); color: rgb(0, 0, 0);">Fine-tune and serve open models without managing a single GPU.</mark></p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p><strong>Accounts are free, and get $5 in credit when you sign up.</strong></p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning &#8594;</a></p></div><h3>Nvidia&#8217;s lawyer contacted the Megaspeed investigators</h3><p>Bloomberg and the New York Times reported Monday that Nvidia general counsel Tim Teter called Commerce Under Secretary Jeffrey Kessler in late summer 2025 to ask about the Megaspeed smuggling probe. According to the reports, Teter then emailed Bureau of Industry and Security investigators directly, saying Kessler had given him permission. One of them turned down his request for a meeting. Kessler told staff to handle the case by the book. The Justice Department later asked Commerce to pause its work, and the case is now a criminal investigation at DOJ.</p><p>Nvidia said Teter did not know about Jensen Huang&#8217;s affair when he made the call, and that it regularly consults officials on export questions. Bloomberg also reports that trade records show Nvidia&#8217;s count of chips sold to Megaspeed through hardware partners came up about 50,000 units short. Nvidia says it later inspected more warehouses holding the missing hardware. Megaspeed is still listed as an Nvidia Cloud Partner.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The story has moved from a CEO&#8217;s private life to whether Nvidia leaned on the export-control system, and that is the version Congress cares about.</mark> House Democratic leaders are asking for answers and tighter China chip rules. A proposed shareholder class action was filed in San Jose five minutes after Monday&#8217;s close, when Nvidia had fallen 4.3%, cutting its market value by about $190 billion. </p><p><a href="https://news.bloomberglaw.com/artificial-intelligence/nvidia-intervened-in-us-probe-of-megaspeed-chip-smuggling">Source: Bloomberg</a> &#183; <a href="https://www.nbcnews.com/tech/tech-news/nvidia-investigation-justice-fbi-commerce-jensen-huang-alice-china-rcna608397">NBC News</a> &#183; <a href="https://www.investing.com/news/stock-market-news/nvidia-ceo-huang-sued-over-disclosures-related-to-suspected-chinese-spy-5013619">Reuters (lawsuit)</a></p><h3>Google adds two TPUs, and AMD gets the host socket</h3><p>Google added TPU 8tx and TPU 8i Max to its eighth-generation lineup, making four new TPU systems this year. TPU 8tx uses the same silicon as TPU 8t, adds hardware confidential computing for regulated customers, and runs on 5th Gen AMD EPYC host CPUs. TPU 8i Max puts two TPU 8i chips in one package, which doubles compute, SRAM and HBM bandwidth in the same rack footprint. Google didn&#8217;t give pricing or availability dates.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Google is now splitting one TPU generation into four systems, each tuned for a different job, from training to secure inference to long-context agents.</mark> The two-chip 8i Max is also a packaging and HBM story: more silicon and more memory bandwidth per rack without a new die.</p><p>AMD says EPYC is now the host CPU at more than 10 hyperscalers and frontier labs, across GPU and custom-ASIC systems. AMD shares rose 13% on Monday.</p><p><a href="https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/">Source: Google</a> &#183; <a href="https://ir.amd.com/news-events/press-releases/detail/1304/amd-epyc-expands-its-leadership-at-the-foundation-of-ai">AMD</a> &#183; <a href="https://www.servethehome.com/google-tpu-8tx-8i-max-amd-epyc-turin/">ServeTheHome</a></p><blockquote><p><strong>Vik: </strong><em>Two training chips, and two inference chips all from the 8th generation TPU. It tells you how much workloads vary, and why specialized silicon variants are required to extract the best possible performance.</em></p><p><strong>Austin: </strong><em>Vik, maybe we should discuss this on the pod, I think it got overlooked given all the other news going on this week.</em></p></blockquote><h3>TSMC posts another record month</h3><p>TSMC&#8217;s October revenue came in at NT$586.89 billion, up 59.7% from a year ago and 14.7% from September. That beat August&#8217;s record of NT$514.81 billion. Revenue for the first ten months is up 43.3%, close to the company&#8217;s guidance of about 44% growth for the full year.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">One month in, TSMC is already running above the monthly pace it needs to hit the top of its fourth-quarter guidance.</mark> TSMC guided to about NT$1.66 trillion to NT$1.73 trillion for the quarter, which works out to roughly NT$575.6 billion a month at the high end.</p><p>The broader Taiwan numbers match. Taiwan&#8217;s chip exports rose 63% in October to $31.17 billion, and total exports hit a record $96.76 billion.</p><p><a href="https://pr.tsmc.com/english/news/3349">Source: TSMC</a> &#183; <a href="https://www.taiwannews.com.tw/news/6474406">Taiwan News</a> &#183; <a href="https://www.mof.gov.tw/Eng/singlehtml/f48d641f159a4866b1d31c0916fbcc71?cntId=25f99a5315f04f15a626000b148eb28d">Taiwan Ministry of Finance</a></p><blockquote><p><strong>Vik: </strong><em>With Intel&#8217;s ongoing struggles, the world has no alternative to TSMC. Demand for chips is still high and TSMC remains key supplier.</em></p><p><strong>Austin: </strong><em>Up 15% month over month!!!</em></p></blockquote><h3>Infineon expects data center power sales to nearly double</h3><p>Infineon guided fiscal 2027 revenue to about &#8364;20 billion, up more than 20% from &#8364;16.35 billion in fiscal 2026. It expects data center power supply revenue of close to &#8364;4 billion, up from &#8364;2.2 billion, which includes about &#8364;500 million from conventional data centers. AI power revenue alone was &#8364;1.7 billion last year, more than double the year before, and Infineon says it still couldn&#8217;t meet all customer requests.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Infineon says customers have signed or are negotiating multi-year supply deals and capacity reservations worth more than &#8364;10 billion.</mark> Power chips for AI racks now look like a capacity-reservation business too.</p><p>Not everything beat. Free cash flow guidance of about &#8364;1.8 billion came in well below the &#8364;2.26 billion analysts expected, because Infineon plans to spend &#8364;3.0 billion to &#8364;3.5 billion on fabs in Dresden, Kulim and Villach. The shares still rose about 9% in early trading.</p><p><a href="https://www.infineon.com/press-release/2026/infpr202611-025">Source: Infineon</a> &#183; <a href="https://www.streetinsider.com/Reuters/Chipmaker+Infineon+forecasts+strong+2027+growth+on+data-centre+demand/27317179.html">Reuters</a> &#183; <a href="https://uk.investing.com/news/earnings/infineon-fy27-revenue-outlook-tops-forecasts-on-ai-demand-cash-flow-disappoints-4970447">Investing.com</a></p><blockquote><p><strong>Vik: </strong><em>Power semis is very important, and I have been saying that it is inevitable that it will become the next most dominant factor in the AI buildout. I have several articles on my substack on why: <a href="https://substack.com/@vikramskr/note/p-224091959?utm_source=notes-share-action&amp;r=222kot">Datacenter constraints and AI power</a>, <a href="https://open.substack.com/pub/viksnewsletter/p/power-delivery-as-the-next-physics-wall?r=222kot&amp;utm_campaign=post-expanded-share&amp;utm_medium=web">Power as next physics wall</a>, <a href="https://open.substack.com/pub/viksnewsletter/p/800vdc-mapping-the-power-semi-space?r=222kot&amp;utm_campaign=post-expanded-share&amp;utm_medium=web">800V DC sideca</a>r, <a href="https://open.substack.com/pub/viksnewsletter/p/understanding-hvdc-architectures-ai-megafactories?r=222kot&amp;utm_campaign=post-expanded-share&amp;utm_medium=web">DC power architectures</a>.</em></p></blockquote><h3>Rapidus is reportedly behind on 2nm</h3><p>Bloomberg reports that Rapidus has not yet shown working SRAM cells, a milestone it was supposed to hit by this summer, which puts it at least several months behind schedule. The report also says the company is dealing with dust and particle contamination in its Hokkaido cleanroom. Rapidus says it is still fully committed to starting mass production in the second half of fiscal 2027, which ends in March 2028.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">SRAM is a basic building block of nearly every leading-edge chip, so missing it now leaves Rapidus little room to make up time before its 2028 target.</mark> Japan has committed more than &#165;2 trillion to Rapidus, and a long delay could put that support and its planned IPO at risk. Customers were already cautious about leaving TSMC.</p><p><a href="https://www.japantimes.co.jp/business/2026/11/10/companies/rapidus-behind-schedule/">Source: Bloomberg via Japan Times</a> &#183; <a href="https://www.trendforce.com/news/2026/11/10/news-rapidus-reportedly-months-behind-on-2nm-sram-development-march-2028-mass-production-target-at-risk/">TrendForce</a></p><blockquote><p><strong>Vik: </strong><em>Let&#8217;s see that 6T SRAM cell working and the resulting density. Very important first step for technology.</em></p><p><strong>Austin: </strong><em>I&#8217;ve been skeptical anyone could take 2nm technology and ramp it to production scale, from scratch. Seems it is, in fact, quite hard. </em></p></blockquote><div class="callout-block" data-callout="true"><h4>Supported by Meter</h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter-daily">See how Meter builds data center networks</a></p></div><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>Nvidia orders</strong> &#8212; Taiwan supply chain sources told Commercial Times that Nvidia may slow chip orders at TSMC and ASE because of the Huang controversy, which would free up capacity for AMD and ASIC designers like Broadcom, MediaTek and Alchip. Server chassis, rack and cooling suppliers say their orders and shipments haven&#8217;t changed (<a href="https://www.trendforce.com/news/2026/11/10/news-nvidia-order-cuts-reportedly-loom-amid-huang-controversy-amd-asic-rivals-may-gain-tsmc-capacity/">TrendForce</a>).</p></li><li><p><strong>Anthropic IPO</strong> &#8212; Anthropic sent banks a filing showing a $69.3 billion IPO at $70 to $80 a share, plus a $26 billion private placement from Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC, then told them the pricing details were inaccurate and an updated filing would follow (<a href="https://www.investing.com/news/stock-market-news/exclusiveanthropic-aims-to-raise-up-to-693-billion-in-blockbuster-ipo-5013861">Reuters</a>).</p></li><li><p><strong>Moonshot B300s</strong> &#8212; Documents seen by Bloomberg show Moonshot AI renting 32 Super Micro servers with Nvidia B300 GPUs inside China in August, valued at 432 million yuan. Moonshot disputes the documents (<a href="https://www.investing.com/news/stock-market-news/moonshot-ai-rents-nvidias-top-chips-in-china-bloomberg-reports-5014089">Bloomberg via Investing.com</a>).</p></li><li><p><strong>Samsung 3D Cube-H</strong> &#8212; Samsung aims to finish its hybrid-bonded logic stacking technology, with about a 6&#181;m pitch, and its first design kit by early 2027, aimed at AI chips such as stacked GPUs or logic on HBM (<a href="https://zdnet.co.kr/view/?no=20261110104822">ZDNet Korea</a>).</p></li></ul><h2>Sector Watch</h2><p><strong>Capital and AI customers</strong></p><ul><li><p><strong>SoftBank</strong> &#8212; SoftBank swung to a &#165;1.1 trillion ($7 billion) quarterly loss, its first in nearly two years, after a $3.7 billion writedown on advance payments for power equipment for US data center projects, including Pike County, Ohio, whose plans were later revised (<a href="https://www.channelnewsasia.com/business/softbank-books-7-billion-quarterly-loss-us-data-centre-hit-6538511">Source: Reuters via CNA</a>).</p></li></ul><p><strong>Equipment and EDA</strong></p><ul><li><p><strong>Onto Innovation</strong> &#8212; The inspection and metrology toolmaker posted record third-quarter revenue of $409.7 million, up 88% on AI chip demand, and guided fourth-quarter revenue to $430 million to $450 million (<a href="https://investors.ontoinnovation.com/news/news-details/2026/Onto-Innovation-Reports-2026-Third-Quarter-Results/default.aspx">Source: Onto Innovation</a>).</p></li><li><p><strong>Siemens</strong> &#8212; Siemens agreed to buy Lauterbach, which makes embedded-software debug and trace tools, to pair with its chip design and hardware-assisted verification tools. The price wasn&#8217;t disclosed and closing is expected in the first half of 2027 (<a href="https://www.prnewswire.com/news-releases/siemens-to-expand-portfolio-for-semiconductor-design-and-verification-with-the-planned-acquisition-of-lauterbach-302934669.html">Source: Siemens</a>).</p></li></ul><p><strong>Substrates</strong></p><ul><li><p><strong>Unimicron and Kinsus</strong> &#8212; Both Taiwanese substrate makers posted their seventh straight monthly revenue record in October on AI demand, with Unimicron up 68% from a year ago and Kinsus up 49% (<a href="https://www.taiwannews.com.tw/news/6474270">Source: Taiwan News</a>).</p></li></ul><p><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better! </mark><a href="http://semidoped.com/poll"><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Give us your feedback</mark></a></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark>:</p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/amd-q3-fy26-annotated">AMD Q3 FY26, Annotated</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/why-ai-data-center-constraints-make">Why AI Data Center Constraints Make Efficiency More Valuable</a></p></li></ul>]]></content:encoded></item><item><title><![CDATA[🎙️ NEW EPISODE: Intel's Earnings Disaster: Bad Luck, Bad Planning, or Both?]]></title><description><![CDATA[Watch now (57 mins) | Intel's 35% stock drop, the three reasons for its weak guidance, chiplet interdependencies across Intel 3 and Intel 7, the Fab 28 explosion, and more.]]></description><link>https://daily.semidoped.com/p/new-episode-intels-earnings-disaster</link><guid isPermaLink="false">https://daily.semidoped.com/p/new-episode-intels-earnings-disaster</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Mon, 09 Nov 2026 22:25:32 GMT</pubDate><enclosure url="https://substack-video.s3.amazonaws.com/video_upload/post/224211134/fe6ce800-d5c9-4bf4-a855-a2113968b6e6/transcoded-1794258247.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Intel just had its worst day on the stock market in 55 years, with shares falling nearly 30%. But here's the strange part: Intel actually reported a strong quarter. Revenue grew, its data center business surged, and the company beat expectations. So what went wrong?</p><p>In this episode of Semi Doped, Vik Sekar and Austin Lyons dig into Intel's Q3 2026 earnings and the problems behind its disappointing outlook.</p><p>Three issues are at the center of Intel's troubles: a weakening PC market driven by rising memory prices, a shortage of advanced packaging substrates that limits server CPU production, and a tragic factory accident in Israel that disrupted chiplet supply.</p><p>But how much of this was bad luck, and how much was poor planning?</p><p>We debate whether Intel should have anticipated the surge in demand for server CPUs from agentic AI, whether its manufacturing strategy can respond quickly enough, and what these problems mean for its ambitions to become a leading-edge foundry alternative to TSMC.</p><p>We also step back to examine what Intel's experience tells us about the broader semiconductor market.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p><strong>Fine-tune and serve open models without managing a single GPU.</strong></p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere.</p><p>Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p>Accounts are free, and <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">get $5 in credit</mark> when you sign up.</p><p><strong><a href="https://semidoped.com/crusoe">Try Serverless Fine-Tuning &#8594;</a></strong></p></div><p><strong>Topics covered:</strong></p><ul><li><p>Why Intel's stock crashed despite strong earnings</p></li><li><p>How the AI memory boom is hurting the PC market</p></li><li><p>Why substrate shortages are constraining server CPU production</p></li><li><p>The consequences of Intel's Fab 28 accident</p></li><li><p>Whether Intel misjudged demand for agentic AI CPUs</p></li><li><p>Intel 18A, 14A, and the long-term foundry opportunity</p></li><li><p>Why strong demand doesn't necessarily translate into revenue</p></li><li><p>Why investors are increasingly demanding returns from the AI infrastructure buildout</p></li></ul><p><strong>The bigger question:</strong> Is Intel facing a temporary supply-chain problem, or are its troubles a warning sign for the broader AI semiconductor industry?</p><div class="callout-block" data-callout="true"><p><strong>This episode is supported by Meter.</strong></p><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks? Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS. One call reaches the team that designed the hardware and wrote the software.</p><p><strong><a href="https://semidoped.com/meter">See how Meter builds datacenter networks &#8594;</a></strong></p></div><p><em>This podcast is lightly edited for clarity.</em></p><h2>The 35% Drop</h2><p><strong>Austin:</strong> All right, hello everyone. Welcome to another Semi Doped episode. I&#8217;m Austin Lyons of Chipstrat. I write a newsletter for institutional investors and technology executives.</p><p><strong>Vik:</strong> Hey, I&#8217;m Vik Sekar and I&#8217;m the founder of semi-exponent, where I work with institutional clients on semis and investing, and I also write a substack called Vik&#8217;s newsletter.</p><p><strong>Austin:</strong> All right, Vik. So today we are talking Intel. Obviously, everyone&#8217;s talking Intel. I was just looking up the stock. So we&#8217;re recording this Tuesday, November 3rd. Everyone&#8217;s voting today or everyone already voted. But in the United States, it&#8217;s a big day for voting. But here we are. We&#8217;re talking Intel. Right now I&#8217;m looking at the stock. It&#8217;s down 35% since the earnings call. So that is why we are talking Intel. You and I have talked a lot about it, have a lot of thoughts. We&#8217;re going to get into that today. So, I&#8217;ll throw it over to you to set the stage.</p><p><strong>Vik:</strong> Yeah, so Intel was quite the recovery story that everybody expected because they had a leadership change sometime back. Lip became CEO and everybody was optimistic that they&#8217;re going to make a comeback in the foundry business. And that is very important for the United States because Intel is a leading edge foundry and the world needs an alternative to TSMC. And especially if it&#8217;s United States based, it&#8217;s a perfect, win-win situation for US companies. So that was the expectation. And the stock ran up recently from like $40 all the way to, I don&#8217;t know, $120, $130, I think. So there was a lot of expectation and stuff. And things were looking okay for quite some time.</p><p>Intel has their substrate business, where they make like advanced packaging, right, through EMIB, which is embedded multi-die interconnect bridge, which is the way you connect a couple of chiplets together is through this technology called EMIB. And the world needs a alternative to TSMC&#8217;s CoWoS as well, like chip on wafer on substrate. So Intel was perfectly poised for advanced packaging. And then a couple of quarters ago, they were also really well poised for CPUs because agentic AI became a thing. And who else could make CPUs than Intel? They&#8217;ve done it for like 50 plus years. So it&#8217;s a perfect position to be in and expectations were running high.</p><p>After the earnings call, even though the numbers were pretty good, the data center segment is growing really, really well. But then, what happened was that their guide for the future quarters to come was kind of dim and the tone was subdued and there were good reasons why Intel ended up in this situation, which we will get into in quite some depth, but I just want to list out the three reasons that Intel&#8217;s guide didn&#8217;t pan out and the investors were disappointed.</p><p><strong>Austin:</strong> Yeah, so and then we&#8217;ll debate them.</p><h2>Reason #1: The PC Market Is Weak</h2><p><strong>Vik:</strong> Yeah, we&#8217;ll go through each one and we&#8217;ll see what that actually means and how we got here. But the first one is that the PC segment has become very weak. And that&#8217;s primarily because of the memory price increases. Data centers are stocking up all the memory. There is no memory left for regular PC consumers to use anymore. So much so that I don&#8217;t know Austin if you followed this, but Sam Altman, there was this like, he said on this cheeky pint podcast that people are wasting RAM on their home PCs. It&#8217;s idle 95% of the time, it&#8217;s probably better if we had it for data centers or something like that.</p><p><strong>Austin:</strong> I saw that. I was shaking my head because I was like, dude, it&#8217;s an interesting argument to ask where should the incremental RAM go, like net new, but to be like, ah yeah, we&#8217;re just wasting all this RAM in our laptops, in our desktops, in our phones. It was like the dumbest thing I&#8217;ve ever heard.</p><p><strong>Vik:</strong> And I&#8217;m sorry, this is a little of a tangent, but it&#8217;s totally worth it. There&#8217;s all these memes around this about how, ways to trap Sam Altman. There&#8217;s like people have put like DRAM chips like candy and then at the end of it, there&#8217;s like a box propped up with a stick, so you can trap Sam Altman when he comes to pick up DRAM chips. So many funny things happening. Anyway, apart from all the fun and jokes, it&#8217;s just that there is not enough DRAM to go around for computers. Anyway. So because of that, memory is very expensive and so basically Intel just passed on all the costs to the customer and what ended up happening was like the demand petered out. Nobody wants to buy PCs. They&#8217;re like, I&#8217;m happy with what I have. If there&#8217;s enterprise or consumers like us, we are like, yeah, it&#8217;s fine. My PC&#8217;s fine. I can&#8217;t buy computers at a premium right now. I&#8217;ll just stick to what I have. So, that segment has been slowing down.</p><p><strong>Austin:</strong> Correct. It&#8217;s a very elastic market as everyone knows, especially consumers, of course. If you were maybe thinking about a new computer, obviously if it&#8217;s $500 more or $1,000 more, you&#8217;re totally not going to buy it. I think enterprise is surprisingly elastic as well. So yes, you&#8217;ve had your laptop for three years and it&#8217;s slow and we want you to be as productive as possible, but if it&#8217;s going to cost an extra $500, $1,000 a person times all the people in your company who are up for a refresh, they&#8217;re just going to push it back. So, it&#8217;s very elastic.</p><h2>A Failure to Plan</h2><p><strong>Austin:</strong> This episode of Semi Doped is presented by Crusoe, where you can fine-tune and serve open models without managing a single GPU. When we sat down with Crusoe&#8217;s serverless fine-tuning product team, we asked the question every AI user wonders, what will you do with my data? Crusoe has a simple answer. They don&#8217;t touch it or see it. Your data is yours to fine-tune models with or simply store for future use. Their fine-tuning is designed so that both your training data and the fine-tune model itself remain your property. Your training data is encrypted, so even Crusoe&#8217;s cloud engineering team can&#8217;t see it. You can deploy it on Crusoe Cloud in one click or download the checkpoints and run it anywhere you&#8217;d like. The fine-tuning process is fully managed, so you upload a data set, pick an open model, and Crusoe takes care of the GPUs, orchestration, and all the infrastructure management. You pay per token processed rather than per GPU hour, so run a few versions, compare them, and download the one that actually works best for your use case. Create a new Crusoe account for free and get $5 in credit to start testing today. This part is super important, guys. Support Semi Doped by going to semidoped.com/crusoe, click on that link in the show notes, type it in yourself, and you&#8217;ll get directed to a sign-in page. Sign in with Google or create a new account, and Crusoe will immediately create that account for you with $5 in it. Boom. Again, just click that link in the show notes, sign in, $5 for you, and you support Semi Doped. So, thank you for your support. Go check out Crusoe, try to fine-tune your own data. Again, the link to try serverless fine-tuning is in the show notes. So, hit that now, click through, and make sure you claim that $5, give it a try, and support us. Back to the show.</p><p><strong>Austin:</strong> Shouldn&#8217;t Intel have seen this coming? How long have we known that memory prices are skyrocketing due to AI and that obviously the same three memory companies that have memory wafers that make DRAM that get used for HBM, stack DRAM, and also for DRAM, it&#8217;s the same wafers that are making DRAM for consumers and shouldn&#8217;t we have all foreseen this? Haven&#8217;t we all foreseen this for a long time? I&#8217;m specifically, if you look at Intel&#8217;s quarterly earnings calls, a year ago, October 23rd, 2025, there was no memory warning back then. I think they just said something like, hey, we&#8217;re prudently preparing for another year of strong demand in 2026.</p><p>So that, one question is like, why did they not call it out then? In January 22 of 2026, we got the first warning in the earnings call where Intel said, industry-wide supply for key components like DRAM, NAND, and substrates has come under increasing pressure due to intense demand to support the rapid expansion of AI infrastructure. So benefit of the doubt, okay, January of 2026, you saw the substrate thing coming, you saw the DRAM thing coming. So how did we get to where we are now? And I&#8217;ll let you continue on of, oh, and I think maybe I buried the lead here a little bit, but we&#8217;ve produced a lot of laptop and desktop chips, client chips for PCs that aren&#8217;t going to sell and so our OEMs, we&#8217;re kind of...</p><p><strong>Vik:</strong> Yeah, so that&#8217;s the whole problem. I was reviewing the earnings calls from the last one to what they said now, and I was trying to understand where things went wrong, and I&#8217;m writing an article on the substack that goes into some of this stuff in more detail. But essentially, the story is, I didn&#8217;t go too far back. I just wanted to see like everything was reasonably upbeat last earnings call, but what&#8217;s happened now? So the whole story is that if you see the last earnings call, Q2, what we just had was Q3. The Q2 earnings call, an analyst asked the CFO, basically, what do you see happening here? Because you say your client segment actually grew in revenue. But did that revenue grow because you actually sold more parts or did you increase the chip prices and then make more money? So the CFO replied and said, yeah, we sold the same number of chips, but we increased the price of each chip.</p><p>And then the analyst goes, hey, what are you going to do for the second half of this year? This is already slowing down, we can see this in Q2. What&#8217;s your plan for Q3 and Q4? You think, can you sell it at this price essentially? And the CFO&#8217;s response was like, yeah, yeah, yeah. We know the demand is going to be soft. So, there&#8217;s one point to take note of. In Q2, definitely they knew that demand is going to be soft. It&#8217;s in the earnings call, you can go read it. But the explanation for why things are going to be A-OK is because they were like, oh, it&#8217;s good, it&#8217;s a good thing that there&#8217;s no demand for client side chips because we don&#8217;t have enough supply anyway. We are not able to make chips fast enough anyway. So it&#8217;s a good thing, quote unquote, good thing that client is weakening. Now we can take all those wafers and build server CPUs with them, which is amazing because everybody wants server CPUs for agentic AI. So it was this actually framed as a don&#8217;t worry about it if the client is weak, we&#8217;ll sell more server CPUs and make up the shortfall. Everything&#8217;s good. But that actually didn&#8217;t happen.</p><h2>Reason #2: Substrates and Chiplet Dependencies</h2><p><strong>Austin:</strong> Yes, yes. I&#8217;m realizing that I kind of interrupted you. You were setting at a high level like the three things that sort of went wrong, right? You were talking about client and how that was very weak. And then I think that I jumped in and went down the rabbit hole right there because that was to me like the most shocking thing, which is back to your point, Intel is an integrated device manufacturer, meaning they have their own foundry. And so they should be able to have tight coordination between their product planning group and their foundry team. Now, of course, their foundry is also working on becoming an external foundry and serving external customers. But even internally, there should be a very tight conversation between the planning groups.</p><p>The client team should have stood up, you would think that they would have stood up a year ago and said, guys, memory prices are really high. We&#8217;re trying to pass it on. Right now some people are still buying, but our volumes are going down, even if the prices are high, so our revenue looks fine, but volumes are going to be soft and they&#8217;re going to be soft for the foreseeable future because come on, every time you and I talk about memory on here, we talk about like, this is not going to go away until 2028 when more supply comes online. And I think even then you can make the argument because of data center AI that it&#8217;s continuing to increase as models continue to get better, agentic AI gets more useful. Meta Muse is spinning up CPU VMs in the cloud and running all this AI. We&#8217;re probably not going to get to a point where everything just even in 2028 is just like, oh, hunky dory, like we have all the memory supply we need. So you could ask yourself as a client team, what should we do? And the answer has to be, well, we should be pretty conservative. And we should not overproduce chips here, even though their client team is pretty strong and they sell I&#8217;ve got a Dell XPS that has 18A in it, which is pretty awesome. It&#8217;s a great chip and everything. And I paid a lot for it. So, it wasn&#8217;t cheap, so they made some money there.</p><p><strong>Vik:</strong> Thank you, Austin. Thank you for bolstering some sales for Intel.</p><p><strong>Austin:</strong> Yeah, totally. Now, but this is where the internal advantage of having a client team, having a server team and having a foundry, like the client team would raise their hand and say like, guys, it&#8217;s not going to help our metrics, but we want to take one for the team. We think things are going to be soft. Let&#8217;s be a little conservative here and let&#8217;s give those wafers to the CPU team, which would have been perfect, you know, if they stand up and say that a year ago, because in the CPU team, the server team could say great. Our life has been really hard for many years now. AMD is as competitive as ever in server CPUs. They&#8217;ve got this high core count product that&#8217;s really popular. Agentic AI is coming, you know, maybe they couldn&#8217;t have even seen it back in October of 2025. It was a little bit before the agentic stuff really took off. That was more of like the January 2026 call. But already lots of X86 CPUs were being used for head nodes.</p><p>And so they could have said amazing. And by the way, Arm chips have been pressuring in the server, in the data center, pressuring Intel as well. And it would have been like, my goodness, this is like divine intervention. We&#8217;ve got this huge thing happening where agentic AI, we need all these CPUs. This is perfect. Give us all of your wafers. Let&#8217;s plan accordingly and let&#8217;s sell as many wafers as we can. Yes, it might hurt our client business a little bit, but we&#8217;re going to make a ton of money in data center. Data center demand is inelastic to some extent, and therefore, we&#8217;ll sell everything we can make. If it&#8217;s on if it&#8217;s made on Intel 7, Intel 3, Intel 18A, when that comes for data center, like we&#8217;re going to sell it all. We&#8217;ll be able to charge a lot. Like let&#8217;s go all in on that.</p><p>And then the foundry could have said, great, guys, we&#8217;re all on the same team. We&#8217;re going to take these wafers here, we&#8217;re going to give them over there. And by the way, we still need to have enough wafers for our external customers because really that is the long-term story for Intel. That is the true growth potential for Intel. The 2030 story is foundry. So foundry, obviously, could look at this opportunity and be like, I want to give as many wafers as possible to the server team, but we do not want to screw our customers that we&#8217;ve worked so hard to convince them. 18AP is great. 14A is coming. It&#8217;s going to be awesome. Trust us. It would be the opposite of what Intel wants to do to be like, but except we&#8217;re giving all of our wafers to Intel server group because we&#8217;ve got this once in a lifetime thing with agentic AI, right? So all of this conversation could have happened. And so I think that&#8217;s really what shocked people on the call was when it basically was, hey, we&#8217;ve got 18A&#8217;s doing great. Austin likes his Dell XPS. We&#8217;ve actually got too much of that. And by the way, we don&#8217;t have enough server CPUs still.</p><p><strong>Vik:</strong> You know, this is a big company. I have never seen a big company like that work in concert across its product lines. It&#8217;s just a sad manifestation of companies that really become large. They kind of fail to work their different business parts in unison sometimes. So maybe it&#8217;s an execution thing. Speaking of execution, there was one thing I was also comparing is basically, Intel 3 yields and cycle times for these wafers, which is basically how long it takes to get wafers through the fab, from start to finish. In the Q2 earnings call, it was like, oh, yeah, it&#8217;s exceeding our expectations. Literally, that was what was on the call. And then this time, you can hear Liputan say, yeah, cycle times are good and yield is good, but I am demanding exceptional execution from the team. I&#8217;m like, wait, I thought everybody was doing fine. Why do you have to go about demanding execution from them? Are they not doing well now?</p><p>So broadly speaking, I like that we&#8217;re going through each reason in depth because I think one step leads to another, and this is a good way to progress through what exactly happened. So it&#8217;s good thing that we break down reason number one, which is the weakness in the PC segment so much because the idea was that you just shift over wafer supply over to the other side. But when we get to the second point, I&#8217;m going to argue that maybe it wasn&#8217;t an execution thing. Maybe they had no choice. And the reason I say that is because you can make as many server CPU core chips as you want to. Like maybe you divert Intel 3 wafers away from client CPUs and you go make server CPUs with them. Great. You&#8217;ve got like a Xeon 6 processor in your hands. Now, the question is what happens after that? And this is where so many other problems came and haunted Intel all at the same time. It was like when it rains it pours kind of thing.</p><p>So the second reason, I think is a good segue into the second reason. See, whenever these chips are made, they&#8217;re useless by themselves. This piece of silicon is basically useless, right? You have got to put it with other chiplets that have other functions like IO die is a chiplet that is not made on Intel 3. It is made on Intel 7, Intel 10, whatever. So those chiplets have to be integrated with the core die, which is built on Intel 3. Now, how are they integrated? Now you need advanced packaging. And that&#8217;s the whole problem. So even if they had directed all their wafers to Intel 3 and made Xeon 6 server CPUs out of them, there are just not enough substrates. And this is Intel&#8217;s reason number two for all their problems. The industry is severely constrained on substrates, and we can cover what these substrates are, but they can&#8217;t make server CPUs without these substrates. So what do you say to that now?</p><p><strong>Austin:</strong> Yes, yes, okay, good. This is a great nuance. So the first nuance you hit, which I think is very important, and we should debate whether Intel should have gone into this level of detail to explain the miss, or not, but it used to be in a monolithic era, maybe back like 2021 and earlier where Intel would just make the entire server CPU SOC on the same node. So, Intel 10 nanometers or even of course back in like the 22 nanometer day, 32 nanometer, 45 nanometer and so on. Let&#8217;s just use all the same size transistors, whether it&#8217;s IO, CPU, GPU, memory controller, what have you.</p><p>Now, we are in an era where the cost, obviously, the cost of every new node is way more expensive. Transistors aren&#8217;t necessarily getting cheaper even as they&#8217;re shrinking. So transistors are pretty expensive. And so for a variety of reasons, it&#8217;s cost effective to say, hey, which by the way, all this IO doesn&#8217;t necessarily, a CPU, of course, you want these high performance part, these digital logic parts, you want the smallest transistors you can, you want as many transistors as possible. Like you definitely get the performance gain from moving to the next node. Some things like IO, you might not actually get as much benefit out of going to that next node. So, hey, keep it on Intel 7. That is the right thing to do. Strategically, that is correct. So this is the chiplet, so the disaggregated design, the tiles. Okay, we&#8217;re going to make our beefiest parts on the latest and greatest node, 18A or three or whatever, and we&#8217;ll keep some of the other stuff on the old nodes. And by the way, maybe those fabs have been like amortized, or we&#8217;re helping amortize them over more and more chips because every CPU we make is using its base nodes on three, and maybe it had the IO die on seven. So this is strategically good, strategically correct. And Intel has been able to do this for three, four, five years now. And partially because they have really great advanced packaging and they have EMIB, and that&#8217;s how they tie this stuff together. We had that nice conversation with Ravi recently from Intel talking about EMIB.</p><p>So all of this is strategically correct. The one nuance, and I never thought about this, is now it means that every CPU that Intel wants to make and sell, whether it&#8217;s going to be their future one on 18A, or on three or on seven, they all depend, for example, on Intel 7, or they a lot of them depend on Intel 3. They might have several process nodes. And so now there&#8217;s this interdependency where it&#8217;s not like, oh, there was a, and we would be remiss not to mention, and people following closely know this, that there was a very, very tragic accident at a fab, I think it was Fab 28 in Israel, there was an explosion, and that impacted some of their supply, and this has ripple effects because now it means that, sure, I maybe I have my dies from Intel 7 or Intel 3 or whatever, but I&#8217;m missing other parts, and so I don&#8217;t even have all the pieces to stitch them together even if I had the substrates. So that is one interesting nuance and a complication. I think there would have been an opportunity to explain that story concisely in a way to show people that, hey, it&#8217;s complicated. We had this tragic accident. It actually has a ripple effect, and it actually impacts like all of our server supply because everyone is depending on on on this node. But then to your point, then there&#8217;s the whole substrate problem, which is, yes, even if we have all of these tiles, and even if we didn&#8217;t have that accident, the question is, can we get the substrates so that we can tie this all together? To that, I would say the same thing that I would say with memory, which is the substrate issue is not a new issue. So the question is, what&#8217;s going on with the planning team? And where when substrates started to become a problem, where did Intel, because if we know this and can explain it, obviously, everyone inside Intel knows the complexities and can explain it. And again, you would have to ask, shouldn&#8217;t people raise their hand and say, we&#8217;re not going to get enough memory, we&#8217;re not getting enough supplies, we&#8217;re not going to get enough DRAM. Okay, everyone, let&#8217;s sit in a room and let&#8217;s plan all of this out. It&#8217;s complicated, but we&#8217;re going to war room, we&#8217;re going to sit down for a week and heads of every team are going to get together and we&#8217;re going to figure this out.</p><h2>Reason #3: The Fab 28 Explosion</h2><p><strong>Vik:</strong> Yeah, that was exactly the conundrum, I think Intel faced, in the last few months. And the reason number three was actually the explosion, and you covered it nicely, and that&#8217;s why I said we should go in a logical order here, because reason number two, which is advanced packaging problems, no substrate supply, leads to the fact that the Fab 28 in Israel blew up because of a gas leak. Two people actually died, so this is like really tragic stuff. And those were running input output die, which is what we&#8217;ve been referring to as IO die. And if you can&#8217;t get them, you can&#8217;t piece together the CPU anyway. But then the argument against that is that, look, it&#8217;s not the only 10 nanometer class fab they have. They I think they have at least four more fabs that are 10 nanometer class. So could they not source the IO die from those fabs, is one question.</p><p>But then for them, this explosion actually happened only in October, right? It&#8217;s literally a few weeks ago. So it&#8217;s still very early for them to make those kinds of transitions. It&#8217;s not like you can switch things over in a day. Frankly, Intel has been known to copy exactly, which means they try to copy fabs exactly the same, so that they all can be like fungible between each other. But then as much as you do it and try to make it exactly the same, transferring one product from one fab to another is always more challenging than it seems on the outside. Because there are differences. You have to make sure that the product is still fine, so you might have to go through a qualification process. Like just because you change fabs and they&#8217;re both 10 nanometer fabs, doesn&#8217;t mean they have to work the same, and you&#8217;ve got to check that they work the same. So there&#8217;s some work involved.</p><p>So things have been so recent that I think they will recover from the supply issue of the 10 nanometer IO die chiplet, supply issue because they&#8217;re going to probably transfer it to somewhere else. Or they will restart Fab 28 and things will go on, as usual, and they&#8217;ll hopefully solve whatever caused that explosion. So that problem will be fixed. But anyway, they have lost $300 million just from that explosion, okay? And that&#8217;s for this quarter. They expect another 300 million damages or opportunity cost, however you put it. I don&#8217;t exactly know how.</p><p><strong>Austin:</strong> Yes, yes. Okay.</p><p><strong>Vik:</strong> in the next quarter.</p><p><strong>Austin:</strong> Yeah.</p><p><strong>Vik:</strong> Yeah, and and more, and more. I mean, 300 plus, okay. And nobody knows the exact extent of the damage because it&#8217;s been like so recent. I&#8217;m not sure how long this will proceed. There&#8217;s no reason to believe that it&#8217;s $600 million and done. If it goes to 700 or 800, I&#8217;m not surprised at all.</p><p><strong>Austin:</strong> Right. So here is probably a missed opportunity as well, which is, hey, by the way, yeah, this happened at the start of October, so we were two months in, so the first two months should have been fine. So there&#8217;s a question of how much of some of the server CPU supply miss is directly attributable to this versus just poor planning.</p><h2>What If? A Counterfactual Bet on Server</h2><p><strong>Vik:</strong> Yeah, that was not mentioned at all. That&#8217;s a good point. Intel internally is in a hard place, right? Their CPU client, the PC market is slowing down. It&#8217;s going to go down 20% or more just in the next quarter and probably 12% down all of 2026. And then the outlook for 2027 isn&#8217;t any better. The memory supply is not going to get better in 2027. And everything is going to still be constrained in the mobile markets and the PC markets. And everybody knows this.</p><p>Now, substrates are of short supply and Intel has been trying to get pre-prepayments and like supply commitments from the Japanese substrate makers like Iden or Unimicron and there are like a few of them like AT&amp;S. All these people can make these substrates, but they&#8217;re making the investments. So it takes time for substrate supply to come back up, right? So puts them in a hard place. Like think about it from Intel&#8217;s point of view. They don&#8217;t have the ability to sell into PC market because memory is wrecking everything. They can&#8217;t make server CPUs, even though the demand is amazing because apart from the plant blowing up thing and the supply of 10 nanometer being disrupted a little bit, which I&#8217;m sure they&#8217;ll recover. The bigger problem is substrates are not going to come online. So, what product do you make as Intel? What do you do? All you have to give them is investors is like, look, we&#8217;re focusing on the foundry business. We don&#8217;t have any solid people announcing, but trust us, bro. We&#8217;ve got customers who are waiting to tape out. It&#8217;s not a strong story. You know what I&#8217;m saying?</p><p><strong>Austin:</strong> Yeah, yeah.</p><p><strong>Vik:</strong> Where is your current story? Like how much money are you going to make in the next three months, next six months? The Intel foundry story is a good one, but like you say, it&#8217;s a 2030 story. You can&#8217;t sell it to anybody in the next two, three months or six months.</p><p><strong>Austin:</strong> Right. And this is the problem with not planning well in advance, which is and maybe for listeners, you might say, oh, Intel doesn&#8217;t have enough Intel 3 supply or they&#8217;ve got lots of extra 18A because they&#8217;re actually ramping yields quicker on 18A than they expected and the PC&#8217;s down and right now all of 18A has been only used to make client CPUs. Can&#8217;t they just like quickly use that 18A to make server CPUs? Well, and I&#8217;m sure people know this, but just as a reminder, there are different chips that are designed for the different process nodes. So Sapphire Rapids and Emerald Rapids are these server CPUs that were designed for Intel 7. And by the way, let me look at my notes. So the Rapids chips, anything with the Rapids on it is a server CPU with the performance core, so P core. So think like beefy, probably good in a head node, good in your data center, for memory, like heavy, like big databases with lots of that need lots of memory, that kind of thing. That&#8217;s Rapids. If it ends in forest, it&#8217;s a server CPU with a lot of efficient cores. So these could be good for like agentic CPUs, for example, or just for like traditional serving APIs and that kind of thing.</p><p>So, back to what I was saying was like Sapphire Rapids, Emerald Rapids, those are on Intel 7. So if you&#8217;re like, oh no, we had this tragic accident, we can&#8217;t get enough Intel 7. You can&#8217;t just picked up those designs and be like, oh, put that on 18A or something. Like no, it was designed for another process node, of course. Then you&#8217;ve got Sierra Forest designed for Intel 3, Granite Rapids designed for Intel 3, Clearwater Forest designed for 18A. So if there&#8217;s lots of extra 18A capacity, you can&#8217;t just immediately go, oh, pull Clearwater Forest forward because yes, that chip has been designed, but there&#8217;s obviously validation, testing, ramping it, that kind of thing. So it&#8217;s not so simple to just like, you can&#8217;t grab CPUs and just kind of print them wherever you want. So the only about the only thing that Intel can do right now is say, oh, can we take some of our tools from if we are able to if what was the term they used that 18A is underloaded. It can make this many chips, but we actually don&#8217;t, we&#8217;re not going to make that many chips, even though it can. About the you can&#8217;t move a design over to 18A. All about the only thing you can do is take the fab equipment from 18A and go use that to bring up more Intel 3 capacity. So like tools that do etching and deposition, that kind of stuff.</p><p><strong>Vik:</strong> I buy that in theory, but normally that would take its own sweet time one.</p><p><strong>Austin:</strong> It does. It does. That is the problem.</p><p><strong>Vik:</strong> That is the problem. The second thing is that if you&#8217;ve got all these machines hooked up and they are currently being used to improve the yield of 18A, which is a very important thing because this is the competitor to TSMC&#8217;s 2 nanometer technology. You don&#8217;t want to stop that train from moving and go backwards by taking away equipment from 18A. That is like you say, the 2030 story has to be remained because the world needs a foundry alternative to TSMC&#8217;s 2 nanometer and 18A is critical to that. So strategically, maybe that&#8217;s not the best idea to do to take away 18A.</p><p><strong>Austin:</strong> Correct. So yes, to your point, 18AP is really the node that externally will be used. So maybe if it was like only for 18A, they could move some of it. But to your point, you do not want to rob Peter to pay Paul. You don&#8217;t want to pull away from the future of the company and what the external foundry is riding on just to make some of these old chips to sell right now, right? So you would not This is where they are and these are some of the decisions they&#8217;re making. But again, you would have rather seen this and predicted this stuff six months ago and had a backup plan so that you didn&#8217;t have to move tools around.</p><p><strong>Vik:</strong> Yeah, yeah. So I&#8217;m trying to think about, maybe this is like an open brainstorm session for...</p><p><strong>Austin:</strong> Yes. What is Okay, you go.</p><p><strong>Vik:</strong> The one possible outcome I could think of is...</p><p><strong>Austin:</strong> if a year ago they made a decision to say, hey, October of 2026, there&#8217;s a lot of stuff they couldn&#8217;t have foreseen. But even if they waited till back in January when agentic stuff was taking off and you could say, holy crap, this is going to use a lot of CPUs. What if in October 26, instead of having too much client and not enough server, what if they had made a bet to have too much server and not enough client? What if they would have reprioritized things then and said, let&#8217;s air on the side of having too many server CPUs and not enough client. What would have happened? Well, if client would have would client is obviously very soft, so they would have ended up fine. Server is very strong, they would have sold it all. I think they would have been totally fine. But even if client wasn&#8217;t to soft, client down 20% this quarter and they thought it was only going to be down 10%. And server&#8217;s crazy hot and even if that wouldn&#8217;t have been quite as hot, I think investors would have given them the benefit of the doubt of saying like, oh, you bet on AI. Good.</p><p><strong>Vik:</strong> Okay.</p><p><strong>Austin:</strong> You missed on client PC. I get it. Okay, right? So if they would have made that bet, they would have won today and even if the future wouldn&#8217;t have turned out the way that it did, I still think investors would have given them the benefit of the doubt as compared to literally it&#8217;s been like five straight quarters or something of saying, maybe even six straight quarters of saying, oh, there&#8217;s more server demand than we can meet. Imagine if this quarter they would have said, we actually had too many server chips. We would have been like...</p><p><strong>Vik:</strong> Man.</p><p><strong>Austin:</strong> Yeah. You&#8217;re paying attention. Good job, right? Yeah, yeah. That&#8217;s the only thing I can think of.</p><h2>The Foundry Story and Long-Term Outlook</h2><p><strong>Vik:</strong> Yes, I agree with that. I think your idea has two downstream consequences. One is that if you planned early enough, first of all, it was very surprising for me that a company like Intel with all its smart people were actually surprised in the earnings call that agentic CPUs had such a demand. I remember like two quarters ago, the earnings call like, oh, wow, look at that agentic CPU demand. We didn&#8217;t expect that. I&#8217;m like, hang on, you guys know stuff about CPUs, right? You&#8217;re the experts. And if you have seen that there is a trend of the open claw and and all this stuff coming up, you kind of have to have an inkling of how it works and what that means for your hardware demand. So that&#8217;s the first thing. Why were you taken aback by agentic demand?</p><p><strong>Austin:</strong> Yes. Okay, let me say something.</p><p><strong>Vik:</strong> Yeah, yeah, say it.</p><p><strong>Austin:</strong> Yes, don&#8217;t forget your second thing. So we&#8217;ll come back to it, but on that one, this could be the biggest downside of Intel not having a competitive GPU. It&#8217;s not that they&#8217;re missing out on all this money, Nvidia&#8217;s obviously crushing it, AMD&#8217;s growing and trying to make money, but they&#8217;re missing out on those conversations with customers to get as early insight as possible of holy crap, they&#8217;re going to need a ton more CPUs. Now, Intel should Intel&#8217;s data center department should have those and do have those kind of relationships with Google, with Meta. I don&#8217;t know if they have it with Open AI and Anthropic. In theory, Open AI and Anthropic are still running on CPUs that their hyperscalers are buying, but I think that maybe it&#8217;s Intel&#8217;s lack of a relationship with the model labs in the same way that Nvidia and probably AMD have that put them further behind here. That&#8217;s my hypothesis.</p><p><strong>Vik:</strong> Could be. Could be. So, they being taken aback by CPU demand was taking I was taken aback by them being taken aback. You see what I mean?</p><p><strong>Austin:</strong> Right, right, right. Yeah, totally. It&#8217;s like you have one business. It&#8217;s CPUs. Everything else is a rounding error.</p><p><strong>Vik:</strong> Okay, so they were surprised by it. This was like, I think February or something this year, like whenever the Q1, they were surprised in Q1. Okay, now we&#8217;ve entering Q4. So like you say, at that point at least, if they had made the decision to go entirely into data center because memory has already been short since the early this year. Everybody knows this memory shortage is not even new this year. We were short of memory last year, actually. And so everybody knows how HBM works. So you know that memory is going to be short. Now, not only do you know memory is going to be short, you know that if you&#8217;re going to make even PC or server CPUs, any CPU needs memory too. So, even if you came across the concept that, oh, wow, we need agentic CPU demand is high. At that time, you should realize that DRAM demand goes hand in hand with server CPU demand. So, DRAM is going to get further constrained. And so what does that mean for the client? The clients cannot pay that price for DRAM, spot prices will increase. And what does that mean for the PC business? Nobody can buy PCs because it&#8217;s too expensive. And therefore, we better double down on servers. And what do servers need, by the way? They need this was my second point. They need substrate supply. So, get on the substrate suppliers a year in advance and invest in them. You see what Nvidia does with Lumentum? They&#8217;re like, oh, we&#8217;re going to optics soon. Here, Lumentum, here&#8217;s $2 billion. Go do something with it, right?</p><p><strong>Austin:</strong> Right. Now, of course, Nvidia has a lot of cash and Intel cash is tight. But if not invest, pay a premium for those wafers. Pay more than TSMC. I don&#8217;t care. You can show up on the earnings call and you can say, I met all the supply, but our ASPs were really high and therefore the market was a little bit softer than we expected. But I had the supply because I got the substrates. How&#8217;d you get the substrates? I paid a premium for them, right? Like take that kind of risk. We&#8217;re going to have the supply, it&#8217;s just going to be expensive. Our costs were too high. Our margins.</p><p>Look, okay, because here&#8217;s let me give some there&#8217;s lots of positives for Intel still. The foundry&#8217;s looking great. We were in a position two years ago where everyone was like, oh, five nodes in four years, I think it&#8217;s going to crash and burn. I don&#8217;t think they can make I don&#8217;t think 18A is going to work. I don&#8217;t think 14A is going to work. And now people are like, no, I think that stuff&#8217;s going to work. The question is just how much can they make of it, right? And so I think the best thing Intel could have done is said, we can make stuff. It&#8217;s really expensive right now and we have to get better at yields and we have to negotiate down our suppliers and that kind of thing, but we can make stuff. But when your whole even 2030 story is about making stuff and you don&#8217;t have enough of the right stuff, that&#8217;s where people are concerned. And then, there&#8217;s the question of is it does that ultimately come down to foundry and then should we be concerned about them not being able to make the right stuff for their customers? Or were they prioritizing their customers and it&#8217;s, this was they purposefully, did this, or is this a product planning thing where the product team, as we just talked about, should have said, we&#8217;re going to need more CPUs, agents, we&#8217;re going to need more memory, we&#8217;re going to need more substrates. So should the product team have been driving this? I don&#8217;t know.</p><p>The best case sort of like quote unquote blame, ultimately at the end of the day, it&#8217;s all Intel&#8217;s problem. They&#8217;re all on the same team and this is all falls onto the. The best thing is that the product team should be if they&#8217;re the ones that were misplanning, then foundry is just like, hey, we&#8217;re trying to serve all of our customers and this is yet another customer who didn&#8217;t give us the best forecast. So we need to get better at like validating our customers&#8217; forecast. But if it has to do with foundry at all, then that starts to get concerning again, which is I thought the 2030 story is foundry. If you can&#8217;t even plan capacity for your biggest customer, Intel, which you guys share up cubicles, how are you going to do this for everyone else?</p><p><strong>Vik:</strong> Yeah, that&#8217;s an argument. But going forward, I think long term there is still, 18A will eventually be adopted somewhere, somehow because if this AI build out continues, I find it hard to believe that the entire thing can all those gigawatts can be fulfilled with just TSMC making chips. Somehow it seems like that&#8217;s not going to happen. Like somebody somewhere has to start making 18A chips. And they will unlock a whole lot of capacity for Intel and hopefully that helps. It&#8217;s just not happened yet, but I hope it happens in the future. 14A is very good. As far as I can tell, it&#8217;s in high demand as well for mostly like mobile applications. Mobile applications are the ones who usually go for the most cutting edge technologies because they the energy saving is the most there, right? So, it&#8217;s very important for phones.</p><p><strong>Austin:</strong> Correct. Yes. That so the yeah, looking forward, unfortunately, 2027&#8217;s not going to look good. We would love if a customer came out and said, I&#8217;m committed to 14A, that would help draw attention over here and not over here on the fact that there&#8217;s still not enough supply and that the client PC is still soft. We probably aren&#8217;t going to get that. But that would be the dream scenario. In the meantime, if I was Intel, yes, they would be able to make mobile chips, which would be great. They also make high-performance CPUs and so they could say, we&#8217;ll make anyone CPU. And I think they actually did. There&#8217;s a little bit of talk of custom ASICs, which is a very broad bucket, by the way. And they talked about making X86 CPU IP that you could someone else could instantiate, I believe. Did you hear that on the call? So it&#8217;s like if I wanted to make an X86 core in a custom product, I could instead of Arm or RISC-V. So there&#8217;s probably like more to lean in there. But I don&#8217;t know. I&#8217;d just be like, hey, we can make high-performance CPUs, we should be able to make high-performance accelerators, like you think they&#8217;d try to go after accelerators on 14A too. And this whole custom ASIC thing, which I don&#8217;t think is XPUs, I think it&#8217;s more like a custom networking chip or a custom CPU, make it tweak it a little bit. Obviously, if they could make XPUs on 14A, that would be amazing.</p><p><strong>Vik:</strong> But you know, as a foundry business, there&#8217;s nothing stopping Intel from making Arm processors, right? Like they just stamp it.</p><p><strong>Austin:</strong> And they said as much. And in fact, when they prove, when they like do test chips to show how good 18 AP is or 14A, they make Arm cores so that you can compare like apples to apples, say built on Intel, built on TSMC, here&#8217;s how this Arm core does.</p><p><strong>Vik:</strong> Yeah, so if any Arm company, Arm itself wants to go and possibly make it on Intel, that&#8217;ll be interesting actually. The dynamic of Arm making a Arm CPU in an Intel foundry is just, I don&#8217;t know, it&#8217;s weird. It&#8217;s just weird.</p><p><strong>Austin:</strong> Today&#8217;s sponsor is Meter. You&#8217;ve got the GPUs and the power. Now, who builds the network and who do you call when it breaks? With Meter, it&#8217;s one call. Tell them your GPU or server count and they design, deploy and run your front-end storage and management networks. Meter builds its own networking hardware, switches and all, and writes its own network OS. So that call goes to the team who built it. Running a neocloud or clusters in a colo, go to semidoped.com/meter. Links in the show notes.</p><p><strong>Austin:</strong> EMIB and advanced packaging. Hopefully, maybe in the next couple quarters we could get some more good news of other people taking even TSMC tiles but having them packaged in the United States. You know, made in TSMC Arizona packaged by Intel in New Mexico and shipped. That those would be hope stories that would be promising and again, continue to show that 2030 thesis playing out.</p><h2>Four Lessons from the Intel Saga</h2><p><strong>Vik:</strong> Yeah, yeah. Since we&#8217;re getting down to the wire here, I wrote down a couple of lessons I learned from this whole episode, because I feel like it&#8217;ll apply everywhere. And it doesn&#8217;t have today it&#8217;s Intel. Tomorrow it could be any company to be honest, because a lot of these themes go across everybody and every company and every investor, right? So the one thing I was thinking about and that came up in my mind was lesson number one I learned was demand itself doesn&#8217;t mean anything. Like Intel still has demand for their CPUs. People want Intel server CPUs. And the market has been pricing in just demand. Oh like, oh, look at how much demand we have. This happens so much across like the optics industry or whatever. We have demand, but we can&#8217;t meet supply. And everybody&#8217;s like stuck up because like, wow, imagine what would happen if they could meet the demand. So much demand, they&#8217;re going to continue to print money because customers are waiting out the door. It&#8217;s exactly like this, right? Like there&#8217;s a restaurant down your road and you go there, it&#8217;s like there&#8217;s a long line out the door. And immediately you go tell all your friends like, oh, I found this great restaurant down the street. But the question is if you when you go to stand in that line, are you ever going to get to eat before you die of starvation? Probably, you&#8217;re not going to get to eat for like one or two hours just the weight line. Then you ask yourself, was it worth it? This is what is happening. Everybody has demand, but no supply. And so this is what happens. Sometimes, Intel is in that situation. So it doesn&#8217;t matter if you have demand. So I think something to keep in mind.</p><p>So second thing is within a company itself, I think AI can help one business, like server CPUs in this case, and hurt another. Like the client business, the PC business. Because the server CPUs are sucking up so much resources from the ecosystem overall that it affects their own business on the other side. It&#8217;s kind of funny how this works, but that can happen to a company, right?</p><p><strong>Austin:</strong> Yes, let me jump in here too. What just popped in my head was the funny thing too is that on the client side, in theory, if AI PC was really a thing, the client side should be hit even worse because you&#8217;d be like, no, I really want to run local AI and therefore I need more memory. But like memory&#8217;s crazy expensive. So this actually, I had a tweet that said Intel should retire the term AI PC and it&#8217;s not just Intel, it&#8217;s Microsoft and everyone. Because no one&#8217;s actually running really LLMs locally and that&#8217;s fine, but don&#8217;t sell the value prop of this is an AI PC and you can do AI locally. Actually this would have not only been a good time to do it just because it&#8217;s not true for the last three years, but also because memory is so expensive, you don&#8217;t want people to think, oh, an AI PC has all this memory because it&#8217;s going to run all this stuff locally and therefore I should definitely not get it. You know, just retire that and come up with something else and make it clear that it doesn&#8217;t have to have a ton of memory, maybe next quarter. But carry on with your lessons.</p><p><strong>Vik:</strong> Yes, lesson number three I came up with that I can see going forward is that just like beating earnings and it used to be enough, company did well and stuff was good. But a lot of times I think nowadays that is not enough and there is an expectation that outruns reality. And I&#8217;m not saying that&#8217;s a good or bad thing. It&#8217;s just like a fact of the investing markets at this present time. Just an observation. It&#8217;s good to keep that in mind. Like good numbers always doesn&#8217;t mean good things for the stock, right? Like from an investor perspective. And finally, the market at this point, we are reaching a boiling point in the industry where people are asking for returns. going to come asking for the receipts and that&#8217;s going to be important, right? So, I think that&#8217;s my big takeaways here. Things are changing. The tides are the excitement is leveling off and people are starting to ask for the receipts overall. And that affects materially how we view a lot of companies in this space, not just Intel.</p><p><strong>Austin:</strong> So to recap, the market is in an interesting position. Lots of things are happening externally. So to Intel&#8217;s credit, yes, they&#8217;re down 30%, but the whole market is also down broadly. Interest rates are high, there&#8217;s a ton of volatility. People are starting to demand returns. You can&#8217;t just promise, hey, we have high demand, you have to meet the demand. You can&#8217;t just have a good quarter, you have to show that you will continue to have a good quarter and that a good quarter now, but a soft quarter tomorrow, investors are smart and they&#8217;re going to see that. I think the same thing could even hit Apple, they&#8217;ll have some good quarters, but memory will impact them too. So it is tough to be Intel. They&#8217;re managing making their own wafers, tragic accidents, all these different process nodes, depending on all of them. They&#8217;re managing substrates, they&#8217;re managing memory. They&#8217;ve got a server business and a client business that are very different in elasticity, and then on top of it they&#8217;re trying to stand up an external foundry, which they can&#8217;t talk about really. So we have to play investigator to try to figure</p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 9th, 2026]]></title><description><![CDATA[Nvidia&#8217;s board investigates its CEO, Mitsui Chemicals buys JSR, and TSMC says it is staying out of DRAM]]></description><link>https://daily.semidoped.com/p/daily-update-november-9th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-9th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Mon, 09 Nov 2026 14:04:09 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/c0a15f39-36d6-4555-80d1-a31713abca68_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Monday, November 9th, 2026</strong></p><p>Nvidia&#8217;s board opened an independent investigation into whether Jensen Huang was the target of Chinese intelligence, about a week before the company reports earnings. Japan got its largest chemicals deal on record as Mitsui Chemicals agreed to buy photoresist maker JSR for &#165;1.84 trillion. TSMC pushed back on a report that it is planning 3D DRAM for Nvidia, Elon Musk&#8217;s Terafab told Japanese toolmakers it wants a million wafers a month, and Microsoft is giving Oracle back its older GPUs.</p><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>3D Stacked DRAM for Inference: d-Matrix&#8217;s Sudeep Bhoja</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=zEqoBSJ4YQ4">YouTube</a>, <a href="https://podcasts.apple.com/us/podcast/3d-stacked-dram-for-inference-d-matrixs-sudeep-bhoja/id1866707196?i=1000800105437">Apple Podcasts</a>, or <a href="https://open.spotify.com/show/0Uuu3s1Nw09f6Xmg24rCZm">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-d-matrix-sudeep-bhoja">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p><mark data-color="#d9d2e9" style="background-color: rgb(217, 210, 233); color: rgb(0, 0, 0);">Fine-tune and serve open models without managing a single GPU.</mark></p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p><strong>Accounts are free, and get $5 in credit when you sign up.</strong></p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning &#8594;</a></p></div><h3>Nvidia&#8217;s board opens an investigation into Jensen Huang</h3><p>On Sunday, Nvidia&#8217;s board said a committee of independent directors, working with outside counsel, will investigate reports that Huang was the target of a foreign intelligence effort. The board said it has no reason to believe U.S. national security or Nvidia secrets were compromised, and that it has confidence in Huang. The same day, Nvidia published an email in which Huang told employees he had been unfaithful to his wife and would cooperate fully with the review.</p><p>The reporting goes back to the New York Times, which said last month that Huang had a relationship with Alice Huang (no relation), a former executive at Singapore&#8217;s Megaspeed whom, according to the Times, U.S. authorities concluded was a Chinese intelligence asset. Megaspeed is an Nvidia cloud partner that U.S. agencies have been investigating over whether it helped get restricted chips into China. The Wall Street Journal reports its name came up in recent Entity List deliberations, though it has not been added. Nvidia had earlier described the pair&#8217;s interactions as not unusual, and Reuters says it could not independently verify the Times&#8217; claims.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The bigger risk for Nvidia is in Washington, where a CEO who has lobbied for looser China export rules now faces questions about a suspected Chinese asset.</mark> House Republicans had already asked Huang for a briefing before Sunday&#8217;s statement.</p><p>Markets split on Monday. Nvidia was down about 5% in premarket trading, while TSMC rose 4.68% and the Taiex closed up 4.94%, its second-largest point gain ever, after Taiwan pledged to support the market. In Seoul, Huang&#8217;s daughter Madison, an Nvidia senior director, cancelled her trip, which scrapped a meeting with Korea&#8217;s science minister and a visit to Rebellions. Korea&#8217;s science ministry said it expects contracted GPU deliveries to go ahead as planned. Nvidia reports third-quarter results on November 17.</p><p><a href="https://nvidianews.nvidia.com/news/statement-from-nvidia-board-of-directors">Source: Nvidia board statement</a> &#183; <a href="https://nvidianews.nvidia.com/news/a-message-to-our-company">Huang&#8217;s email to employees</a> &#183; <a href="https://www.investing.com/news/stock-market-news/explainerwhat-we-know-about-nvidias-probe-into-ceo-jensen-huang-5010580">Reuters explainer</a> &#183; <a href="https://www.cnbc.com/2026/11/09/asia-chip-stocks-south-korea-taiwan-nvidia.html">CNBC</a> &#183; <a href="https://focustaiwan.tw/business/202611090017">Focus Taiwan</a> &#183; <a href="https://biz.chosun.com/en/en-it/2026/11/08/VD6LU5L75JE2PAECBSWKDL4KVM/">Chosunbiz</a></p><blockquote><p><strong>Vik: </strong><em>Going to pass on commenting about this as it involves personal lives. We focus on business and tech here at Semi Doped. Nvidia makes some amazing hardware, nothing changed overnight on that front.</em></p></blockquote><h3>Mitsui Chemicals buys JSR for &#165;1.84 trillion</h3><p>Mitsui Chemicals agreed to buy JSR from funds run by JIC Capital, part of the state-backed Japan Investment Corp., for &#165;1.84 trillion (about $11.6 billion), a figure that includes &#165;447.9 billion of debt. The Development Bank of Japan is putting in &#165;200 billion for non-voting preferred shares. Mitsui will fund the rest with a &#165;1.64 trillion bridge loan and plans to issue up to &#165;150 billion in new shares. Closing is targeted between October 2027 and March 2028, pending antitrust and foreign-investment approvals.</p><p>JSR has been private since JIC took it over in 2024, and its semiconductor materials sales reached &#165;196.0 billion last fiscal year. <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Mitsui is paying for JSR&#8217;s EUV resist position: by Mitsui&#8217;s account, JSR is No. 2 in photoresists worldwide and the only supplier offering both chemically amplified and metal-oxide resists for EUV.</mark> Mitsui&#8217;s own chip materials business is much smaller. Its EUV pellicles, ICROS wafer tape, and one film product are its only lines with annual sales above &#165;10 billion.</p><p><strong>Reuters says it is the largest deal ever in Japan&#8217;s chemicals sector</strong>. Mitsui is targeting &#165;40 billion of extra annual profit by fiscal 2035, and it is now considering splitting its petrochemicals and mobility businesses into a separate polymer company.</p><p><a href="https://jp.mitsuichemicals.com/content/dam/mitsuichemicals/sites/mci/documents/release/2026/261109e.pdf">Source: Mitsui Chemicals disclosure</a> &#183; <a href="https://jp.mitsuichemicals.com/content/dam/mitsuichemicals/sites/mci/documents/release/2026/event_261109e.pdf">Mitsui Chemicals investor presentation</a> &#183; <a href="https://www.channelnewsasia.com/business/mitsui-chemicals-buy-chip-materials-firm-jsr-116-billion-6535076">Reuters via CNA</a></p><h3>TSMC says it is staying a logic foundry after a 3D DRAM report</h3><p>The Information reported Friday, citing two people, that TSMC has started drafting plans to make 3D DRAM at Nvidia&#8217;s request. The options reportedly include building new fabs or buying an existing plant and converting it. The effort is said to be led by Hsu Kuo-chin, the vice president over advanced packaging and a former head of Micron&#8217;s Taiwan operations. No timeline or capacity was reported.</p><p>TSMC&#8217;s answer: &#8220;TSMC focuses on our core business at all times, which is dedicated IC foundry. Our mission is to be the trusted technology and capacity provider of the global logic IC industry for years to come.&#8221; It did not directly address whether Nvidia asked. Winbond chairman Chiao Yu-Cheng said process and patent barriers would make it hard for TSMC to enter DRAM directly.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Even if the report is right, 3D DRAM is still years from volume at Samsung, SK hynix and Micron, so it would not ease today&#8217;s shortage.</mark> What the report does show is how hard Nvidia is pushing for more memory supply. A DRAM business would also complicate TSMC&#8217;s role making HBM base dies for SK hynix.</p><p><a href="https://www.tomshardware.com/pc-components/dram/tsmc-denies-exploring-3d-dram-at-nvidias-request-advanced-memory-tech-is-still-years-from-market-viability">Source: Tom&#8217;s Hardware (TSMC statement)</a> &#183; <a href="https://www.investing.com/news/stock-market-news/tsmc-considers-making-advanced-memory-chips-the-information-reports-5008786">Reuters (The Information report)</a> &#183; <a href="https://www.taipeitimes.com/News/biz/archives/2026/11/09/2003867384">Taipei Times</a> &#183; <a href="https://www.businesskorea.co.kr/news/articleView.html?idxno=280765">BusinessKorea</a></p><blockquote><p><strong>Vik: </strong><em>Unsure if these reports are true. But I am a big fan of 3D dram. It has the opportunity to unlock new levels of performance. This is why we interviewed d-Matrix CTO Sudeep Bhoja on Semi Doped. Watch it <a href="https://www.youtube.com/watch?v=zEqoBSJ4YQ4">here</a>.</em></p><p><strong>Austin: </strong><em>That d-Matrix episode was one of my favorite yet. Really insightful.</em></p></blockquote><h3>Terafab tells toolmakers it wants a million wafers a month</h3><p>Nikkei reports that Elon Musk&#8217;s Terafab told Japanese equipment makers this autumn it is targeting about 1 million wafers a month from 2032. <strong>About 700,000 would be memory and more than 300,000 logic</strong>. A Terafab procurement team visited Japan earlier this month and asked some suppliers to deliver within months, aiming for mass production in the second half of 2028. Some suppliers were confused by demands that cut against standard practice.</p><p>The total would be about 40 times the planned capacity of Rapidus&#8217; Hokkaido fab. By one estimate, the memory share alone roughly matches Samsung&#8217;s entire DRAM capacity today. Jiji Press has reported that Intel is coordinating technical support for logic on its 2nm- and 1.4nm-class processes, and that Terafab is weighing a Taiwanese partner for memory. No partner has been confirmed.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Terafab has a capacity target and a supplier list, but it has not said where its DRAM process technology will come from.</mark> Advanced fabs usually take more than three years to go from planning to volume production. For Japanese toolmakers, the near-term question is whether the rush requests turn into firm orders.</p><p><a href="https://www.nikkei.com/article/DGXZQOUC026RB0S6A101C2000000/">Source: Nikkei</a> &#183; <a href="https://www.trendforce.com/news/2026/11/09/news-terafab-reportedly-eyes-1m-wafers-monthly-by-2032-with-700k-for-memory-near-samsungs-total-dram-capacity/">TrendForce</a> &#183; <a href="https://en.fnnews.com/news/202611090902383508">Financial News</a></p><blockquote><p><strong>Vik: </strong><em>That number seems far too high to me. 2032 is 5 years away and Elon wants to launch from the ground up to match Samsung&#8217;s entire DRAM capacity today? I&#8217;m staying skeptical.</em></p></blockquote><h3>Microsoft hands Oracle back its older GPUs</h3><p>Bloomberg reported Friday that Microsoft is winding down a 2023 deal to rent AI computing from Oracle after deciding not to renew it, and has been switching off the capacity in recent months. The servers ran older Nvidia GPUs and were originally used for Bing&#8217;s AI features. According to Bloomberg, they no longer fit Microsoft&#8217;s specifications. The deal&#8217;s size was never disclosed.</p><p>Oracle said demand for its AI infrastructure &#8220;remains strong, including for earlier-generation GPUs.&#8221; Microsoft declined to comment. Bloomberg says the move is part of a review of more than $80 billion Microsoft has committed to outside capacity from CoreWeave, Nscale, Nebius and others. In the same week, IREN said it had delayed some deliveries after Microsoft asked for network design changes.</p><p><mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">When an older-GPU lease expires, one of the biggest renters of AI capacity is choosing to walk away rather than renew.</mark> Rented GPU capacity is usually financed on the assumption that the chips keep earning for years. Nebius reports on November 12 and CoreWeave also reports this week.</p><p><a href="https://www.bloomberg.com/news/articles/2026-11-06/microsoft-is-winding-down-oracle-deal-to-rent-ai-cloud-capacity">Source: Bloomberg</a> &#183; <a href="https://www.investing.com/news/stock-market-news/microsoft-is-winding-down-oracle-deal-to-rent-ai-cloud-capacity-bloomberg-news-reports-5008946">Reuters</a></p><blockquote><p><strong>Vik: </strong><em>Who walks away from compute? Older Nvidia GPUs can still run inference of some kind. Maybe Microsoft has no need for that inference?</em></p></blockquote><div class="callout-block" data-callout="true"><h4>Supported by Meter</h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter-daily">See how Meter builds data center networks</a></p></div><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>Neocloud orders</strong> &#8212; An expert at a global AI server OEM says only 30&#8211;40% of first-phase MOUs from newer GPU-rental customers turn into binding purchase orders, and follow-on orders from many of them remain tentative through the first half of 2027. Hyperscalers, which pay from their own cash, convert at very high rates. The same expert says Vera Rubin&#8217;s volume ramp has moved from December to the first quarter of 2027, with this OEM expecting 5,000&#8211;6,000 Rubin racks next year at about $8 million each (<a href="https://wukong123.substack.com/p/channel-check-from-oem-expert-rubin">WK channel check</a>).</p></li><li><p><strong>Samsung HBM mix</strong> &#8212; Industry participants expect Samsung to start moving some wafer capacity from HBM back to conventional DRAM in the fourth quarter, because DRAM margins have nearly caught up with HBM. In that scenario, memory prices level off in the first quarter of 2027 and ease in the second (<a href="https://wukong123.substack.com/p/channel-check-from-oem-expert-rubin">WK channel check</a>).</p></li></ul><h2>Sector Watch</h2><p><strong>Compute</strong></p><ul><li><p><strong>Rebellions / SK Telecom</strong> &#8212; Rebellions will connect its next-generation R200 NPUs to Nvidia&#8217;s platform through NVLink Fusion and MGX racks, and SK Telecom plans to deploy the systems with Nvidia GPUs running prefill and Rebellions NPUs running decode. The R200 systems are expected to be ready in 2028 (<a href="https://nvidianews.nvidia.com/news/nvidia-rebellions-and-sk-telecom-announce-collaboration-to-advance-sovereign-ai-infrastructure-with-nvidia-nvlink-fusion">Source: Nvidia</a>).</p></li><li><p><strong>Groq</strong> &#8212; Groq is pitching investors on a Series B at an $8 billion pre-money valuation, about 2.5 times its August mark, to fund inference capacity sold on short- and medium-term contracts, Bloomberg reports (<a href="https://www.investing.com/news/stock-market-news/nvidiabacked-groq-seeks-8-billion-valuation-in-new-funding-round--bloomberg-5009486">Source: Bloomberg via Investing.com</a>).</p></li></ul><p><strong>Servers</strong></p><ul><li><p><strong>Quanta</strong> &#8212; October revenue doubled from a year earlier to NT$353.25 billion, a record for the month, as AI servers more than offset falling notebook shipments (<a href="https://tw.stock.yahoo.com/news/%E5%85%AC%E5%91%8A-%E5%BB%A3%E9%81%94-2026%E5%B9%B410%E6%9C%88%E5%90%88%E4%BD%B5%E7%87%9F%E6%94%B63532-5%E5%84%84%E5%85%83-%E5%B9%B4%E5%A2%9E103-061213536.html">Source: Quanta filing via Yahoo Taiwan</a>).</p></li><li><p><strong>Wiwynn</strong> &#8212; The Wistron server unit posted record third-quarter revenue of NT$434.6 billion, up 62.9%, with after-tax profit up 51.6% (<a href="https://www.wiwynn.com/news/wiwynn-reports-third-quarter-2026-financial-results">Source: Wiwynn</a>).</p></li></ul><p><strong>Equipment</strong></p><ul><li><p><strong>Kokusai Electric</strong> &#8212; The Japanese deposition-tool maker&#8217;s fiscal second-quarter adjusted operating profit rose 56% on AI-driven demand, and it kept full-year guidance of &#165;340 billion in revenue and &#165;86 billion in adjusted operating profit (<a href="https://www.nikkei.com/markets/ir/irftp/data/tdnr/tdnetg3/20261109/g81k9g/140120261109568660.pdf">Source: Kokusai Electric results slides</a> &#183; <a href="https://www.tradingview.com/news/urn:summary_document_slides:quartr.com:4268385:0-kokusai-electric-strong-ai-led-demand-lifts-results-fy27-guidance-remains-unchanged/">TradingView summary</a>).</p></li><li><p><strong>SMEE</strong> &#8212; At the China International Import Expo, China&#8217;s main lithography maker showed KrF and i-line scanners it says use more than 90% domestic key parts and have processed over 10 million wafers. These are mature-node tools, not EUV (<a href="https://www.trendforce.com/news/2026/11/09/news-chinas-smee-showcases-lithography-systems-with-90-localized-parts-claims-10m-wafers-processed/">Source: TrendForce</a>).</p></li></ul><p><strong>Foundry</strong></p><ul><li><p><strong>Samsung Taylor</strong> &#8212; Samsung&#8217;s Texas fab could start trial runs in mid-to-late December ahead of 2027 mass production, which includes Tesla&#8217;s AI6 chip. Samsung hasn&#8217;t confirmed a date (<a href="https://www.koreaherald.com/article/10929626">Source: The Korea Herald</a>).</p></li><li><p><strong>TSMC Tainan</strong> &#8212; Taiwan plans to spend about NT$30 billion on 50 hectares in Tainan to bring TSMC&#8217;s 3nm fab under construction there into the Southern Taiwan Science Park (<a href="https://udn.com/news/story/7240/9865346">Source: UDN</a>).</p></li></ul><p><mark data-color="rgb(217, 234, 211)" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better!</mark> <a href="http://semidoped.com/poll">Give us your feedback</a></p><p><mark data-color="rgb(207, 226, 243)" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark>:</p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/amd-q3-fy26-annotated">AMD Q3 FY26, Annotated</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/intel-why-didnt-we-see-this-coming">Intel: Why Didn&#8217;t We See This Coming?</a></p></li></ul>]]></content:encoded></item><item><title><![CDATA[Sunday Pub Quiz]]></title><description><![CDATA[After the war, two engineers set up shop in a bombed-out Tokyo department store.]]></description><link>https://daily.semidoped.com/p/sunday-pub-quiz-cb6</link><guid isPermaLink="false">https://daily.semidoped.com/p/sunday-pub-quiz-cb6</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Sun, 08 Nov 2026 14:32:15 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!pSMR!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd1d09112-7a08-4b61-9599-274d200a77cf_1254x1254.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>After the war, two engineers set up shop in a bombed-out Tokyo department store. Their first product was an electric rice cooker: a wooden tub with aluminium electrodes that cooked rice by passing current through it. It mostly burned or undercooked the rice. They then paid Western Electric $25,000 for the right to make Bell Labs&#8217; new transistor. Then in 1955 they sold a pocket radio. Name the company.</p><h3>CLUES:</h3><blockquote><p>The name combines the Latin root for sound, with a common American diminutive for the same word. </p></blockquote><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!5vc8!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg 424w, https://substackcdn.com/image/fetch/$s_!5vc8!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg 848w, https://substackcdn.com/image/fetch/$s_!5vc8!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!5vc8!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!5vc8!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg" width="250" height="187" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:187,&quot;width&quot;:250,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!5vc8!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg 424w, https://substackcdn.com/image/fetch/$s_!5vc8!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg 848w, https://substackcdn.com/image/fetch/$s_!5vc8!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!5vc8!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F581134c9-4388-4324-8ba1-d6605d0d371a_250x187.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div><figcaption class="image-caption">By &#12502;&#12523;&#12540;&#12494;&#12539;&#12503;&#12521;&#12473; - Own work, CC0, https://commons.wikimedia.org/w/index.php?curid=145222506</figcaption></figure></div><blockquote><p>The brand appeared on its radios in 1955, but the company didn&#8217;t take it as its own name until 1958.</p></blockquote><h3>SPOILER ALERT: ANSWER BELOW</h3><div class="pullquote"><h4><strong>Answer: Sony</strong></h4></div>]]></content:encoded></item><item><title><![CDATA[TIL: The First Contract in Silicon Valley Was Ten Dollar Bills]]></title><description><![CDATA[How a letter to a New York securities firm got eight Shockley defectors $1.38 million, and why the deal left them as employees.]]></description><link>https://daily.semidoped.com/p/til-the-first-contract-in-silicon</link><guid isPermaLink="false">https://daily.semidoped.com/p/til-the-first-contract-in-silicon</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Sat, 07 Nov 2026 18:01:41 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/60dfc136-a6a0-469f-8e3a-ef1a2fe2adaf.svg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><a href="https://en.wikipedia.org/wiki/Fairchild_Semiconductor">Fairchild Semiconductor</a>&#8217;s first contract was ten one-dollar bills. In 1957, eight engineers met two New York bankers at the Clift Hotel in downtown San Francisco and agreed to start a company together. Each man signed every one of the ten bills, and for the time being, those bills were the only paperwork the company had.</p><p>The eight engineers worked for <a href="https://en.wikipedia.org/wiki/William_Shockley">William Shockley</a>, and their reasons for leaving started with him. In 1956, the year he shared the Nobel Prize for the transistor, Shockley recruited a group of young PhDs to <a href="https://en.wikipedia.org/wiki/Shockley_Semiconductor_Laboratory">Shockley Semiconductor Laboratory</a>, set up in a former Quonset hut on San Antonio Road in Mountain View. Candidates sat a psychological test before he would interview them, and when a secretary cut her hand on a pin, Shockley suspected sabotage and ordered lie detector tests. The lab&#8217;s direction worried his staff more than his suspicions did, because Shockley kept pulling people off silicon transistors to work on his four-layer diode, a device few of them thought was worth the effort.</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Thanks for reading! Subscribe for free to receive new posts and support our work.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div><p>In May 1957, a group led by <a href="https://en.wikipedia.org/wiki/Gordon_Moore">Gordon Moore</a> took the problem to <a href="https://en.wikipedia.org/wiki/Arnold_Beckman">Arnold Beckman</a>, whose company funded the lab, and asked him to move Shockley into an advisory role under a professional manager. Beckman offered a manager but kept Shockley as director with full authority. After that, the question was no longer whether to leave but how and when.</p><p>Seven of them decided to go, and their first idea was to find an employer that would hire them as a team. <a href="https://en.wikipedia.org/wiki/Eugene_Kleiner">Eugene Kleiner</a> wrote to <a href="https://en.wikipedia.org/wiki/Hayden,_Stone_%26_Co.">Hayden, Stone &amp; Co.</a>, a New York securities firm where his father had an account, to ask whether any company would take them on. The letter reached <a href="https://en.wikipedia.org/wiki/Arthur_Rock">Arthur Rock</a>, a 30-year-old analyst at the firm, who persuaded his boss, <a href="https://computerhistory.org/events/an-evening-with-legendary-venture/">Alfred &#8220;Bud&#8221; Coyle</a>, to fly to California and meet the group.</p><p>Before Rock could pitch the group to anyone, he needed to know who would lead it, since investors would expect one person in charge. The seven had someone in mind, but he was not one of them. <a href="https://en.wikipedia.org/wiki/Robert_Noyce">Robert Noyce</a> had apparently sided with Shockley in the lab&#8217;s disputes, and Shockley considered him his one supporter on staff. <a href="https://en.wikipedia.org/wiki/Sheldon_Roberts">Sheldon Roberts</a> persuaded Noyce to come to the meeting with the bankers anyway, which is how eight men came to be sitting at the Clift.</p><div class="captioned-button-wrap" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/til-the-first-contract-in-silicon?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="CaptionedButtonToDOM"><div class="preamble"><p class="cta-caption">Thanks for reading! This post is public so feel free to share it.</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/til-the-first-contract-in-silicon?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://daily.semidoped.com/p/til-the-first-contract-in-silicon?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p></div><p>At the Clift, Rock and Coyle proposed a different plan. Instead of looking for an employer, the eight would start their own company, and Hayden, Stone would find a large corporation willing to pay for it. Everyone agreed. Coyle took ten new dollar bills from his wallet, asked each man to sign every one, and explained that the bills would be their contracts with each other.</p><p>Finding the backer took far more work than signing the bills, because venture capital did not yet exist as an industry, so Rock drew up a list of about 35 established companies and asked each of them to fund a group of researchers with no product and no experience in running a business. All of them &#8212; rightfully &#8212; said no.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!BEae!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9c3c477-5076-495e-8938-7e62770aa30d_250x89.png 424w, https://substackcdn.com/image/fetch/$s_!BEae!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9c3c477-5076-495e-8938-7e62770aa30d_250x89.png 848w, https://substackcdn.com/image/fetch/$s_!BEae!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9c3c477-5076-495e-8938-7e62770aa30d_250x89.png 1272w, https://substackcdn.com/image/fetch/$s_!BEae!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9c3c477-5076-495e-8938-7e62770aa30d_250x89.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!BEae!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9c3c477-5076-495e-8938-7e62770aa30d_250x89.png" width="250" height="89" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/c9c3c477-5076-495e-8938-7e62770aa30d_250x89.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:89,&quot;width&quot;:250,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!BEae!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9c3c477-5076-495e-8938-7e62770aa30d_250x89.png 424w, https://substackcdn.com/image/fetch/$s_!BEae!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9c3c477-5076-495e-8938-7e62770aa30d_250x89.png 848w, https://substackcdn.com/image/fetch/$s_!BEae!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9c3c477-5076-495e-8938-7e62770aa30d_250x89.png 1272w, https://substackcdn.com/image/fetch/$s_!BEae!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9c3c477-5076-495e-8938-7e62770aa30d_250x89.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div><figcaption class="image-caption">By Fairchild Semiconductor - http://www.fairchildsemi.com/, Public Domain, https://commons.wikimedia.org/w/index.php?curid=26763160</figcaption></figure></div><p>The backer Rock eventually found was <a href="https://en.wikipedia.org/wiki/Sherman_Fairchild">Sherman Fairchild</a>, IBM&#8217;s largest individual shareholder (his father had co-founded the company) and the owner of <a href="https://en.wikipedia.org/wiki/Fairchild_Camera_and_Instrument">Fairchild Camera and Instrument</a>, a Long Island maker of military aerial cameras. Fairchild Camera agreed to lend the new company $1.38 million over 18 months. Each of the eight put in $500 for 100 shares &#8212; about two to three weeks&#8217; salary. Noyce had to write to his parents to ask whether his grandmother could lend him his share. Hayden, Stone took 225 shares, and 300 more were held back for future managers. In return for the loan, Fairchild Camera controlled the company through a voting trust and held an option to buy all of its stock for $3 million.</p><p>With the money arranged, the eight resigned from Shockley&#8217;s lab on September 18, 1957, and became known as the &#8220;<a href="https://en.wikipedia.org/wiki/Traitorous_eight">Traitorous Eight</a>.&#8221; According to one account, Shockley&#8217;s notebook records the day in a single line: &#8220;Group resigns.&#8221;</p><p>Fairchild Semiconductor set up at 844 Charleston Road in Palo Alto and landed its first customer within months. In early 1958, IBM&#8217;s Federal Systems Division needed a high-voltage silicon transistor to drive the magnetic core memory in the B-70 bomber&#8217;s onboard computer, and no established manufacturer could meet the spec. IBM ordered 100 of them from Fairchild at $150 each, and Moore&#8217;s team delivered a double-diffused silicon mesa transistor.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!EoZ9!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!EoZ9!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg 424w, https://substackcdn.com/image/fetch/$s_!EoZ9!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg 848w, https://substackcdn.com/image/fetch/$s_!EoZ9!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!EoZ9!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!EoZ9!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg" width="344" height="498.112" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:362,&quot;width&quot;:250,&quot;resizeWidth&quot;:344,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!EoZ9!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg 424w, https://substackcdn.com/image/fetch/$s_!EoZ9!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg 848w, https://substackcdn.com/image/fetch/$s_!EoZ9!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!EoZ9!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3bb0dc3e-bb63-4bf2-a9d0-447946f22a00_250x362.jpeg 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The first advertisement of a silicon transistor made in Silicon Valley, by Fairchild Semiconductor in 1958. The ad appeared in various publications; this scan is from <em>Electronics</em>, August 15, 1958. The 2N696 and 2N697 transistors described differ only in their current gain (2N696 15 to 30, 2N697 30 or higher). By Unknown; presumably the copyright, if any, was owned by Fairchild Semiconductor Corporation. - https://www.americanradiohistory.com/Archive-Electronics/50s/Electronics-1958-08-15.pdfone page out of this online magazine scan, Public Domain, https://commons.wikimedia.org/w/index.php?curid=72050388</figcaption></figure></div><p>Orders like that made the $3 million option look cheap to Fairchild Camera, and it exercised the option in 1959. Each founder&#8217;s $500 stake became about $250,000 in Fairchild Camera stock, and each founder became an employee of a division run from New York.</p><p>Under New York&#8217;s ownership, Fairchild Semiconductor kept losing people. By later accounts, the parent company would not extend stock options beyond the founders and moved semiconductor profits into its other businesses. Engineers left to start their own firms, and Fairchild&#8217;s founders and employees went on to create at least 65 companies over the next two decades. In 1968, Noyce and Moore left too. They called Arthur Rock, who had been doing venture deals from San Francisco since 1961, and he raised $2.5 million to start Intel.</p><p>Shockley&#8217;s one loyalist became general manager of the company the defectors started, and then he defected from that one too.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://daily.semidoped.com/p/til-the-first-contract-in-silicon/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://daily.semidoped.com/p/til-the-first-contract-in-silicon/comments"><span>Leave a comment</span></a></p><p></p>]]></content:encoded></item><item><title><![CDATA[Daily Update - November 6th, 2026]]></title><description><![CDATA[A midterm-driven AI rout, Taiwan&#8217;s market rescue fund, and a packaging build-out that didn&#8217;t blink]]></description><link>https://daily.semidoped.com/p/daily-update-november-6th-2026</link><guid isPermaLink="false">https://daily.semidoped.com/p/daily-update-november-6th-2026</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Fri, 06 Nov 2026 13:28:34 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/8a7950e3-9dd3-49b7-8ede-03b111bffd1d_1200x630.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>Hello world, its Friday, November 6th, 2026</strong></p><p>Democrats won Congress, and the market decided data centers were next on the ballot. The Nasdaq had its worst day in 19 months, Taiwan summoned its rescue fund, and TSMC briefly hit limit-down. Meanwhile ASE wrote a $3.7 billion packaging check and an optics maker signed a billion-dollar deal on the same day. Fear and orders are moving in opposite directions.</p><p><em>Let&#8217;s get into it. --Austin &amp; Vik</em></p><p><mark data-color="#cfe2f3" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Latest episode</mark>: <strong>3D Stacked DRAM for Inference: d-Matrix&#8217;s Sudeep Bhoja</strong> &#8212; listen on <a href="https://www.youtube.com/watch?v=zEqoBSJ4YQ4">YouTube</a>, <a href="https://podcasts.apple.com/podcast/id1866707196">Apple Podcasts</a>, or <a href="https://open.spotify.com/show/0Uuu3s1Nw09f6Xmg24rCZm">Spotify</a>. Full transcript: <a href="https://daily.semidoped.com/p/new-episode-d-matrix-sudeep-bhoja">daily.semidoped.com</a>.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p><mark data-color="#d9d2e9" style="background-color: rgb(217, 210, 233); color: rgb(0, 0, 0);">Fine-tune and serve open models without managing a single GPU.</mark></p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. </p><p>Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p><strong>Accounts are free, and get $5 in credit when you sign up.</strong></p><p><a href="https://semidoped.com/crusoe-daily">Try Serverless Fine-Tuning</a></p></div><h3>Wall Street prices in a data center slowdown</h3><p>The Nasdaq Composite fell 5.5% on Thursday, its worst day since April 2025&#8217;s tariff shock, and slipped into correction territory, nearly 12% below its October 6 record. The Philadelphia Semiconductor Index dropped about 12% to a six-month low. Nvidia lost 9%, AMD 17%, and SanDisk 22%. The Nasdaq 100 shed roughly $2.5 trillion in value.</p><p>The trigger was political. Democrats won both chambers of Congress and flipped governorships in Texas and other data center states, and traders concluded that permitting, power, and local opposition will slow U.S. buildouts. UBS estimated that about 30% of expected 2028 U.S. data center capacity additions could be at risk from construction pauses. Reuters also noted the AI trade was already shaky: neocloud financing costs have climbed, and CoreWeave&#8217;s five-year credit default swaps hit a record.</p><p><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Analysts are pricing in a slowdown in capacity that has not been cancelled, and the reaction is running well ahead of anything Congress has actually done.</mark> Nothing has been passed. By early Friday, Nvidia, AMD, Intel, Broadcom and Micron were bouncing 4&#8211;8% in premarket trading.</p><p><a href="https://za.investing.com/news/stock-market-news/nasdaq-100-eyes-18-trillion-loss-as-democratic-win-sparks-ai-delay-fears-4555558">Source: Reuters</a> &#183; <a href="https://www.theguardian.com/business/2026/nov/05/nasdaq-heading-for-worst-day-in-18-months-as-wall-street-sustains-widespread-losses">The Guardian</a> &#183; <a href="https://www.investing.com/news/stock-market-news/nasdaq-futures-rise-after-steep-ai-selloff-as-chip-stocks-stabilize-5007026">Reuters (Friday premarket)</a></p><blockquote><p><strong>Vik: </strong><em>The election results are a scare for the datacenter buildout. Definitely not worth the massive shake up. Until something passes from Congress about actually limiting buildouts, they are only election promises for now. We know how those things go. If it does pass, then we will see a trajectory of growth that tries to squeeze the performance of what we have already. There may be some caveats in the latter scenario.</em></p><p><strong>Austin: </strong><em>If there happens to be a slow down, it&#8217;s a boon for air-cooled CPU/GPU/AI ASIC offerings.</em></p></blockquote><h3>Asia takes the second leg, and Taiwan calls in the cavalry</h3><p>Asia opened Friday into the same fear. Korea halted KOSPI trading for 20 minutes after an 8% plunge, then closed down 6.02%. SK hynix ended 10.4% lower and Samsung 6.35% lower. Tokyo&#8217;s Advantest fell 12.2% and SoftBank 11.6%. In Taipei, TSMC opened at its 10% limit-down price, then clawed back to close down 4.69% at NT$2,135. The Taiex finished down 4.51% at 39,701.64 after a low of 37,837.</p><p>By Friday afternoon, Taiwan&#8217;s National Stabilization Fund had authorized its executive secretary to deploy money, the tenth time it has done so, with up to NT$500 billion available. It said the Taiex was down 12.19% on the week and about 20.5% from its 2026 high. Foreigners sold a record NT$994 billion of Taiwan stocks in October and another NT$524 billion this month. The fund didn&#8217;t say how much, if any, it had bought.</p><p><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">Taiwan&#8217;s regulators are treating a chip selloff as a financial stability event, which tells you how much of that market is one sector.</mark> TSMC alone is more than 40% of Taiwan&#8217;s market capitalization. A Daishin analyst quoted by Yonhap called the reaction &#8220;excessive&#8221; and said a full halt to U.S. data center buildouts is unlikely.</p><p><a href="https://www.cna.com.tw/news/afe/202611060270.aspx">Source: CNA (stabilization fund)</a> &#183; <a href="https://focustaiwan.tw/business/202611060016">Focus Taiwan</a> &#183; <a href="https://en.yna.co.kr/view/AEN20261106007852320">Yonhap</a> &#183; <a href="https://www.cnbc.com/2026/11/06/tech-stocks-sk-hynix-samsung-softbank.html">CNBC</a></p><blockquote><p><strong>Vik: </strong><em>Yeah, remember this month as Black November for AI. Nov 5th, 2026 to be exact. Everything turned deep red, and investors rotated out of AI semis.</em></p></blockquote><h3>TrendForce&#8217;s 2027 forecast arrives in the worst possible week</h3><p>At its Taipei forum on Thursday, TrendForce forecast that AI server shipments will grow 36% in 2027 to about 3.83 million units, ahead of 2026&#8217;s pace. It sees combined capex at the nine biggest cloud providers rising 60% to about $1.52 trillion. Nvidia should keep roughly 62.6% of the AI server market, and Vera Rubin racks could pass the GB series in Q2 2027. AMD&#8217;s first rack-scale design, Helios, could see order delays like those that hit Nvidia&#8217;s ramps.</p><p>The forecast carries a caveat that matters this week: TrendForce said it could be revised for macro or geopolitical shifts. It was prepared before the political shock hit the market.</p><p><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The most-cited industry forecast for 2027 was finalized before the selloff, so the real test is whether hyperscaler capex plans survive the next round of permitting fights.</mark> Early reads will come from the next round of cloud capex commentary and from permitting decisions in the states that just flipped.</p><p><a href="https://www.trendforce.com.tw/presscenter/news/20261105-13295.html">Source: TrendForce</a> &#183; <a href="https://www.taipeitimes.com/News/biz/archives/2026/11/06/2003867228">Taipei Times</a> &#183; <a href="https://www.trendforce.com/presscenter/news/20261014-13269.html">TrendForce (Oct. capex outlook)</a></p><blockquote><p><strong>Vik: </strong><em>Talk about timing a report badly &#8212; right in the middle of $2.5T leaving the station. LOL Trendforce.</em></p></blockquote><h3>ASE commits $3.7 billion to packaging as Nvidia reportedly funds SPIL</h3><p>Taiwan&#8217;s Ministry of Economic Affairs approved ASE&#8217;s plan to invest NT$123.7 billion (about $3.7 billion) in new automated packaging and test lines in Kaohsiung and Tainan, adding 3,553 local jobs. It is the largest case under the government&#8217;s Invest Taiwan program in recent memory, and the lines are aimed at AI and high-performance computing demand. ASE shares still fell 7.9% on Friday with the rest of the market.</p><p>Separately, Reuters reported, citing an industry executive, that Nvidia is providing billions of dollars to expand CoWoS-class capacity at SPIL, an ASE unit. The funding is for new factory space and equipment. Neither Nvidia nor ASE has confirmed it. The same Reuters piece says advanced packaging revenue should pass traditional packaging for the first time in 2026, with capex near $28 billion (Yole Group).</p><p><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The scramble for packaging capacity is outside TSMC now, and the biggest customer is helping pay for it.</mark> The ASE approval is on the record. The Nvidia funding rests on a single anonymous executive.</p><p><a href="https://www.taiwannews.com.tw/news/6472728">Source: Taiwan News / CNA (MOEA approval)</a> &#183; <a href="https://chipdispatch.com/articles/ase-puts-nt-123-7-billion-behind-ai-packaging-push-in-southern-taiwan">Invest Taiwan details via Chip Dispatch</a> &#183; <a href="https://www.investing.com/news/stock-market-news/ai-boom-turns-chip-packaging-into-the-next-battleground-for-semiconductors-5006549">Reuters</a></p><blockquote><p><strong>Vik: </strong><em>Look past the AI doom and gloom seriously. There is still packaging capacity being built out. Just because the stock market had a rout does not mean we stop making chips overnight.</em></p></blockquote><h3>Applied Optoelectronics books a 1.6T deal, with a bigger one &#8220;within weeks&#8221;</h3><p>AOI reported record Q3 revenue of $269.9 million, more than double last year, and a return to GAAP profit. It also disclosed an October three-year agreement with an existing Tier 1 hyperscaler covering 1.6T transceivers, which the customer has qualified. The customer agreed to buy at least 10% of its 1.6T needs from AOI from June 2026 through May 2029, worth an expected $1&#8211;1.5 billion. Q4 revenue is guided to $385&#8211;420 million.</p><p>There are strings. AOI committed to agreed pricing that includes some price cuts, the customer made no capital commitment, and CFO Stefan Murry said the locked pricing could cap upside if market prices rise. Capacity is also slipping: the U.S. expansion will take longer because of power and construction delays. <mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The CEO said a second agreement, </mark><strong><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">potentially five times larger</mark></strong><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);"> and including 800G and lasers, is in its final stages and could come within weeks</mark>. It has not been signed.</p><p>The first deal is real, signed and disclosed, while the second, bigger one is a promise from the CEO. AOI also said it is qualifying CPO/NPO lasers with four to five customers and that lasers carry margins above 60%.</p><p><a href="https://investors.ao-inc.com/2026-11-05-Applied-Optoelectronics-Reports-Third-Quarter-2026-Results">Source: AOI Q3 2026 release</a> &#183; <a href="https://www.fool.com/earnings/call-transcripts/2026/11/06/applied-optics-aaoi-q3-2026-earnings-call-transcript/">Earnings call transcript (Motley Fool)</a> &#183; <a href="https://www.marketbeat.com/instant-alerts/transcript-what-applied-optoelectronics-aaoi-said-on-its-q3-earnings-call-2026-11-05/">MarketBeat</a></p><blockquote><p><strong>Vik: </strong><em>Applied Optoelectronics stock has been beaten down more than most. The underlying fundamentals seem okay; not sure why the hard hit.  </em></p><p><strong>Austin: </strong><em>Guide was down&#8230; significantly compared to earlier-in-the-year expectations. Ramp is slow, margins not expanding. </em></p></blockquote><div class="callout-block" data-callout="true"><h4>Supported by Meter</h4><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks?</p><p><strong>Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS.</strong></p><p>One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://www.meter.com/?utm_source=semidoped&amp;utm_medium=podcast&amp;utm_campaign=semidoped-q4-2026">See how Meter builds data center networks. </a></p></div><h3>NTT DATA lines up 900 MW of onsite power</h3><p>NTT DATA agreed to deploy 900 MW of Bloom Energy fuel cells across multiple global data center projects, with initial shipments expected in 2027. Bloom calls it its largest international project. The companies did not disclose a price or a site-by-site breakdown.</p><p>The fuel cells will serve as primary power, generating electricity at the data centers rather than relying entirely on a grid connection. That addresses one constraint behind the slowdown fears: getting enough power to a new campus.</p><p><mark data-color="#fce5cd" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">The deal provides a concrete example of operators finding another route to power, even as investors question the pace of data center construction.</mark></p><p>Source: <a href="https://www.nttdata.com/en-us/news/2026/ntt-data-announces-agreement-with-bloom-energy">NTT DATA</a> &#183; <a href="https://www.bloomenergy.com/news/ntt-data-announces-agreement-with-bloom-energy-for-900-mw-of-onsite-power-generation-across-multiple-data-center-projects/">Bloom Energy</a></p><h2>Industry Chatter</h2><p>These are just unconfirmed industry whispers.</p><ul><li><p><strong>Anthropic run rate and IPO</strong> &#8212; Anthropic&#8217;s annualized revenue topped $80 billion at the end of October, up from roughly $65 billion in July, according to the Wall Street Journal. IPO investors are pushing back on a $1.5 trillion valuation, with some favoring about $1 trillion or a smaller raise, and the selloff makes a delay a live option. Anthropic is one of the biggest buyers of compute in the market, so its listing timeline matters to the chip supply chain (<a href="https://www.wsj.com/finance/major-investors-are-skeptical-anthropic-ipo-can-achieve-1-5-trillion-valuation-ca965d8b">WSJ via X</a>).</p></li><li><p><strong>KYEC / Rubin / HBM4</strong> &#8212; A post from @jukan05 links a major-GPU delay at KYEC to HBM4 yield problems, with Nvidia&#8217;s Rubin named as the affected chip. That puts the memory ramp back in focus as a possible constraint on GPU shipments. <a href="https://x.com/jukan05/status/2118667945455001937">Source</a></p></li></ul><h2>Sector Watch</h2><p><strong>Packaging &amp; Substrates</strong></p><ul><li><p><strong>JNTC</strong> &#8212; The Elec reports that Korean glass substrate maker JNTC was registered in October as a through-glass-via supplier to Nvidia, its second major U.S. chip customer after Broadcom, and is shipping paid 2mm glass-core samples. The company said it can&#8217;t discuss customers, and it plans pilot-line samples through 2027 and initial volume from mid-2028 (<a href="https://www.thelec.kr/news/articleView.html?idxno=64955">Source: The Elec</a>).</p></li><li><p><strong>Powertech</strong> &#8212; In the Reuters packaging feature, Powertech says it will invest $400 million in an advanced packaging joint venture with Broadcom in Singapore, with production possible in 2028, and is betting on panel-level packaging (<a href="https://www.investing.com/news/stock-market-news/ai-boom-turns-chip-packaging-into-the-next-battleground-for-semiconductors-5006549">Source: Reuters</a>).</p></li></ul><p><strong>Optics</strong></p><ul><li><p><strong>Santec</strong> &#8212; The Japanese optical test equipment maker reported first-half operating profit of &#165;7.7 billion and raised full-year guidance to &#165;50 billion in revenue and &#165;18 billion in operating profit, citing data center optics testing demand (<a href="https://japanir.jp/wp-content/uploads/2026/11/6777-20261106-02.pdf">Source: Santec filing</a> &#183; <a href="https://en.kabutan.com/jp/news/k202611060213">Kabutan</a>).</p></li></ul><p><strong>Memory &amp; Materials</strong></p><ul><li><p><strong>YMTC</strong> &#8212; Reuters, citing two people, reports that YMTC wants 60&#8211;70% domestic equipment for a planned fourth Wuhan fab making 300-plus-layer NAND for data centers. It would be up from just over half at the third fab. Construction is expected to start next year and no orders have been placed (<a href="https://www.devdiscourse.com/article/international/4002430-exclusive-chinas-top-flash-memory-chipmaker-targets-more-homegrown-tools-for-new-factory-sources-say">Source: Reuters</a>).</p></li></ul><p><strong>AI Infrastructure</strong></p><ul><li><p><strong>Neocloud credit</strong> &#8212; Reuters notes that CoreWeave&#8217;s five-year credit default swaps hit a record 1,267.9 basis points and Intel&#8217;s hit a record 149.7. Both CoreWeave and Nebius have lost half or more of their value since October 6 (<a href="https://za.investing.com/news/stock-market-news/nasdaq-100-eyes-18-trillion-loss-as-democratic-win-sparks-ai-delay-fears-4555558">Source: Reuters</a>).</p></li><li><p>OpenAI &#8212; Its new National Community Compact commits future campuses it owns and operates to cover their electricity and infrastructure costs, disclose water and noise impacts, and avoid NDAs with public officials, subject to narrow security and trade-secret exceptions. <a href="https://openai.com/index/setting-new-standard-responsible-ai-infrastructure/">Source</a></p></li></ul><p><strong>Equipment</strong></p><ul><li><p><strong>Lam Research</strong> &#8212; The board authorized another $25 billion in share repurchases, with no termination date; the authorization does not obligate Lam to buy the shares. <a href="https://investor.lamresearch.com/2026-11-05-Lam-Research-Corporation-Announces-25-Billion-Share-Repurchase-Authorization-and-Declares-Quarterly-Dividend">Source</a></p></li></ul><p><strong>Power</strong></p><ul><li><p><strong>Infineon / SOLUM</strong> &#8212; The companies signed a memorandum of understanding to develop high-voltage direct-current power systems for AI data centers, pairing Infineon semiconductors and reference designs with SOLUM&#8217;s power-supply design and manufacturing. <a href="https://www.infineon.com/technology-news/2026/inftn202611-027">Source</a></p></li></ul><p><mark data-color="#d9ead3" style="background-color: rgb(217, 234, 211); color: rgb(0, 0, 0);">Tell us how to make Semi Doped better!</mark> <a href="http://semidoped.com/poll">Give us your feedback</a></p><p><mark data-color="#cfe2f3" style="background-color: rgb(207, 226, 243); color: rgb(0, 0, 0);">Also worth a read</mark>:</p><ul><li><p>Austin&#8217;s latest on Chipstrat, <a href="https://newsletter.chipstrat.com/p/when-agents-are-the-interface">When Agents Are the Interface</a></p></li><li><p>Vik&#8217;s latest on Vik&#8217;s Newsletter, <a href="https://www.viksnewsletter.com/p/intel-why-didnt-we-see-this-coming">Intel: Why Didn&#8217;t We See This Coming?</a></p></li></ul>]]></content:encoded></item><item><title><![CDATA[🎙️ NEW EPISODE: d-Matrix - Sudeep Bhoja]]></title><description><![CDATA[Watch now (76 mins) | Raptor's 3D stacked DRAM, 144 chips in an Nvidia Vera Rubin rack, 0.37 picojoules per bit, speculative decoding, and more]]></description><link>https://daily.semidoped.com/p/new-episode-d-matrix-sudeep-bhoja</link><guid isPermaLink="false">https://daily.semidoped.com/p/new-episode-d-matrix-sudeep-bhoja</guid><dc:creator><![CDATA[Semi Doped]]></dc:creator><pubDate>Thu, 05 Nov 2026 20:23:36 GMT</pubDate><enclosure url="https://substack-post-media.s3.amazonaws.com/public/images/c438736d-cf54-401b-b9a5-49a4ac28a687_2000x1116.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Vik Sekar talks with Sudeep Bhoja, co-founder and CTO of d-Matrix, about the company&#8217;s journey from analog in-memory compute to its current digital approach. Sudeep explains the architecture of the Corsair and Raptor chips, detailing how 3D stacked DRAM achieves SRAM-like bandwidth with much greater capacity and power efficiency. They also discuss how d-Matrix integrates into the Nvidia ecosystem to build multi-rack systems for massive AI models.</p><div class="callout-block" data-callout="true"><h3>Presented by Crusoe</h3><p><strong>Fine-tune and serve open models without managing a single GPU.</strong></p><p>When we sat down with Crusoe&#8217;s product team, Austin asked: what will you do with my data? Their answer: They don&#8217;t touch or see it.</p><p>Your training data is encrypted so even Crusoe&#8217;s own admins can&#8217;t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere.</p><p>Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best.</p><p>Accounts are free, and <mark data-color="rgb(252, 229, 205)" style="background-color: rgb(252, 229, 205); color: rgb(0, 0, 0);">get $5 in credit</mark> when you sign up.</p><p><strong><a href="https://semidoped.com/crusoe">Try Serverless Fine-Tuning &#8594;</a></strong></p></div><p><strong>Things we cover:</strong></p><ul><li><p>From analog to digital in-memory compute</p></li><li><p>The Corsair PCIe card&#8217;s SRAM and LPDDR</p></li><li><p>Raptor&#8217;s 3D stacked DRAM architecture</p></li><li><p>Speculative decoding and workload disaggregation</p></li><li><p>Building multi-rack systems with Nvidia&#8217;s Vera Rubin</p></li><li><p>The future Lightning architecture with four-high DRAM stacks</p></li></ul><div class="callout-block" data-callout="true"><p><strong>This episode is supported by Meter</strong>. </p><p>You have the GPUs and the power. Now who builds the network, and who do you call when it breaks? Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS. One call reaches the team that designed the hardware and wrote the software.</p><p><a href="https://semidoped.com/meter">See how Meter builds datacenter networks</a><strong><a href="https://semidoped.com/meter">&#8594;</a></strong></p></div><p><em>This podcast is lightly edited for clarity.</em></p><h2>The Inference-First Approach</h2><p><strong>Welcome to the Semi Doped podcast where we have conversations with some of the best minds in the semiconductor industry. I&#8217;m Vik Sekar and I&#8217;ve worked in the semiconductor industry for like 20 years, and today I&#8217;m an independent researcher and the founder of Semi Exponent. I also write a substack newsletter called Vic&#8217;s newsletter. Welcome to the show, Sudeep. How are you?</strong></p><p><strong>Sudeep</strong>: Good, good. It&#8217;s a pleasure to be on your show, Vic. I&#8217;ve always been a fan of all your work, and it&#8217;s great to be on your podcast.</p><p><strong>Thank you for coming on. We have so much to chat about today because I am quite familiar with d-Matrix. I&#8217;ve covered it on the substack before and we&#8217;ve spoken at conferences. There&#8217;s so much exciting technology to get into, so we should just get right into it. Let&#8217;s start right at the beginning because the story here is very interesting and that&#8217;s because d-Matrix actually was founded in 2019 and that was well before the AI thing even took off. You and your CEO, Sid, were both at Infi before starting d-Matrix. And it&#8217;s really interesting because right after that, actually Marvell acquired Infi. So that was really cool to see how you guys decided that you&#8217;re going to go and do inference architecture that early in the boom when something as exciting as Marvell&#8217;s acquisition of Infi was in the works. So how did you foresee all of this happening?</strong></p><p><strong>Sudeep</strong>: Of course, we did not know the acquisition was in the works at the time. But we were working at Marvell in the data center business unit and the interconnect products we were building were being used by the hyperscalers&#8212;Google, Microsoft, Facebook at that time, Amazon. And we were seeing that the AI training clusters were pulling our interconnect solutions. And they were in fact driving speeds and feeds, which was a change from what it used to be before that. And then we realized that they were training these models, these amazing models. ImageNet had come out in 2012 and there was breakthroughs after breakthroughs already. And late 2017, the transformer paper had been released.</p><p>We decided to focus on inference because we realized that the problems of inference were very, very different from what everyone else was working on at that time, which was training models. Training models is about high batches, compute, right? Because it&#8217;s a compute-bound problem. You want to pack as much compute as possible. Whereas inference is a memory-bound problem. Latency is very critical. And so, that&#8217;s what we focused on.</p><p><strong>That is a unique decision to make at that time because if you actually see all the other, I&#8217;d say pre-ChatGPT companies, whether it&#8217;s Groq or Cerebras, all of them went for the training problem first because all of them wanted to go and train these models. And I don&#8217;t think at that time the industry realized how big these models are going to get and how many chips are going to be required. So a lot of them did start with the SRAM-based approach, which actually d-Matrix also has in their technology and we&#8217;ll get to that. But clearly you went straight for inference. So that is a unique problem that you identified early on for sure.</strong></p><p><strong>Sudeep</strong>: Yes. Specifically, we recognized that once you train these models, you have to run inference on them. And the characteristics of inference was very different from training. To do an inference, you have to read all of the weights for every token and read all of the user&#8217;s KV cache. So if you want to keep latency small, then you can&#8217;t batch too many users. And that means that the memory subsystem is very pressured. There is not sufficient memory bandwidth to feed any sort of compute. And having worked in the semiconductor industry for many years, we recognized that the compute itself was very cheap. It was in the femtojoules per operation. The data to feed the compute that had to come from HBM or other memory subsystems was in the picojoules. So there was this three orders of magnitude gap, which is the technical problem that we wanted to focus on and solve. And that turned out to be a very good match for inference.</p><h2>From Analog to Digital In-Memory Compute</h2><p><strong>So that power delta that you identified told you that this is the problem you need to solve. So who was the first person you convinced that this problem is worth solving? Who decided to fund the effort first?</strong></p><p><strong>Sudeep</strong>: The initial check was written by Infi, the company we worked at before we started d-Matrix. So they were the first investor in d-Matrix in the seed round. And then on series A, we had Microsoft M12 and Playground Global as our lead investors.</p><p><strong>That&#8217;s fascinating. It&#8217;s fascinating you could get Infi as your first investors. It&#8217;s amazing. But your first chip looked nothing like what your technology looks today, right? So what was the first kind of chip that you went after as soon as d-Matrix was founded?</strong></p><p><strong>Sudeep</strong>: We wanted to build a proof-of-concept silicon before we scaled the company in our seed round. So we essentially went after trying to co-locate compute and memory together. After many years of transistor scaling, we knew that the compute is very small, but the wires had not scaled at the same rate. So it was data movement that was a problem. Our initial idea was, can we take an SRAM array and add compute into the SRAM array itself? So unlike a traditional SRAM where we only activate one word line at a time, we ended up designing an array where we activated all of the word lines. That was a pretty good operation for a matrix calculation, matrix-vector specific calculation. And then we activated, instead of the bit line, we used an analog scheme there using capacitors to do the accumulation of the bit line.</p><p><strong>Okay, awesome. So, just for some background, SRAM generally is arranged, let&#8217;s say in a grid, and you can activate it with word lines and bit lines. You can store something in an SRAM cell by applying the right voltages at these levels and reading out with the right sense amplifiers on either side and you can get what bit was stored or what was not stored. That&#8217;s typically how SRAM is used to store information. It&#8217;s been around a very long time. Typically there are six transistors that build up this unit cell which you access with bit lines and word lines. That&#8217;s the basic SRAM approach. But you were not using SRAM as a storage unit as much as you were trying to compute with it, right? You were trying to do analog compute using this six-transistor SRAM cell.</strong></p><p><strong>Sudeep</strong>: Yes. The weights were being stored on the SRAM cell. And then we added a few transistors to convert that into a multiplying bit cell. So if you insert on the word line, if you were to send activations one bit at a time, then you could essentially multiply right at the place where the data is stored. So there&#8217;s no movement of data. And then instead of the bit line, we ended up putting capacitors on top of the SRAM array and shorting the bit line. And the capacitors would essentially do the addition.</p><p><strong>Awesome. So at the heart of all of this is the multiply and add function, the MAC operation that you&#8217;re doing, but you&#8217;re doing it now in the initial version of the d-Matrix, the first version, it was all analog. And it was done by holding the weights in the SRAM cell and using an extra transistor to do the multiplication and then a capacitor to do the charge addition of all of this. And then the result you read it out and then you have your MAC operation, which means now you can multiply vectors together.</strong></p><p><strong>Sudeep</strong>: That&#8217;s correct. That&#8217;s exactly correct. In the original versions, we have digitized a lot of it. Obviously, the storage is binary, so it&#8217;s a bit-serial, bit-parallel architecture. We have removed the DAC function that traditional in-memory compute circuits had by sending one bit at a time. And only the accumulation was analog. That was the first test chip that we built, the proof-of-concept chip. But that proof-of-concept chip was a full quad, if you will, of the Corsair product itself. It had a RISC-V processor, it had all of the architecture built around these compute engines. It had a software stack, it had a runtime. So we were able to prove out the whole stack using these basic fundamental building block elements that did matrix-vector multiply, SIMD operations, reshapes, data movement, and so on.</p><p><strong>What was the first version of this chip called?</strong></p><p><strong>Sudeep</strong>: It was called Nighthawk. That was the chip that we taped out in our seed stage of the company with very few people. And then we realized along the way that it was very difficult to maintain precision of the bit line. Customers are very used to seeing 32 bits of precision. And so we had to maintain that level of fidelity. And occasionally we would see a bit error that we worked very hard to correct, but it was very hard to maintain 32 bits of precision across parts when you scale this product. And that&#8217;s the reason that we went down later on on Corsair, we went down the digital in-memory compute approach. What we did is we just took that analog bit line and digitized that bit line using an adder tree. And we had worked on low-voltage circuits and we looked at the energy consumption of that versus analog and then we decided that it was far easier to productize the digital in-memory compute.</p><h2>The Corsair Architecture</h2><p><strong>Okay, that&#8217;s a perfect segue to go into the next product you have because clearly when customers wanted that level of precision, you can&#8217;t do it with analog compute. Today, the fundamental method of computing you have is digital, right? So now you have digital in-memory compute. Can you walk us through what the transition is now so that you have it all digital? How that looks?</strong></p><p><strong>Sudeep</strong>: The matrix multiply units are exactly the same, which is you have the SRAM array and as part of Corsair, we have all of the word lines that we activate and we perform this bit-serial, bit-parallel multiply and add circuitry in the tensor engine piece of it. And instead of an analog accumulation, we put an adder tree and that&#8217;s how we are able to perform matrix multiply in digital and we get to very, very good energy efficiency by doing that.</p><p><strong>I see. Okay, that&#8217;s good. So now your entire compute is still digital in-memory compute, but now you use entirely digital techniques, so you get the kind of accuracy that people were looking for.</strong></p><p><strong>Sudeep</strong>: Correct. We got to 32 bits of accuracy, so we maintain FP32 precision on the accumulation data path. So it&#8217;s just like any other digital engine from that perspective. It&#8217;s just that the computation happened right at the bit cell. So there&#8217;s no moving of data inside the tensor engine. And for that reason, those tensor engines are very efficient. They get to 25 to 40 TOPS per watt, depending on the voltage and the frequency we choose. Some of this we presented at Hot Chips last year.</p><p><strong>Nice. I had one question though, because if you have FP32 and that was probably the initial implementation of what people thought inference should look like. But today we have other formats. We can do FP16, FP8, FP4. Do we still need that level of accuracy when it comes to these operations? Or could you get away with something far less accurate?</strong></p><p><strong>Sudeep</strong>: Even on Corsair, the operands we supported block floating point format. So we supported something called MX int8 and MX int4. So those are like 8-bit and 4-bit formats that shared an exponent across a whole vector. Now it&#8217;s part of an OCP standard. So the precision we support is 8 bits and 4 bits. But only the accumulation. The activations could be in 8 bits, the weights are in 8 bits, but when you add a lot of them up, the accumulation is now in 32 bits. That&#8217;s what we used. So we certainly don&#8217;t need 32 bits for weights and activations.</p><p><strong>I see. Okay, that&#8217;s a good distinction to make. The thing with this digital approach, which is now your Corsair platform, is still completely SRAM-based, right? And so what that fundamentally puts you in is now in the class of the other SRAM-based accelerators, which all have a fundamental limitation and that is capacity. So how much capacity does Corsair have and what is the implication of that limitation?</strong></p><p><strong>Sudeep</strong>: Here&#8217;s Corsair. This is a PCIe card and we have two gigabytes of SRAM on this card.</p><p><strong>So that Corsair PCIe card you&#8217;re holding up essentially has not just one chip in it, right? Like how many chips does it actually have?</strong></p><p><strong>Sudeep</strong>: We have, if you take all the sheet metal out, this is what the card looks like. So there are two chips. Here we have taken out the lid of the other chip and you can see the four chiplets on the second chip. Each of the chiplets is half-reticle. So there&#8217;s a sum total of two reticles on each chip. And together they form two gigabytes of SRAM capacity. But more importantly, they can deliver 150 terabytes per second of memory bandwidth, which is orders of magnitude more than what is available using HBM. Even Blackwell GPUs are 18 terabytes per second and this is 150. So this can deliver a lot more bandwidth because it&#8217;s all on-chip. And at the back of the card, you can see all of the LPDDRs. So that&#8217;s what we use to store as bulk storage.</p><p><strong>Is it the black chips that are the LPDDR in that image?</strong></p><p><strong>Sudeep</strong>: Correct. Black chips are the LPDDRs.</p><p><strong>Okay, so that&#8217;s interesting that you have both SRAM and LPDDR and we&#8217;ll get to what the benefit of that is. If you have two reticles worth, I&#8217;m guessing this is a TSMC leading-edge process. What node is it built on? The Corsair.</strong></p><p><strong>Sudeep</strong>: TSMC 6 nanometer.</p><p><strong>6 nanometers. Okay. That&#8217;s a good node in a sense SRAM kind of has stopped scaling much beyond that. So if you go down to like maybe two nanometers or three nanometers, you don&#8217;t get as much SRAM capacity anyway and then you get the benefits of better wafer cost at 6 nanometers and therefore you still get 2GB SRAM out of this.</strong></p><p><strong>Sudeep</strong>: That&#8217;s exactly correct. The SRAM roadmap like you pointed out has stopped scaling. You only see maybe 10% better density as you keep shrinking process nodes from 6 nanometers to 4 nanometers, three and two and so on. So 6 nanometer is a very sweet spot in terms of SRAM capacity.</p><h2>Disaggregation and Use Cases</h2><p><strong>It&#8217;s useful to put what this PCIe card is capable of because it has only 2GB of SRAM. Clearly to run a model that&#8217;s bigger than that, you&#8217;re going to have to put more of these cards together, right?</strong></p><p><strong>Sudeep</strong>: Yes. Exactly. We use it for many different applications. We build up whole racks. We take a server with eight of these cards, for example. We have a Supermicro X14 server with eight of these cards. And we also have low-latency networking that we built called Jetstream. And we use that to build up an entire rack that we call Squadrack with an Arista top-of-the-rack switch or any Ethernet top-of-the-rack switch that can host the entire model, all in SRAM. So per rack, we can get 128 gigabytes of SRAM.</p><p><strong>That&#8217;s pretty good. So for sure you can run probably not like the biggest models out there at 128GB, but you can run some quite capable models at blazing fast speeds in a single rack.</strong></p><p><strong>Sudeep</strong>: Correct. Exactly right. There are different parts of the workload that we disaggregate and run on our cards. So we think of our deployments as heterogeneous deployments. There&#8217;s still a GPU. And we accelerate the GPU, we turbo-boost the GPU. And for example, one use case is something called speculative decoding where we have a draft model that guesses what the target model running on the GPU is and helps accelerate the GPU&#8217;s target. For example, the GPU could be running Kimi K3 or GLM 5.2, a very large model. But since the GPU is running at one token at a time, it&#8217;ll be relatively slow. But if a drafter that&#8217;s running on Corsair on the SRAM can guess, let&#8217;s say eight or 16 tokens at a time and provide that guess to the GPU, the GPU can verify that in a single shot. So that&#8217;s the speculative decoding use case that we have talked about.</p><p>The other application is something called AFD, attention feed-forward disaggregation, where we take the experts&#8212;the experts are very memory-bound&#8212;and we move the experts to Corsair, to the SRAM chips, but we leave the attention and the pre-fill on the GPU. So that&#8217;s a second use case. And the third use case is you leave the pre-fill on the GPU and only have Corsair cards to decode. So there are different applications, each one disaggregating a portion of the workload that works well with these fast-decode, limited-capacity chips.</p><p><strong>Okay, that&#8217;s a great description because what the framing that comes out of all this is that you have to use this chip where its strengths lie and its strength lies in providing an enormous amount of memory bandwidth, 150 terabytes per second, which is like 10 times higher than what GPU HBM-based systems give you. And I believe there is also a benefit that comes from not having to move bits as much and that saves a lot of energy as well, if you don&#8217;t have to shuttle bits far away, right? So we&#8217;ll get to the energy efficiency aspect. But because of this capability of Corsair and SRAM and the limited capacity, you can actually use it in unique places.</strong></p><p><strong>The speculative decoding was one example you gave where, I&#8217;m just trying to explain to myself so I understand it as well, is that you use a smaller model. The drafter model you mentioned is usually a much smaller model than like a Kimi K2 big, trillion parameter. This is not a big model. If my understanding is the speculative decoding drafter model doesn&#8217;t even have to be that accurate. So it&#8217;s okay to use a small model because you&#8217;re going to be spraying out like eight tokens and the GPU will decide which of these eight tokens are the correct one by the process of verification. So the idea is that you have to keep the GPU fed with speculative decoder and the drafter, which is being really fast, which is where the Corsair fits in. So that the GPU can run in parallel verifications. It can run 8 to 16 tokens in parallel verification as long as it gets enough tokens from SRAM to verify, right?</strong></p><p><strong>Sudeep</strong>: Exactly correct. Some of the work we did with Gimlet, we have a 1.6 billion draft feeding a 120 billion target model. Typically it&#8217;s a percent or less. It&#8217;s a much smaller model. For example, in the Kimi K3 work that we&#8217;re doing, we would use a 4 billion drafter for a 3 trillion model. That&#8217;s a very small draft compared to the target model.</p><p>And the draft would be able to get many tokens right, and it&#8217;s always exactly the same as what the GPU would output because it&#8217;s always verified by the GPU. If at all the drafter makes a mistake, the GPU tells it what the mistake is and then the drafter corrects and continues from there. So there&#8217;s no difference in the output tokens that the GPU produces compared to suppose it was not running any speculative decoding. Using these rejection sampling approaches, we make sure it&#8217;s exactly the same accuracy. We don&#8217;t lose any accuracy.</p><p><strong>And what is the bandwidth benefit you get by using this drafter-verifier approach rather than just running decode on a fast GPU?</strong></p><p><strong>Sudeep</strong>: The fast GPU would be generating one token at a time. And if you were to get let&#8217;s say five tokens correct on the average&#8212;let&#8217;s say the drafter produces eight tokens or 16 tokens and feeds that in a chunk to the GPU and on the average, let&#8217;s say we get five out of eight correct&#8212;you would get much better speed-ups. So we were able to demonstrate three and a half to four X speed-ups on the GPT-J model.</p><p>There are two ways you could use it. One is you could get to regimes you could not get with a GPU. In the work we did with Gimlet, for example, we could get to 1600 tokens per second per user. Something that the GPUs could not get to. The GPUs could only get to say 400 tokens per second per user. So we get this 4X higher interactivity. Another use case is for the same interactivity, you can get much higher throughput. So both modes are possible. And it&#8217;s up to the user how they want to trade off throughput versus interactivity. We essentially shift the whole Pareto curve to the right.</p><p><strong>You did mention two other forms of disaggregation, which is like attention feed-forward disaggregation and then just disaggregating the prefill and decode. The broad idea is that even the process of inference has many ways to break up the problem and you could use SRAM-based inferencing for parts of them where it really benefits the most. All of those things, just like the speculative decoding example you gave, should speed up the inference and move the Pareto frontier of the interactivity curve in a good way. So these are all ways you can use the SRAM approach. This is a very good setting the stage for what comes next. But before we do that, we have to address what the LPDDR on the card is meant to do.</strong></p><p><strong>Sudeep</strong>: In the performance versus TCO curves, you can have super-fast these SRAM chips like d-Matrix or Cerebras or Groq. They&#8217;re super fast, but you sometimes require a very large footprint. That&#8217;s at the top of the tier. Then there&#8217;s the mid-tier, which is essentially all of the GPUs, and that&#8217;s where the bulk of the market is. And then there is another tier where you can store large models very efficiently on LPDDR. But you may not get the interactivity advantages of the GPU or of course the SRAM. But you could use that to reduce cost. You could use that to shrink the number of cards because you can pack a lot of memory on one card.</p><p>For example, the card we&#8217;re showing, you can put 250 gigabytes and this is commodity LPDDR. The same LPDDR that is found in cell phones and it doesn&#8217;t require three times as many wafers to produce the same bits as HBM. So it&#8217;s much cheaper because of that. So we cover both one and three. Obviously we don&#8217;t use any HBM on Corsair, we don&#8217;t touch two, but we cover the other two.</p><h2>Raptor and 3D Stacked DRAM</h2><p><strong>So you have one approach to go really fast and another approach to go really slow. The really slow approach gives you more capacity, but not as much speed and that&#8217;s useful for some workloads. Maybe like offline batching, you could process all your day&#8217;s work at night and then next morning you can come back to inference results. You can run those cheaply on a widely available DRAM. Although I don&#8217;t even know what that means anymore. Is DRAM even widely available? There&#8217;s no such thing called cheap DRAM or widely available DRAM anymore. But anyway, it&#8217;s DRAM. It&#8217;s not HBM, which is even harder to get a hold of possibly. So it serves two different areas of high speed or high capacity. Because you don&#8217;t have HBM, you don&#8217;t have the medium-speed, medium-capacity tier, but you have the other two extremes. This is good. I think the next step we should go into is actually the meat and potatoes of all of the d-Matrix stuff that&#8217;s been going on at Hot Chips, which brought all the excitement and was all over Twitter after your talk, which is basically 3D DRAM. And I think it solves this high-speed tier somehow by increasing capacity as well. But before all that, we should explain what 3D DRAM means in this context because there&#8217;s a lot of confusion as to whether this is actually 3D DRAM from like the transistor level. So we should set the stage of what actually 3D means in this situation.</strong></p><p><strong>Sudeep</strong>: Yes. This is our Pavehawk proof-of-concept silicon of 3D stacked DRAM. What we mean by 3D DRAM is an inference accelerator where one layer is logic, which is where the compute happens, and another layer is the memory, DRAM, and those are vertically stacked. So unlike HBM which sits 2D, where you have DRAM that sits next to a GPU, here we vertically stack compute and memory.</p><p><strong>This is helpful because now we can discuss basically why we had to stack DRAM on top of logic. And I have maybe several questions around this. The first one is, we will directly be able to increase capacity because DRAM is about an order of magnitude better bit density than SRAM is. So immediately you will go from the hundreds of megabytes per chip. In your case, you have multiple chips that give you 2GB of SRAM in the Corsair, right? But now, what kind of capacity do you get by just putting one chip of DRAM on top of your logic chip?</strong></p><p><strong>Sudeep</strong>: We get 32GB of DRAM. In the specific construction, we have DRAM at the bottom and logic at the top. And this is what we mean by 3D DRAM. You can see the cross-section on the right where they&#8217;re connected using micro-bumps. And the power to the logic die is delivered through the TSVs. The logic die itself in this case is a TSMC N4 logic die. And the bottom die is a custom-designed DRAM. And we&#8217;re using 36-micron face-to-face stacking. So going back to your question, Vic, we can pack 32GB of capacity at 100 terabytes per second of memory bandwidth.</p><p>Essentially, we got 16 times the capacity of SRAM, while we retained SRAM bandwidth. And in this particular demonstration, we measured 0.37 picojoule per bit because the bits only go down and traverse very short distances as opposed to HBM where it has to go down the entire stack, and then it has to go sideways, and then it has to traverse tens of millimeters on the logic die. All of that power adds up and you will end up burning about two to three picojoules, 2.5 picojoules to be exact on the HBM itself, HBM4 itself. But another two and a half to three picojoules on the GPU. So that five picojoules became 0.37 picojoules. And that&#8217;s the big advantage of this technology. For the same power, you can deliver a lot more bandwidth&#8212;in this case, 6X bandwidth&#8212;by using this technology.</p><p><strong>So to put that in perspective, if you took your previous Corsair card, which had 2GB of SRAM, you need 16 PCIe cards, which you probably have to put in across two different servers. You can put only eight PCIe cards in a single server of Corsair. And then you&#8217;re going to have two full servers to get 32GB of SRAM. Now, Corsair required all that, while Raptor, just by putting logic on top of DRAM, you ended up getting basically two servers&#8217; worth of capacity while retaining the bandwidth that you had in SRAM. So you had free lunch and you ate it too. Yes. This is very rare in engineering. That&#8217;s great. And the energy efficiency too. The distance that the electrical signals have to cross between the compute chip and the DRAM chip are so close by that you don&#8217;t spend as much energy shuttling bits between memory and logic as you would in HBM. Now, you have a micro-bump that is like, what was the size of that thing? 36 micron pitch. It&#8217;s a pitch, but the height of the bump itself is like tens of microns, right? So instead of going millimeters, you&#8217;re going tens of microns, so the energy required to shuttle bits is now 1/10th of what you would have to do with HBM. Or 1/20th even based on your five picojoules per bit number.</strong></p><p><strong>Sudeep</strong>: Exactly right. Because vertically, of course, you have to make sure that you&#8217;re not going to move a lot of data on the logic die. So we custom-designed and pitch-matched the compute engines directly above the DRAM banks, so we call it micro-banks, and we design a channel such that three micro-banks fit underneath the compute engine, and they&#8217;re pitch-matched. You have to co-design the compute and the DRAM.</p><h2>Stacking Order and Future Bonding</h2><p><strong>Is there a specific reason you chose to put compute above the memory? I think there are two ways to do this. There is two options. One, you could have the compute chip on the bottom and put the DRAM chip on top, which seems to be kind of the approach that Cerebras wants to do at wafer scale or Qualcomm wants to do with their high-bandwidth compute where you put DRAM on top of logic. The other approach, which is what d-Matrix has, at least on the Raptor version of this, is logic that is sitting on top of DRAM. So, there are definitely some trade-offs here that we need to talk about. And why you chose this approach.</strong></p><p><strong>Sudeep</strong>: Yes. You can pick your poison. On the left picture here, you see the logic on the bottom, the DRAM on the top. And of course, since the liquid cooling cold plates are on the top, the heat has to conduct through the DRAMs. And the logic is the hottest die here and it has to traverse the DRAM. So the DRAM temperature will rise and thermally this is not as good a solution as one on the right. On the one on the right, the heat sink is right on top of the logic die because the logic is on the top. It doesn&#8217;t have to conduct through the DRAM. So the DRAMs don&#8217;t see as much of a temperature gradient as on the left. So from a thermal standpoint, the solution to the right is much better than the one on the left. But from a power delivery standpoint, the solution on the left is better than on the one on the right. So those are the fundamental trade-offs.</p><p><strong>Yeah, so putting logic on top of DRAM means that you can just put the cold plate right on top of the logic chip and extract heat before it upsets the DRAM die too much. DRAM really doesn&#8217;t like heat, does it? And the problem with that approach, the poison you have chosen here is you have to deliver power to the logic chip through the DRAM die, which is a power delivery challenge because you want to make sure that there is no power drop across these vias that are going through the DRAM die to the logic die. So, clearly you decided that thermal was a bigger problem to fix early on.</strong></p><p><strong>Sudeep</strong>: That&#8217;s correct. For a one-high stack, we found that we could in all our modeling, we could reliably cool the DRAM to be less than 95&#176;C for power densities of the order of 0.5 to 0.7 watts per millimeter square. And that&#8217;s the power density of our logic die. So we could solve both the power delivery problem and the thermal problem using this approach.</p><p><strong>In the future, I know right now you have micro-bumps connecting these things together, but in the future, do you think that going to maybe hybrid bonding will help in the bandwidth aspect because the pitch can get closer than 36 microns? But do you think that helps the thermal aspect if you go to hybrid bonding in the future?</strong></p><p><strong>Sudeep</strong>: Yes, it does help the thermal aspects, especially if you do multi-high stacking. But the fundamental problem is still reducing the energy consumption on the DRAM. Because you still have some routing on the DRAM and it still takes like in our case, 0.37 picojoule per bit. So you can pack with wafer-on-wafer, you can get a lot more I/O than say 100 terabytes per second, right? It could be 400 terabytes per second. But if you don&#8217;t reduce that I/O power from say 0.37 picojoule to 0.1, the total power that you&#8217;ll dissipate just in the interconnect is so high that that is no longer the limiting factor.</p><p><strong>Yeah, that&#8217;s a very good point and to put some more light on that, the way you can easily calculate how much power this kind of a chip at a given memory bandwidth will consume is if you take the bandwidth, let us say 100 terabytes per second, and you multiply it by the number of picojoules per bit, it automatically spits out the power consumption in watts. So if you go four times higher in the memory bandwidth by doing some techniques, you want to be sure that you reduce the energy per bit four times to keep the same level of power dissipation at the chip level. Otherwise, what ends up happening is if you boost the bandwidth, but you&#8217;re still spending energy shuttling bits, you&#8217;re just shuttling more bits and that&#8217;s going to cost you in terms of power.</strong></p><p><strong>Sudeep</strong>: Yes, so 100 terabytes, so you convert it to terabits. So 100 * 8 * 0.37, that I think is about 280 watts. Just in the memory subsystem to move those bits into the compute unit. And now if you start putting stacking more, maybe that number will go up. If the number, let&#8217;s say, even remains the same, but if you want four times the bandwidth, then that&#8217;ll end up burning upwards of 1,000 watts. So we have to first figure out from an accelerator standpoint, what&#8217;s the bandwidth we desire, how much power we want to spend on the memory subsystem. Because there&#8217;s still compute you have to spend power on, interconnect you have to spend power on. Typical rule of thumb we use is 60% on compute, 30% on memory, 10% on interconnect. And then figure out what technologies can help you get there.</p><h2>Building Racks with Nvidia</h2><p><strong>That&#8217;s fascinating. What comes next after this? The really exciting part of the news that recently came out from d-Matrix is the use of the Raptor chips in the Nvidia NVLink Fusion ecosystem. What&#8217;s the idea and what are you thinking going into putting Raptor chips into the ecosystem? What is the benefit of that? And really how easy is it to integrate Raptor chips into the Nvidia ecosystem?</strong></p><p><strong>Sudeep</strong>: The rack is our product. At the end of the day, like you&#8217;d pointed out, models are very large and let&#8217;s take an example like Kimi K3. Kimi K3 is the largest open-source model out there, 2.8 tera. It&#8217;s a huge model. And it&#8217;s got 1 million context. Obviously, it won&#8217;t fit in one chip, GPU or otherwise. So the whole idea is to put many of these chips, in our case, 144 of these R4 packages. The Raptor package looks exactly like this. There are four chiplets in a package. And there are 144 of those packages inside a single rack. And for that, we chose the Vera Rubin rack. And we use NV switches. So we have 18 compute trays and nine switch trays. And each compute tray has eight of these packages. So that&#8217;s how we get 18 * 8, 144. So we have 144 Raptors inside a VR rack, Vera Rubin rack.</p><p>What we do is we keep the Vera CPU the same, all of the networking, the CX and the Bluefield networking the same, and replace the Rubin GPU with Raptors. And instead of four Rubins, we have eight Raptor packages. So a sum total of 144 chips that&#8217;s all-to-all connected. And this rack then can talk to a Rubin rack because they use the same networking technology.</p><p><strong>So you just have to pull out all the Rubin GPUs, there&#8217;ll be like four of them, put in like eight Raptor chips. You keep the Vera CPU and you build a rack with 18 computer trays. Each one has eight Raptor chips, that&#8217;s 144 chips. Each one, what is the total capacity of the 3D DRAM in a single rack?</strong></p><p><strong>Sudeep</strong>: 2.3 terabytes of capacity.</p><p><strong>So that&#8217;s pretty big to hold an entirely large model all by itself actually, if you wanted to do that.</strong></p><p><strong>Sudeep</strong>: Correct. We were showing at Hot Chips that we could hold the Kimi K3 model at 4-bit precision and have sufficient space for many users&#8217; KB cache space and we could hit 1,000 tokens per second per user. It&#8217;s a performance metric that you can&#8217;t achieve with traditional memories. The 3D DRAM, because it&#8217;s such high bandwidth, you can get these levels of interactivity at high batch all in one rack. In a single rack, we can get to a 3T model at 1 million context at 1,000 tokens per second per user on Kimi K3. And on GLM 5.2, we could get to 3K. So these are really fast decode speeds.</p><p>And because in the Nvidia racks, you can do multiple racks, there is an NVL 576 option where you have four racks and the four racks are connected using optics, optical scale-up. And using that now, you have not just 2.3 terabytes of capacity, but 2.3 terabytes times four. So 9.2 terabytes of capacity, which is sufficient to, at FP4 precision, serve even a 10T model. So we can serve out the whole frontier at unprecedented speeds because of our rack-level, rack-scale solution as well as the fundamental technology that delivers much higher bandwidth.</p><p><strong>So you&#8217;re talking about basically putting four racks of Raptor chips, each with 144. So that would make it 576. Hook them all with optics or any form of very high-speed interconnect, and then now you have a large scale-up domain with basically nearly 10 terabytes of SRAM-like bandwidth, but with 3D DRAM capacity, which can serve any large frontier model of the future at ridiculous speeds.</strong></p><p><strong>Sudeep</strong>: Exactly.</p><h2>Software and The Future</h2><p><strong>Seems like a winner. How does the whole software stack work though? So if you put in a rack of Raptor chips next to an Nvidia rack running GPUs, all of that stuff runs on CUDA. How does the software work? How do you do the disaggregation?</strong></p><p><strong>Sudeep</strong>: With Dynamo and frameworks like that would handle the disaggregation. sglang or vLLM, any of these frameworks would partition the work even today between say a pre-fill cluster and a decode cluster. Same thing happens here. But underneath that, when it comes to generating code for our chips, we have our own software stack. We provide something called a Model Builder SDK. And the Model Builder SDK essentially we expose a set of primitive kernels&#8212;MatMul, SIMD, various shapes and so on&#8212;and a way to stitch together models using those primitive kernels. And a programmer can choose to write new kernels if they want to, because we expose the ISA of the chip, or just to take these building blocks and compose models. And we provide a lot of example models. And the AI kernel generation tools have gotten so good these days that using these examples and using these agentic harnesses, you can quickly come up with a new model and perform a new model in a couple of days.</p><p>So that&#8217;s the lowering stack. And then we have our own execution stack where the device runtime, there&#8217;s a runtime that runs on our chip. And that runtime runs on the embedded RISC-V processors that are part of each of those chiplets. It&#8217;s the same on Corsair, it&#8217;s the same on Raptor. And everything essentially is targeted on that RISC-V.</p><p><strong>I see. It&#8217;s fascinating. What comes next after this? Clearly if you can have one DRAM stack, what are you going to try next? Put two, three, four? What&#8217;s the roadmap on that?</strong></p><p><strong>Sudeep</strong>: We are trying to build up to a four-high stack. On Lightning, which is our next product, we have our logic on the bottom and then four DRAM on the top, and we&#8217;re trying to build up to that. Now, we have obviously lots of challenges in terms of thermals, in terms of the bandwidth we can deliver, in terms of the availability of hybrid bonding. As you know, hybrid bonding of DRAM and logic hasn&#8217;t yet been successful, mainly because of contamination or warpage. These are fab-grade processes that you cannot take a wafer into a clean room and contaminate, for example, a logic fab. So all of these are challenges that need to get worked through.</p><p>But more importantly, from an architectural standpoint, if you cannot test any of these things and you&#8217;re just going to hook this up, how are we going to solve this problem? How are we solving the yield problem? How are we solving redundancy problems? If you look at even Raptor, we do refreshes that are eight times faster, 4-millisecond refreshes compared to 32-millisecond refreshes, precisely because of these thermal concerns. And now, what is the solution for a multi-high stack? So stay tuned. We have solutions to many of these problems, but it&#8217;s definitely a work in progress.</p><p><strong>I think to me that the inference world right now is kind of the wild west. Like making training chips seems to be like there&#8217;s one or two methods to do it. Most of them are GPU-based. But inference is all over the place. You&#8217;ve got wafer-scale engines, you&#8217;ve got SRAM-only accelerators, you have SRAM plus LPDDR, and then now you have 3D DRAM, and then you&#8217;ve got other companies who use all three tiers. Everybody has their own approach. So in the future, say a year from now, do you see any kind of a convergence into what is the right architecture we should be doing? What is your take on that?</strong></p><p><strong>Sudeep</strong>: It is what we&#8217;ve been discussing. To us, it&#8217;s fairly obvious that the densest memory today is DRAM. And because it&#8217;s 1T1C, you can get 40, 50 megabytes per millimeter square, versus four or five on SRAM. So you get much denser storage and you can get the bandwidth by stacking. So with this heterogeneous integration approach where you stack, the benefits are so obvious and specifically the metrics we called out, like on bandwidth and power, is so compelling that it&#8217;s only a question of when, not if. So we feel like in five years, this is the wave of the future. Everyone will be doing 3D stacked DRAM. Now, there are engineering challenges to solve along the way and we&#8217;ll all figure out. There&#8217;s a lot of problems to solve and we&#8217;ll figure out how to solve them, but we believe this is the end state of inference.</p><p><strong>That&#8217;s amazing. I&#8217;m only going to leave you with one contest that I think somebody in the future is going to come over and say, hey, why don&#8217;t we stack high-bandwidth flash on top of logic? That&#8217;ll give us even more capacity, right? And then you open up a whole new world of problems and then we can all work on it for another five years. This is great.</strong></p><p><strong>Sudeep</strong>: Yes, exactly. Talking about wild west, it&#8217;s a really exciting space. There are so many new ideas, there are so many avenues to explore. I&#8217;m sure creative people will find new ways to solve problems. It&#8217;s just a really, really exciting time to be in this space.</p><p><strong>Yes, absolutely. Thank you so much for doing this podcast with us. As always, whether it&#8217;s your talks or meeting you in person or on the podcast, I always learn something new. So I hope our listeners learned a whole lot of stuff about d-Matrix and how everything has evolved from your initial analog in-memory compute days, far before even GPT was a thing. You&#8217;ve started building something for inference, which is clearly well over half a decade out, and you have working silicon that is now going into racks, and it&#8217;s linking up to Nvidia&#8217;s ecosystem and in the future, you&#8217;re going to have even more memory per chip and so you can build even more models and bigger things and smarter things. So that&#8217;s all very exciting. So thank you very much.</strong></p><p><strong>Sudeep</strong>: Thank you, Vic. I&#8217;ve always been a fan. I enjoy your writings and I enjoy your podcasts. It&#8217;s great to be here and talk about what we love to do. So, thank you for having us.</p><p><strong>All right, that&#8217;s it for this episode of the Semi Doped podcast. If you enjoyed it, leave us a five-star review on Apple Podcasts. It&#8217;s very important. Rankings are always important. We always want to rank higher than somebody else. So, get that and you can always watch this on YouTube or any podcast platform of your choice. And if you enjoy it, share it with a friend. And we also have a daily newsletter on daily.semidoped.com where we keep track of all of the daily news. So, check that out as well. Catch you on the next one.</strong></p>]]></content:encoded></item></channel></rss>