Samsung Electronics posted a record 107.4 trillion won in Q3 2026 operating profit, a 783% year-on-year surge that Reuters called the highest quarterly profit ever reported by any tech company at roughly $80 billion. Elsewhere, Wolfspeed landed a conditional $1.5 billion U.S. government loan to build out a domestic SiC supply chain, and AMD CEO Lisa Su was in Seoul on October 7 pressing Samsung and SK Hynix for supply deals spanning three to five years.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Samsung and TSMC Print All-Time Records as AI Spending Lands in Earnings

Samsung Electronics reported a 783% year-on-year surge in Q3 2026 operating profit to a record 107.4 trillion won, with Reuters noting the result represents the highest quarterly profit ever posted by any technology company at roughly $80 billion. The memory windfall drove almost all of the gain, as hyperscaler AI infrastructure buildouts translated directly into HBM and DRAM demand. TSMC’s quarterly revenue jumped 51% year-on-year, with Bloomberg reporting that AI demand held firm through the period. Both companies reported these results within days of each other, yet their stocks barely moved, reflecting, per Bloomberg, investor expectations already calibrated to torrid AI-driven growth.

Austin: More operating profit than Nvidia!? Nvidia did $63.7B in its July quarter. And that’s with Samsung’s phones, TVs, and foundry reportedly losing money, so memory is carrying all of it.

Vik: $80B a quarter is a stupidly large number. This memory craze has to end at some point, right? This is literally the trillion dollar question.

Zeiss and ASML Map Hyper-NA EUV Path as Sell-Side Upgrades 2027 Outlook

Carl Zeiss and ASML have publicly outlined a Hyper-NA EUV roadmap extending beyond the High-NA tools currently entering production, confirming that chipmakers are already planning for the generation after next. The technical work centers on pushing numerical aperture past the 0.55 ceiling of High-NA systems, a step that requires new optical architectures Zeiss is developing alongside ASML. Wall Street is paying attention. JP Morgan says ASML could signal a stronger-than-expected 2027 revenue outlook at its next investor event, while Morgan Stanley has separately raised its earnings estimates on the basis of ASML’s pricing power and accelerating EUV demand. Customers apparently aren’t waiting to find out.

Vik: Everyone’s planning the node after next, but I called hyper-NA just pointlessly expensive, EUV pushed past its economic limits. The 2027 money is booked anyway - JP Morgan says most EUV orders for the year are already committed. Hyper-NA revenue is a long way off.

Wolfspeed Wins $1.5B Defense Loan to Build Domestic SiC-GaN Supply Chain

The U.S. Department of War’s Office of Strategic Capital has issued a conditional $1.5 billion loan commitment to Wolfspeed (NYSE: WOLF) to finance domestic production of silicon carbide materials and wide-bandgap power devices, with an explicit national-security mandate covering applications in drones, radar, and missile defense. The commitment advances key Department of War initiatives alongside the domestic wide bandgap supply chain. The commitment is conditional, meaning Wolfspeed must satisfy further requirements before funds are disbursed. Demand from AI data centers is running alongside the defense rationale: higher-voltage power networks in large-scale compute facilities have accelerated adoption of SiC and GaN devices, giving Wolfspeed’s expansion a dual commercial and military customer base.

Vik: Wolfspeed overbuilt SiC for EVs and got caught when Chinese capacity took demand down, so a conditional 30 year defense loan matters a lot here. What interests me is the GaN-on-SiC RF epi for electronic warfare, not the data center angle everyone’s tacking on. SiC demand is legit though.

AMD CEO Lisa Su Presses Samsung, SK Hynix for Multi-Year Supply Deals

Lisa Su traveled to Seoul on October 7 to open partnership talks with Samsung Electronics and SK Hynix, telling Korean media at Seoul Dragon City in Yongsan that supply chain complexity requires cooperation plans spanning three to five years. The meetings covered HBM memory and advanced packaging allocations that AMD needs ahead of its next product ramps. Su also met Ryu Je-myung, South Korea’s second vice minister of the Ministry of Science, and announced a dedicated AI research center to deepen AMD’s presence in the country. AMD separately said it plans to integrate Korean NPUs into its own hardware, broadening the scope of the Seoul talks well beyond memory supply. The research center commitment and NPU integration pledge together give Korean counterparts reasons to prioritize AMD alongside existing Nvidia obligations.

Vik: This is AMD trying not to end up last in line for HBM behind Nvidia. Memory is over half a rack’s cost now, so a 3 to 5 year lock makes sense. The AI research center is the sweetener you dangle to get bumped up the allocation queue.

Austin: First time AMD has disclosed SK Hynix memory in its AI systems. HBM4 for MI455 was known to be Samsung only, so a second source is the real news here with supply this tight.

Turba Labs Raises $52 Million to Double Compute by Optimizing Existing Data Centers

Palo Alto-based Turba Labs has raised nearly $52 million in combined seed and Series A financing to optimize AI infrastructure without building new data centers. The two-year-old startup creates a digital twin of data center infrastructure, models customers’ AI use cases, and routes workloads to hardware best suited for their complexity. Co-founder and co-CEO Hans-Juergen Schmidtke said the platform could double available compute by customizing how AI workloads interact with underlying hardware, though the exact capacity unlocked will vary by infrastructure.

Sources: wsj.com

Vik: Huh. Interesting idea. Run and optimize a datacenter buildout using computer simulation before deploying billions. There is a lot of economic value to that.

Austin: Double is a big claim. What’s the baseline, i.e. how much compute is sitting idle in these data centers today?

Applied Digital Posts Q1 2027 Results as HPC Lease Demand Accelerates

Applied Digital reported fiscal first-quarter 2027 earnings as demand for high-performance computing infrastructure continued building across its data-center portfolio, with utilization rates and lease pricing drawing close attention from analysts tracking the broader HPC operator cycle. The company, regarded as one of the first AI-native data-center operators to report this cycle, provides an early read on how AI cloud infrastructure is absorbing capacity. Pentagon interest in AI cloud services has added a new demand vector for operators like Applied Digital, expanding the customer base beyond traditional hyperscalers. As subsequent HPC operators prepare to report, Applied Digital’s figures on contract pricing and facility utilization now serve as the opening data point for the quarter.

Vik: Applied Digital’s whole pitch is durable long term contracts with tier-one investment grade hyperscalers. Do not underestimate military interest in AI compute, especially after the OAI-HuggingFace incident has shown what AI agents can do in unison. Rest assured governments will militarize that.

Sector Watch

Foundry & EDA

• Siemens EDA and TSMC are jointly applying agentic AI to semiconductor design workflows, targeting faster RTL-to-GDSII iteration on TSMC process nodes. (The Elec)

• Rapidus enlists four Malaysian IC design houses as it actively builds a 2nm customer pipeline ahead of commercial production. (The Business Times)

• Rohm expands back-end chip manufacturing outsourcing, signaling a structural shift in how the Japanese IDM manages packaging capacity.

Memory

• SK Hynix’s NAND flash memory unit is said to have selected banks for a 2027 U.S. IPO, putting a valuation process in motion for one of the world’s largest flash suppliers. (Bloomberg.com)

• Biren shares slump 6 percent following a $4 billion share sale. (The Standard (HK))

• BE Semiconductor shares slide as analysts flag memory-shortage pressure on its hybrid-bonding ramp, per Bloomberg coverage. (Bloomberg Tech)

Packaging & Semiconductor Equipment

• Aehr Test Systems receives $6M follow-on Sonoma burn-in production orders from its lead hyperscale customer for a next-generation AI processor, extending its burn-in footprint. (Aehr Test Systems)

• Mitutoyo opens a dedicated semiconductor metrology center in Dresden, expanding precision measurement support for European chip fabs. (The Elec)

• Boss Semiconductor and Japan’s Mirise Technologies partner to develop chiplet interconnect technology for automotive applications. (The Elec)

Power

• Renesas adds 100V GaN FETs targeting AI data-center and industrial power conversion, expanding its wide-bandgap device lineup.

• LS Electric establishes a Taiwan joint venture to expand its power-infrastructure business on the island. (The Elec)

• HD Hyundai Infracore wins a 390 billion won US data-center gas generator engine order, its largest disclosed stateside energy contract. (The Elec)

Data Centers & Cloud Infrastructure

• CoreWeave enters India with a 240 MW AI cloud region at AdaniConneX campus, its first international deployment outside North America and Europe. (CoreWeave)

• San Francisco passes a data-center moratorium, restricting new builds in the city that is home to many of the leading AI companies. (Bloomberg.com)

• Firmus and CDC abandon their planned $51 billion Australian AI data-center project as Firmus shifts focus to Asian expansion. (Light Reading)

Compute & Edge

• Microsoft launches the $2,599 Surface Laptop Ultra with an Nvidia RTX Spark chip, its first co-engineered AI PC with Nvidia. (CNBC)

• AMD will launch a next-generation APU with FSR 4 support by year-end, expanding its consumer AI compute lineup. (The Elec)

• Dell Technologies unveils new Windows PCs and workstations targeting the agentic AI era at a product launch event. (Business Wire)

Policy & Trade

• Amazon vows to stop signing data-center NDAs with local governments, following Microsoft’s similar pledge amid scrutiny of secret deals. (Law360)

• FCC approves SpaceX to deploy a 15,000-satellite D2D LEO constellation, with a nine-year completion deadline through October 2035. (Light Reading)

• SEMI Europe joins the European Commission’s STEAM Executive Panel to address the semiconductor industry’s growing skills shortage. (EE News Europe)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Elon Musk ruled out any operating role for TSMC at Terafab and Intel stays in the venture, with Musk’s companies building and running the fab on licensed TSMC process technology. Elsewhere, SpaceX is in talks to borrow $40 billion through Apollo to buy Nvidia chips, Applied Materials and Intel formalized an R&D pact on next-generation transistors and packaging, and Samsung Electro-Mechanics landed a $216 million AI server MLCC order.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Musk Confirms Intel Stays in Terafab as TSMC Operating Role Ends

Elon Musk has ruled out TSMC taking any operational role at Terafab, the Texas AI chip complex he is building, saying his own companies will construct and run the facility while TSMC contributes process technology under license. The clarification came after Intel shares initially slid on fears Musk might hand TSMC a broader position, then rebounded once the structure became clearer. Bloomberg confirmed that Intel will keep working on the Terafab venture, preserving its place in a project that had briefly looked like it might sideline the company entirely. Terafab thus takes shape as a Musk-owned fab running TSMC process nodes with Intel in a supporting capacity, according to Bloomberg.

Austin: Note to readers, this whole convo started recently with a scoop from Tim Culpan:

Culpium, by Tim Culpan

Exclusive: TSMC Exploring Plans to Work With Elon Musk’s Terafab

Read more

6 days ago · 32 likes · 1 comment · Tim Culpan

Vik: TSMC licensing its process for someone else to run? That seems like news in itself.

SpaceX Pursues $40B Apollo Financing to Stockpile Nvidia Chips

SpaceX is in talks to borrow $40 billion, led by Apollo, specifically to purchase Nvidia chips for its AI buildout. The scale is hard to ignore: forty billion dollars in debt, raised not for rockets or satellites but for GPUs. Doosan Enerbility has separately completed production of a 380-megawatt gas turbine destined directly for SpaceXAI, holding a shipment ceremony at its Changwon headquarters with roughly 70 attendees. SpaceX is sourcing chips through private credit and commissioning bespoke power infrastructure without routing either through a hyperscaler. The financing round, if closed, would rank among the largest single AI hardware purchases on record.

Vik: They’re building the whole stack themselves. Chips through private credit, and a 380MW Doosan turbine shipping straight to SpaceXAI, no hyperscaler in between. Compute’s still short, so owning your own supply makes sense right now.

Austin: At Jensen’s $40B per GW for Vera Rubin, this loan buys roughly 1 GW of compute. One 380MW turbine is a start, but they’ll need a lot more power than that.

Penguin Solutions lifts fiscal 2027 revenue target to $2.43 billion

Penguin Solutions raised its fiscal 2027 revenue outlook to $2.43 billion, up from a prior target of $2.2 billion, after reporting a Q4 earnings and revenue beat driven by demand for AI systems integration and memory subsystems. The company now guides for fiscal Q1 2026 revenue between $480 million and $520 million, with full-year adjusted operating margins expected to expand alongside the revenue ramp. The stock climbed on the news. Analyst reactions split: Stifel raised its price target on the strength of the revised guidance, while Barclays cut its rating, citing valuation concerns as the revenue trajectory steepens. The divergence puts investors in the familiar spot of deciding whether the growth is worth the multiple.

Vik: Memory is doing the heavy lifting here. Integrated Memory was over nine-tenths of the year over year gain, so this is a memory subsystem story dressed up as AI systems.

Samsung Electro-Mechanics Cements AI Server MLCC Lead With Fresh $216M Order

Samsung Electro-Mechanics has landed a $216 million MLCC contract for AI server supply, adding to a position that already covers more than 40% of the AI server MLCC market, according to Seoul Economic Daily. The company now holds over 40% share in AI server MLCCs, a component that has become a chokepoint in next-generation server builds as GPU-dense configurations demand far higher capacitor counts per board than conventional rack servers. Digitimes reports that the new deal locks in 2027 delivery commitments as supply pressure in the segment continues to build, with lead times tightening across the MLCC tier. The $216 million figure represents a discrete order on top of existing volume, not a cumulative tally, reinforcing how concentrated procurement has become around a single supplier.

Vik: 600K MLCCs in a single rack. That’s where the demand is coming from.

Applied Materials and Intel formalize next-generation transistor and packaging R&D pact

Applied Materials and Intel have announced a formal R&D collaboration targeting next-generation transistors, interconnects, and advanced packaging technologies, drawing on Applied’s EPIC Center in Silicon Valley and Intel’s R&D campus in Hillsboro, Oregon. The goal is to shorten time-to-market for semiconductor processes central to AI computing. The two companies will combine capabilities across both facilities to accelerate chipmaking innovations that span transistor scaling and advanced packaging in a single coordinated program. Applied and Intel describe the relationship as built on decades of prior work together, now structured to move faster at the leading edge. Besi and Applied have separately deepened their own packaging collaboration around the same period, extending Applied’s co-development reach across the packaging supply chain.

Vik: The Applied-Intel headline is Foveros and transistors, but the Besi extension is the part I’d watch - die-on-panel and photonics-enabled interconnect.

Austin: Intel joins Samsung, Micron, SK Hynix, TSMC, and Kioxia at EPIC. All three leading-edge foundries doing R&D at Applied’s site now. Good spot for Applied to be in.

AI-Powered Hack Hits Seven South Korean Banks, Exposes 68,000 Records

Hackers used a Chinese AI cybersecurity tool called Artex AI to breach at least seven South Korean financial firms and steal personal information belonging to 68,000 people, South Korean officials said. The attacks were initially detected last week, and the National Police Agency’s cyber terror unit opened a formal probe on Tuesday. Investigators said the intrusion is one of the first known AI-agent-assisted attacks on the global financial system.

Sources: wsj.com

Austin: AI-agent assisted you say? Crime even requires GPUs now…

Sector Watch

Memory

• Micron union in Taiwan receives official authorization to strike, introducing potential production disruption risk at a critical DRAM manufacturing site. (Reuters)

• Samsung Electronics Q3 profit seen jumping nine-fold year-on-year though analysts warn chip margins may remain flat amid HBM yield pressure. (Reuters)

• Kingston retains 62% memory-module market share as B2B AI server demand drives record shipments, per DigiTimes analysis. (digitimes)

Foundry & Packaging

• Towa announces a new Japan plant to reshore strategic chip-packaging equipment production as AI packaging demand accelerates. (digitimes)

• NEXTIN wins a 25.44 billion won wafer inspection equipment order from SK Hynix, its largest single order disclosed this cycle. (The Elec)

Compute & Edge

• AMD will report fiscal Q3 2026 financial results on an unspecified date, with no additional details provided in the announcement. (AMD)

• Cirrus Logic acquires INVENTVM Semiconductor, expanding its mixed-signal and audio silicon portfolio with an undisclosed-value deal. (Business Wire)

• Silicon Labs launches new IoT developer platform tools at its annual summit, adding edge-AI capabilities for connected hardware builders. (EE Times)

Optics & Networking

• Openreach deploys Ciena’s Waveserver 1.6 Tbps platform across the UK national fibre network, the operator’s largest coherent upgrade to date. (Comms Business)

• CIG and ml&s Group will establish high-speed optical-module production in Arteaga, Mexico, targeting AI data-center interconnect demand. (Business Wire)

• AT&T forms a new 50%-owned fiber JV combining Gigapower and Forged Fiber 37 with Global Infrastructure Partners and CPP Investments to accelerate wholesale open-access fiber expansion. (Light Reading)

Power

• Bloom Energy adds a Silicon Valley manufacturing facility as AI-era demand for solid-oxide fuel-cell power systems outpaces existing capacity. (CoStar)

• Infineon and ZuriQ expand their partnership to develop scalable quantum-computing control electronics, deepening the chipmaker’s quantum hardware push. (Bisinfotech)

• LS Power closes a $6 billion flagship infrastructure fund, citing AI-driven power-demand growth as the primary investment thesis. (Bloomberg.com)

Data Centers

• LG Electronics signs a 5 GW cooling-solutions supply agreement for North American AI data centers, its largest cooling commitment to date. (The Elec)

• Lambda targets a $4 billion fundraise ahead of a planned IPO as the Nvidia-backed GPU cloud expands capacity to meet hyperscaler overflow demand. (Reuters)

• Huawei details global expansion plans for its Ascend 950 supernode cluster, outlining deployment roadmap for large-scale AI training infrastructure. (Huawei Central)

Policy & Trade

• Amazon warns publicly that more than 100 proposed US data-center restriction measures risk ceding AI infrastructure leadership to rival nations. (Yahoo Tech)

• Zhone acquires Tellabs in a telecom-equipment deal that consolidates two access-networking vendors amid carrier spending shifts. (Light Reading)

• Microsoft and Nvidia CEOs are set to jointly unveil a new AI laptop at a San Francisco event, marking a high-profile co-launch for Copilot+ PC hardware. (Reuters)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Google locked up 3.6 GW of nuclear power from Constellation, the largest corporate nuclear deal on record and 5x the Amazon agreement from last week. Marvell raised its FY28 guide to roughly $20 billion and now sees $70-90 billion in revenue by FY31, and Tokyo Electron is forecasting operating profit above ¥1 trillion on surging AI tool orders.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Google Locks Up 3.6 GW of Nuclear Power in Constellation Megadeal

Google has signed a 3.6-gigawatt power purchase agreement with Constellation Energy, the largest single corporate clean-energy deal on record, securing baseload nuclear capacity across multiple plants over a long-term horizon. The contract runs well into the next decade and spans several Constellation facilities, giving Google a guaranteed electricity supply as its data-center buildout accelerates. Morgan Stanley, per Reuters, now identifies power availability rather than silicon supply as the binding constraint on AI infrastructure expansion, a view this deal appears sized to address. Nvidia and Broadcom retain relative insulation from the crunch, the bank argues, but hyperscalers face a harder ceiling. The agreement is valued in the billions of dollars, according to Bloomberg.

Vik: The deal between Google and Constellation is in response to PJM’s “bring your ​own power” proposal. This deal is for nuclear power but still relies on the grid. This is a much longer term play, but I think that grid interconnection is still going to be an issue. BTM power with gas turbines is the only near term play (not solid oxide fuel cells btw, lots of raw material availability issues).

Austin: That’s 5x the Amazon deal from last week (690 MW, also Constellation), and Oracle is lining up Point Beach too. How much nuclear is left to sell?

Marvell Sees $70-90B in Revenue by 2031

Lots of positive growth forward for Marvell according to CEO from their analyst day today.

Tokyo Electron Forecasts Operating Profit Above ¥1 Trillion as AI Tool Orders Surge

Tokyo Electron has forecast operating profit exceeding ¥1 trillion for the current fiscal year, driven by surging orders for deposition and etch tools used in advanced AI-chip fabrication nodes. That threshold, rarely crossed by equipment makers, reflects chipmakers’ accelerating capital expenditure on the infrastructure behind large-scale AI workloads. The company’s Korean subsidiary is moving in lockstep: Tokyo Electron Korea has opened triple-digit hiring across 19 roles with a January start date, a recruiting push that points to a service and support backlog extending well into 2027. Korea sits at the center of that demand, home to the leading-edge fabs of Samsung and SK Hynix. Headcount at the local unit expanding at this pace while headquarters clears the trillion-yen profit bar tells you where the orders are actually landing.

Vik: High tool demand like this means more hardware is coming. Note the hiring - field engineers who install and maintain the tools, triple-digit across 19 roles, January start. That’s a service backlog stretching into 2027…. WFE 🚀🌙

Seagate and Toshiba Compete to Acquire TDK’s Magnetic-Head Business

TDK is weighing a sale of its hard-drive magnetic-head unit, with Seagate and Toshiba both in the running as potential acquirers in a deal that Bloomberg reports could reach into the billions of dollars. Magnetic heads are the read-write components at the heart of every hard-disk drive, and TDK is one of the few remaining suppliers. Whichever company lands the unit would absorb a critical piece of HDD supply that currently serves the broader industry, folding an independent component source into a vertically integrated rival. Seagate already dominates nearline HDD capacity sold into hyperscale data centers; Toshiba competes in the same market. TDK has not confirmed a timeline for the process or a final shortlist of bidders.

Vik: TDK is the only independent head supplier, and Toshiba leans on it entirely. If Seagate lands the unit, Toshiba’s sole source ends up sitting inside a rival.

Awesome AI Slop: Jensen as Michael Jackson

Sector Watch

Compute & Edge

• AMD CEO says the company plans to substantially increase chip supply in 2027 and sees very high demand, underscoring continued AI accelerator capacity expansion beyond Nvidia. (Reuters)

• MediaTek launches Dimensity 9600 Pro on a 2nm process in India, its first 2nm flagship chip with agentic AI features targeting 2027 handsets. (Gizbot)

Memory

• Teradyne introduces the Titan HP burn-in test platform for advanced AI data-center devices, expanding its semiconductor test coverage for high-power accelerators. (Teradyne)

• YMTC raises NAND prices and warns internally that the memory shortage will persist until 2029, the most bearish supply timeline disclosed by any major flash maker. (TweakTown)

• CXMT begins mass production of its first DRAM platform while 27 engineers recruited from Samsung and SK Hynix join the Chinese startup, accelerating its ramp. (Network World)

Foundry & Packaging

• Hanmi Semiconductor wins a 24.5 billion won order from Samsung Electro-Mechanics for AI substrate post-processing equipment, signaling rising CoWoS-adjacent tooling demand. (Chosunbiz)

• DISCO reports Q2 parent sales surging 39% to 118.5 billion yen, driven by advanced dicing and grinding demand tied to AI packaging ramp. (Smartkarma)

• SUSS MicroTec enters the wafer-cleaning market with its GreenTec Solutions line and the first GT200 system, expanding its process-equipment footprint at advanced nodes. (PR Newswire)

PCB & Components

• Samsung Electro-Mechanics secures customer backing for a US$5B substrate expansion. (digitimes)

• Magnachip expands its Gen6 super-junction MOSFET lineup with 15 new devices, broadening its power-semiconductor portfolio for industrial and EV applications. (Semiconductor Digest)

• BizLink reports revenue growth driven by its European site restart and integration of Interplex, strengthening its connector supply chain for AI server and automotive markets. (digitimes)

Power

• Infineon completes acquisition of Bengaluru-based C2i Semiconductors, adding AI data-center power management silicon to its portfolio. (Semiconductor Digest)

• Microchip Technology and Navitas Semiconductor develop an 800V DC-to-6V reference design targeting high-voltage AI data-center power architectures. (EE News Europe)

• Infineon opens a new Thailand chip hub with a workforce approaching 1,000 and a Ho Chi Minh City office, deepening Southeast Asia manufacturing and design capacity. (TechRepublic)

Optics & Networking

• MaxLinear introduces its Topaz PON HGU chip family, one of the industry’s most advanced AI-ready platforms supporting 10G and beyond for next-generation broadband. (MaxLinear)

• Nokia is selected by Brazil’s RNP to modernize the national research and education network, deploying optical and IP infrastructure for the AI era. (Nokia)

Data Centers

• Applied Digital secures up to 1 GW of potential power capacity in Finland, marking the company’s first international data-center development outside North America. (Applied Digital)

• Oracle plans a nuclear power offtake agreement with Point Beach Nuclear Station to supply a data-center campus, adding to hyperscaler nuclear procurement momentum. (NucNet)

• DayOne Data Centers files for a US IPO, becoming the latest Singapore-headquartered operator to tap public markets amid sustained hyperscale demand in Asia. (Bloomberg Tech)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Thanks for reading! This post is public so feel free to share it.

Share

Former Groq engineers are suing Nvidia over its $20 billion license-and-hire deal, claiming the structure left them holding equity in an empty shell. Cloudflare shipped Clef, an open-source decision model that is Jev-API compatible just weeks after Typesafe created the category. Elsewhere, Elon Musk confirmed TSMC is in early talks on Terafab, Schneider Electric is buying PTC for $22.6 billion, and Foxconn’s Q3 cumulative sales topped NT$1.15 trillion on AI server demand.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Former Groq Engineers Sue Over Nvidia’s $20 Billion Licensing Deal

A group of former Groq engineers has filed suit against Nvidia over its $20 billion license-and-hire agreement with Groq, claiming the deal was structured to cut them out of any compensation. The plaintiffs allege that Nvidia effectively acquired Groq’s inference chip intellectual property and key personnel while leaving behind a hollowed-out company with no assets to satisfy employee equity claims. The transaction, one of the largest AI chip IP deals on record, saw Nvidia license Groq’s architecture and bring aboard critical engineering staff. Groq’s remaining entity, the engineers argue, amounts to an empty shell incapable of honoring prior financial obligations to them.

Austin: The structure dodged merger review, but can it dodge employee equity claims? Didn’t see that coming but it makes sense. Whatever the courts decide here reads straight across to Microsoft + Inflection and Google + Character.AI… again, those guys set the precedent here.

Cloudflare launches Clef, open-source AI decision model platform

Cloudflare has released Clef, an open-source platform pairing decision models with a reinforcement learning fine-tuning layer aimed at producing bounded structured outputs cheaply and consistently inside automated workflows. The project enters a small but active category seeded by Typesafe AI’s Jev System One, which introduced the decision model concept to a wider audience in recent weeks. Where traditional classifiers return a label, decision models are designed to slot directly into a workflow at the point where a choice must be made, returning structured outputs with enough speed and cost efficiency to run inline rather than as a batch afterthought. Clef’s RL fine-tuning component lets teams adapt those models to their own decision criteria without rebuilding from scratch.

Austin: That was fast. Jev-API compatible and Apache 2.0, weeks after Jev introduced the category. Cloudflare hosting it on Workers AI puts it right next to where the action runs. What’s Typesafe’s moat now? BTW I wrote about the usefulness of these models and their impact on the AI infra buildout here on Chipstrat

Musk Confirms TSMC Talks for Terafab

Elon Musk has publicly confirmed that TSMC is in early discussions to play a role in Terafab, the $119 billion fab project slated to serve Tesla, SpaceX, and xAI exclusively, with Intel named as the only other partner in the venture. No deal is in place, according to tbreak.com, and the talks remain at a preliminary stage. Korean MCE firms have already secured a major stake in the project. Intel stock fell on the news, per qz.com, as investors weighed whether a TSMC-Musk alignment would pull leading-edge US fab investment away from Intel Foundry. TSMC shares gained 1.13 percent on the day. TrendForce lists the Terafab talks as one of five key themes heading into TSMC’s next earnings call.

Schneider Electric’s $22.6B PTC Deal Reshapes Industrial Software Stack

Schneider Electric has agreed to acquire PTC for $22.6 billion, folding the Boston-based maker of Windchill PLM and ThingWorx IIoT software into a company whose core business is power distribution hardware for factories and data centers. The combined entity would span physical infrastructure and the software layer that manages it, a pairing Schneider is positioning directly at AI data center buildouts and smart manufacturing customers. Schneider shares fell on the announcement, an unusual market verdict on a deal the company framed as strategically essential. Siemens and Rockwell Automation both compete across segments the merged company would now cover end-to-end, leaving neither rival with an obvious clean response.

Foxconn Q3 Revenue Hits Record as AI Server Assembly Drives Demand

Foxconn posted third-quarter revenue that beat market forecasts on surging AI server assembly demand, with the company recording its first-ever monthly figure above NT$1.15 trillion. Hon Hai Precision, Foxconn’s listed entity, saw cumulative Q3 sales top NT$1.15 trillion, driven by hyperscaler orders flowing through its AI server integration lines. Bloomberg reported the results as a beat on sales estimates, attributing the outperformance to what it called an AI frenzy among Nvidia’s supply chain partners. The record monthly threshold, crossed for the first time in the company’s history according to Taiwan News, lifted guidance expectations across the broader electronics manufacturing services sector for the quarter ahead.

Austin: Foxconn’s an ODM, so this is a good a read on hyperscaler order flow. Up and to the right.

Scale AI’s Muse opens ESP32 firmware to third-party hardware makers

Alexandr Wang announced Muse Gadgets, releasing an open-source ESP32 firmware and Linux SDK so any developer can build hardware that connects to the Muse AI platform. Scale AI is also shipping its own first-party device under the initiative: Muse Home Link, designed to let Muse interact with consumer smart-home equipment including TVs. The move follows a pattern familiar from voice-assistant ecosystems, where a platform owner publishes a hardware abstraction layer and then seeds the market with a reference product. Releasing on ESP32, the low-cost Espressif microcontroller that already powers millions of IoT devices worldwide, keeps the barrier to entry low for independent makers.

Austin: Dude, agentic IoT? Amazing. This is huge. Qualcomm bought Arduino, they should jump on getting Muse on Arduino!

And this suggests easier agentic physical AI more broadly?

Nathan Lambert launches Trillium Labs to open-source frontier AI research

Nathan Lambert, announcing via X, unveiled Trillium Labs, a new non-profit organization aimed at fostering open science around frontier artificial intelligence. The lab’s initial focus sits on building open post-training recipes, with plans to expand into open infrastructure for studying recursive self-improvement, reward hacking, and multi-agent systems. Lambert, known for his work on reinforcement learning from human feedback, described the effort as built around open principles from the ground up. The organization operates under the handle @trillium_labs. No funding figures or institutional backers were disclosed in the announcement, though the non-profit structure distinguishes it from the commercial AI labs whose closed development practices it implicitly pushes back against.

Austin: A non-profit for open frontier AI, starting with open post-training recipes and later open infra for things like reward hacking and multi-agent systems. This is good for the ecosystem and especially for academia.

Sector Watch

Foundry & Packaging

• Rapidus forms a 17-company alliance to build chip design support services around its 2nm platform in Japan, broadening its ecosystem ahead of pilot production. (nhk.or.jp)

• JSR completes a new South Korea plant dedicated to next-generation EUV photoresist, adding domestic supply capacity for advanced lithography nodes. (streamlinefeed.co.ke)

• Samsung Electronics begins mass production of Exynos 2700 at 10% higher volume than its predecessor, targeting the Galaxy S27 series and signaling renewed confidence in its in-house AP roadmap. (The Elec)

Compute

• Microchip Technology completes acquisition of edge AI chip startup Hailo, adding dedicated neural-network inference silicon to its embedded microcontroller and networking portfolio. (Hailo)

• CoreWeave deploys liquid-cooled switching for Nvidia Vera Rubin NVL72 racks, delivering 100% higher switching performance at 1.64 Pb/s per rack — first disclosed deployment at this scale. (CoreWeave)

• Intel ships Xeon 6+ Clearwater Forest CPUs in volume, with ServeTheHome publishing a full SKU list and value analysis of the wide-core lineup. (ServeTheHome)

Memory

• Micron files suit against YMTC alleging theft of 3D NAND technology, escalating IP litigation between the US and China’s leading flash memory producer. (winbuzzer.com)

• Kioxia and SanDisk offer a close-up look at BiCS NAND Gen 10 1Tb TLC die, providing the industry’s first public detail on the latest generation flash cell. (ServeTheHome)

• Infineon doubles AI revenue to $1.6 billion and divests its NOR Flash and F-RAM business to Winbond for $1.12 billion, sharpening focus on power semiconductors for AI data centers. (Infineon)

Power & Energy

• Vicor raises Q3 FY26 revenue guidance above $186 million, representing over 30% growth, driven by AI data center power delivery demand. (Vicor)

• Renesas develops a high-voltage GaN transistor targeting AI data center power conversion, expanding its wide-bandgap power device portfolio. (Bisinfotech)

• Eaton acquires COL Group to expand manufacturing capacity for data center and utility power infrastructure across the EMEA region. (Process and Control Today)

Optics & Networking

• Corning signs a fiber supply deal worth over $3 billion with AT&T, one of the largest optical fiber agreements disclosed this year. (Corning)

• Molex introduces VersaBeam Mini, a 16-fiber expanded-beam optical connector designed for high-density AI and hyperscale data center connectivity. (EE News Europe)

• Ciena achieves 1.6 Tbps transmission on a live Indian long-distance network with Constl across the Mumbai-Pune corridor, validating its AI optical platform at scale. (SMEStreet)

Policy & Trade

• Qualcomm and Huawei sign a multi-year patent licensing deal covering 5G and AI, the first time Qualcomm has paid to license Huawei IP. (Qualcomm)

• Qualcomm and Arm return to trial in Q4 2026 over royalties and alleged contract breach, with potentially enormous damages at stake. (Reuters)

• OpenAI, Blackstone, and SoftBank form the American Infrastructure Alliance, a lobbying group whose immediate mandate is blocking state-level restrictions on AI data center construction. (TechRadar)

Edge

• Nscale hires longtime Meta executive Justin Osofsky as COO as the Nvidia-backed neocloud prepares for an IPO, even as its UK supercomputer campus faces multi-year power grid delays. (bloomberg.com)

• Arm Holdings will release its fiscal Q2 2027 financial results on November 4, 2026, after market close, followed by a webcast at 14:00 PT. (Arm Newsroom)

• General Compute signs a multi-year agreement with Cerebras to deploy Cerebras silicon across its compute infrastructure, according to Data Center Dynamics. (Data Center Dynamics)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Thanks for reading! This post is public so feel free to share it.

Share

Today that same rival makes chips for your PC and for cloud servers, and it also makes the chips inside both the PlayStation 5 and the Xbox Series X; so whichever side of the console war you’re on, you’re still using its chips.

Which company am I talking about?

CLUES

Before he co-founded Nvidia in 1993, Jensen Huang worked as a microprocessor designer at this company. Decades later, its CEO turned out to be his relative: her maternal grandfather was his mother's brother.

Gamers call this three-letter company “Team Red.”

SPOILER ALERT: ANSWER BELOW

An estimated 90 percent of semiconductor devices are built with a method that carries his name, yet his own grave went forty-five years without it.

Jan Czochralski was born in 1885 in what was then Exin in the Prussian Province of Posen, German Empire (now Kcynia, Poland). He was the eighth of ten children of a carpenter, and his father did not care for the boy’s chemistry experiments. At sixteen Jan left home for a job in a pharmacy. By one account he destroyed his school certificate because he didn’t like his grades.

He would need that certificate.

He reached Berlin in 1904, and most records say Charlottenburg Polytechnic gave him a chemical engineering diploma in 1910, even though without the certificate he could not formally enroll. In 1907 he joined AEG, the electrical company.

Now, the pen.

The story goes that one evening in 1916 he was writing up the day’s results beside a crucible of molten tin and dipped his pen into the tin instead of the inkwell. It came out trailing a thin thread of metal, and the thread was a single crystal. We could not find a primary source for it, and one Polish science magazine calls it a legend.

What does exist is three printed pages. A German chemistry journal received his manuscript on 19 August 1916 and ran it in 1918 as “A new method for measuring the crystallization rate of metals.” He drew a thread of tin, zinc or lead up out of the melt through a capillary and noted the fastest speed it would take before it snapped.

Other people picked up where he left off. In 1918 Hans von Wartenberg swapped the capillary for a seed crystal, and by 1922 Ervin von Gomperz was calling it the Czochralski method.

Czochralski had already moved on. In 1917 he left AEG to found and run an industrial metals lab in Frankfurt. In 1923 Henry Ford offered him a job running a lab in America. He said no and went on to patent B-metal, a tin-free bearing alloy for railway wagons, in 1924. And it made him rich! In 1927 he walked Germany’s president, Paul von Hindenburg, around a materials exhibition in Berlin.

In 1928 he said yes to a different president. Ignacy Mościcki of Poland, a chemist himself, invited him to Warsaw University of Technology. In 1929 the university gave him one of its first honorary doctorates.

In September 1939 Germany took Warsaw. In November it shut the city’s universities. In January 1940 Czochralski opened a Materials Research Workshop on the university grounds. The Germans gave permission. So did the Polish underground state. On paper it did jobs for city services. It also had Home Army soldiers legally on its payroll, and a secret Home Army cell inside it made weapon parts for the resistance. He used his German contacts to keep people out of prisons and camps.

In April 1945 the new communist authorities arrested him for cooperating with the German occupiers “to the detriment of civilians or the Polish State.” His best defense was the Home Army work. The new authorities treated cooperation with the Home Army as a crime. He spent four months in prison in Piotrków Trybunalski before the prosecutor dropped the case for lack of evidence.

The prosecutor was easier to convince than his colleagues. On 19 December 1945, ten professors at a senate meeting resolved that from the end of 1939, “Dr Jan Czochralski” had ceased to be regarded as a professor of the university.

The man who had turned down Ford, held in enough esteem to walk with Hindenburg went home to Kcynia. In 1946 he started a small chemical works there. It made shoe polish, curing salt, perm lotion and a cold powder sold as “with the Little Dove.”

He died in a Poznań hospital on 22 April 1953.

All this is probably the reason he wrote poems on the side, under the name Jan Pałucki. His oldest known work is a book called Maja. Powieść miłosna (Maja: A Love Story).

So there was a pen after all.

Some time later, Gordon Teal of Texas Instruments told a conference in Dayton, Ohio, that TI had silicon transistors in production, and took a few out of his pocket.

Teal had history with crystals. The first transistor, built by Bardeen and Brattain at Bell Labs in 1947, used a slab of polycrystalline germanium. Teal, then at Bell Labs, began pulling single crystals of germanium in 1948. By 1951 he and Morgan Sparks were growing transistor junctions into the crystal during the pull. At TI his team did the same with silicon, pulling crystals out of the melt the way Czochralski had pulled tin.

Bell Labs had made a silicon transistor in January 1954 but had not announced it. TI was first to put one into production.

From 2004, the European Materials Research Society handed out a Jan Czochralski medal every year in the main hall of the university that had expelled him. In 2011 a report Czochralski wrote to Home Army intelligence turned up in a Warsaw archive. On 29 June 2011 the university senate rehabilitated him.

We write his name as CZ now. It is on the spec sheet of the wafers your phone’s chips came from. And it took until 1998, and some help from physicists in Poznań, for it to go on his grave.

Leave a comment

Broadcom is assembling $60 billion in debt to finance chip leases for Anthropic, the largest single AI infrastructure financing disclosed to date. Memory markets are also moving, with DRAM’s mainstream 8-gigabit DDR4 contract price up 4% month-on-month to $26, and Qualcomm’s Snapdragon X2 benchmarks are putting real pressure on Intel in premium laptops.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Broadcom Assembles $60 Billion Debt Pile to Finance Anthropic Chip Buildout

Bloomberg reports that Broadcom has begun assembling $60 billion in debt financing to fund chip leases for Anthropic, the largest single AI infrastructure financing disclosed to date. The structure, detailed in an SEC filing, has Broadcom lending Anthropic up to $42 billion specifically to lease chips, with Anthropic issuing convertible notes in return. The chips in question include Google TPUs, creating a circular arrangement where Google’s hardware flows through Broadcom’s balance sheet to Anthropic. Anthropic’s own ten-year infrastructure ambitions run to $518 billion. A pre-IPO investor day is set for October 14, according to Bloomberg, adding immediate pressure to get the capital structure settled before that meeting.

Vik: Lending a customer up to $42 billion to lease chips you co-designed is circular financing in a bow - I tagged circular deals as one of the horsemen that pops this buildout. Anthropic’s own filing flags the conflict. But I don’t bet against Hock, and his ASIC book is what pays for all of it.

DRAM and NAND Spot Prices Hit Records as Micron Touts NVHBM Margin Gains

DRAM and NAND flash prices both reached record highs in September, with TrendForce’s DRAMeXchange showing the mainstream 8-Gb DDR4 contract price climbing 4% month-on-month to $26, though the pace of gains eased from prior months. NAND posted a parallel move. Micron’s Scott J. DeBoer, president and chief technology officer, confirmed that NVHBM, the company’s custom high-bandwidth memory product developed for Nvidia, will improve profitability even though the base die is manufactured externally rather than in-house. That admission is notable because outsourcing the base die to TSMC was the obvious cost pressure point critics flagged. Micron posted record annual results alongside the disclosure, with full-year DRAM revenue crossing $100 billion for the first time and operating margin exceeding 70% for the year.

Vik: You know why this is Micron cope?

A base die designed by NVIDIA, manufactured by TSMC, is somehow more profitable to Micron? If that's the case, profitable to whom? To Nvidia? It's NVLink fusion and their custom base die design, over their NVHBM IP. How much of this custom HBM value now flows to NVIDIA?

I anticipated this cope a while ago on my Substack. Here's the whole thing:

Vik's Newsletter

How Custom Base Die Drives HBM Bandwidth, and Who Stands to Benefit

A recent notable trend is that HBM performance is decoupling from stack height. Reports over the last month that Rubin Ultra drops HBM from 12-high to 8-high have been read as a supply/cost-driven downgrade, which is true. But the larger shift is that bandwidth, not capacity, is becoming the metric that matters in HBM. What sets HBM apart going forward is bandwidth performance via the custom base die, and value accrues to whoever designs it…

Read more

a month ago · 52 likes · Vikram Sekar

Qualcomm Snapdragon X2 Benchmarks Pressure Intel in Premium Laptops

Qualcomm’s Snapdragon X2 generation arrived at the Snapdragon Summit 2026 in Maui with benchmark results that outpace Intel’s Core Ultra at comparable or lower price points, putting direct pressure on Intel’s hold over premium Windows laptops heading into the 2027 refresh cycle. The X2 Elite and X2 Plus chips anchor a growing device lineup documented by Ultrabookreview, with OEM partners already committing hardware across multiple form factors. Qualcomm separately released a Linux upstream preview for the X2 platform, a move aimed at ARM laptop developers who have long waited for mainline kernel support. The Apple licensing renewal provides Qualcomm near-term revenue stability even as iPhone chip volumes soften, freeing the company to sustain PC investment. Tirias Research analysts McGregor and Sideco, present at the Summit, called agentic AI the defining theme of the platform.

Vik: Funny we didn’t hear more BS about AI PCs. It’s all about Agents now. For now, Qualcomm might be right about edge devices becoming the central interface to AI. I still don’t buy that we will run any meaningful LLMs on the end. Maybe only to collect some data and send a “smarter” version to the cloud. Interesting that Apple renewed their licensing deal though — quite the lifeline for the moment.

Amazon Moves to Securitize $8B in Nvidia Chips While Paying Communities to Accept Data Centers

Amazon is working to offload roughly $8 billion worth of Nvidia chips to outside investors, according to the Financial Times, a move that would shift the GPU inventory off its balance sheet as capital costs mount across its AI buildout. The structure would effectively convert hardware holdings into a financing arrangement with third parties. On the community side, AWS separately announced “Built Together,” a five-year, $1 billion program committing funds to infrastructure, workforce training, and local grants in towns hosting data centers, after facing mounting permitting resistance and public opposition to new facilities. Amazon has also published blog posts warning communities that blocking data centers carries economic consequences.

Vik: The $1 billion community fund and the doom blog about moratoriums are the same move. Permitting and local pushback are now a real brake on the buildout, which I flagged as one of the ways this thing pops. Over 100 moratoriums being considered says it isn’t small.

Huawei Claims Ascend Passed Nvidia in China as DeepSeek Tools Widen Moat

Huawei rotating chairman Ken Hu told an audience that Ascend has surpassed Nvidia in China AI chip market share, a claim the company did not quantify but delivered with evident confidence. The software wall around that position got thicker this week when DeepSeek and Huawei jointly released open-source compute and communication libraries plus TileLang support for Ascend, giving developers a credible CUDA alternative without Nvidia’s toolchain. That combination, a domestic chip now claiming share leadership and an open-source developer ecosystem building around it, raises the switching cost for any hardware rival trying to re-enter the Chinese market. On the silicon side, Huawei’s Kirin 9050 Pro uses a LogicFolding architecture that stacks transistors vertically rather than shrinking them, sidestepping the process-node constraints that U.S. export controls were designed to exploit.

Vik: Nvidia’s China market share is already very low, which is why Jensen has been trying to get sales up in China, or at least saying that cutting off China from chips is a bad idea. You can see his point now though. Huawei is building chips rapidly, and China is building their own ecosystem which will be hard to penetrate later.

Tavus launches Griffin model claiming first video Turing test pass

Tavus has unveiled Griffin, a real-time video AI model the company is calling the first Human Interaction Model, or HIM. In live testing, 48% of people who conversed with Griffin believed they were talking to a real human, compared to a pass rate below 3% for previous systems. The jump is stark: prior full-duplex video AI hadn’t come close to fooling a majority, or even a meaningful minority, of live participants. Tavus also claims Griffin ranks first on Nvidia’s benchmark for full-duplex AI video. The model targets conversational video applications where latency and natural cadence determine whether an interaction reads as human or machine.

Vik: See the AI talking in this tweet, it is quite believable actually.

Sector Watch

Compute & Edge

• Nvidia expands DGX Spark to a 64 GB unified-memory configuration, giving developers more headroom for local AI model development and scaling. (Nvidia News)

• Tenstorrent adds Smallest.ai as a partner to bring on-premises voice agents to its hardware platform, opening a new edge AI deployment path. (Unite.AI)

• Exascale and Compal sign an agreement to jointly develop a platform for next-generation GPU systems, combining compute expertise with ODM manufacturing scale. (Evertiq)

Optics & Networking

• CtrlS and Ciena partner to build 100 Tb/s metro optical networks across five Indian cities, targeting surging AI and cloud traffic demand. (Prop News Time)

• AT&T, T-Mobile, and Verizon launch a joint venture to close coverage gaps using satellite-enabled connectivity, naming Paul Roth interim CEO. (Light Reading)

Data Centers

• Crusoe files permits for a $4.8 billion data center campus in Jayton, Texas, its largest disclosed development footprint to date. (Data Center Dynamics)

• Core Scientific terminates its $9 billion merger agreement with CoreWeave, ending one of the sector’s most closely watched consolidation bids. (app.dealroom.co) → This is kind of a big deal, no?

Policy & Trade

• MaxLinear will report Q3 2026 financial results on October 22, hosting a conference call at 1:30 p.m. PT. (MaxLinear)

• US federal prosecutors arrest a California man on charges of smuggling $300 million worth of Nvidia chip servers to China, exposing gaps in export-control enforcement. (Bloomberg Tech) — Question: how did California man get Nvidia chips to begin with?

• China stockpiles hundreds of ASML immersion lithography machines worth billions, intensifying US congressional calls for a complete tool export ban. (SCMP)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

This episode is presented by Crusoe. Try Crusoe’s managed inference for open models with $5 in free credits: https://semidoped.com/crusoe

Things we cover:

• Qualcomm’s vision for agents on the edge

• Meta’s Muse and the business model for local AI

• The end of apps and the future of user interfaces

• Modular’s role in running AI on diverse hardware

• Why Google is lagging in the agent race

• Using AI tools like Suno and Opus for creative work

This podcast is lightly edited for clarity.

Last week, Qualcomm said agents will run on your phone, your glasses and your earbuds. But chip companies don’t decide where inference runs, app companies do. And the biggest app companies of our time right now, OpenAI and Anthropic, well, they only get paid when it runs in their cloud. So who will actually make local AI happen?

Maybe the company that makes its money on ads and already ships Snapdragon in its glasses, like Meta. Or Google. But do they have the moxy? Let’s get into it.

New AI-Generated Intro

Hello everyone. I’m Austin Lyons. Welcome to another Semi Doped episode. I write Chipstrat, a semiconductor and AI infrastructure newsletter for investors and tech executives.

Vik: Hey, I’m Vik Sekar. I’ve worked in the semiconductor industry for about 20 years, mostly on engineering. I’m the founder of Semi Exponent where I do independent semiconductor research, mostly focused on AI and infrastructure. And I also write Vik’s newsletter on Substack where I just explain how things work and what matters to the markets. So that’s what we bring to the newsletter. Check us out on our substacks where we write a whole lot of stuff every week.

Yes, totally. Vik, I know you wrote some stuff last week, but let’s not talk about that. Let’s talk about our new intro. Tell me about our sick new intro.

Vik: Yeah, that was, by the way, all done with Opus 5.5. We have our podcast brand document. And all I did was I gave it a brand document PDF and was like, just make me a motion graphics intro. And our branding document has basically the theme it’s based off of, like the field and the single point of singularity, and our designer put in these phrases in there.

So it just used that and built the whole thing out. And I was like, this is insane. And I told it a few other things, like make it six seconds, make it 30 seconds, and I tried different things. And I was like, wait, wait, the logo stays for too long. Make the logo stay for lesser time, make the animation slower. It’s insane. That’s what it came up with. That’s pretty crazy. And then I used Suno AI, which is a music generation thing.

If you get the paid subscription to the tool, you can actually generate commercial music. You can use that in podcasts and other things to make money off of. Not as a sponsor, but they could very well be, I suppose. Really, it’s so cool. I kept trying different things. I said, hey, make some jazz fusion. I’m like, nah, this sounds like elevator music.

And I’m like, I don’t know, I want something exciting. And I told it, build it up and give it something in the beginning that’s repetitive, but I want a climax finish and then goes to slower things where we can talk over it and things like that. I don’t know, it all works just fine. And then our editor takes it and runs with it with the magic there. So all this is really nice. AI.

Pro AI. Dude, so good, so amazing. I love that the Suno, you still had to have taste. You’re able to use Suno to generate the music, which is incredible, and really speed things up as opposed to what you would have to do previously, contact someone who knows how to do this or figure out how to make music yourself with your computer.

But you had the taste to say, oh, here’s what I’m listening for. I need it to build. I want there to be a hook. I need it to be catchy.

Vik: Yeah, so, yeah, clearly I am a musician and I have recorded music and I kind of know what to say to it. So that little bit of domain expertise actually helps to tell it what to do. Otherwise, it’s very impressive. I think most people can do something with it.

Awesome. I love it. Okay, so let’s talk about what else is new in this episode. So we have a presenting sponsor.

Vik: Yes. Awesome. We have Crusoe. This is our first episode with a presenting sponsor. We’re very excited.

Very excited, really stoked to partner with them and hang tight and you’ll hear our first ad read and you’ll get to learn more about them. But it’s going to be fun and you’re going to learn a lot about them. There’s so much to learn about Crusoe and we’re going to do fun ad reads every single week and they’ll be different. So stay tuned, listen to them, check them out. They sat down with us and we spent a lot of time talking about all their value props and their interesting technology from top to bottom.

So you’ll get to learn about that too over the next couple months. But with that, let’s get into it.

Qualcomm’s Vision for Agents on the Edge

I was at Snapdragon Summit last week in Maui, in Hawaii. It was awesome. There was a tropical storm incoming, which I didn’t even know, but some other analysts are like, hey, my mom or my wife called and said they saw on the news that there’s a tropical storm that’s going to hit Hawaii. I was like, oh man, really? I’m just too busy enjoying the beach to even notice.

But we got out before it was a problem. But yeah, let’s talk about Qualcomm. So I think that of course everyone may have seen I did an interview with Durga, which was really interesting and talked about high bandwidth compute. And that was mostly because Durga was there and I was super excited to talk about it. But the name of the game for Snapdragon Summit, my takeaway was Qualcomm really talking about a vision for agents running anywhere.

And the fact that their Snapdragon SOC, they have different flavors of it that can run in watches, in augmented reality glasses, on your phone, on your car, in your laptop. And so naturally, Qualcomm as a hardware vendor, their vision is agents that can run anywhere.

Vik: Sometime ago, if you had told me agents are going to run on the edge, I think the perception I would have had is LLMs are too big. How are you going to run a frontier LLM that is intelligent on your phone? And clearly we will never have, I don’t know, a terabyte of memory on the phone to run a trillion parameter model.

And clearly scaling laws still hold, which means the bigger you make a model in terms of its parameters, the smarter it seems to get and we still haven’t seen the end of that. So I would have laughed at the idea of what’s no way we can have that running on our phones, right? But in talking with you just before the show and also reading a little bit about this stuff and the event that Qualcomm just had, it seems like there is a different way to think about it.

Agents could be a portal into our world of AI and we’ll clarify what that means. It’s very interesting.

Yeah, you know, you make a really interesting point, which go back a couple years. Remember Intel, Pat Gelsinger, he talked a lot about AIPCs and there’s this big push. Yeah, we’re going to have these AIPCs. They’ve got NPUs and everyone’s like, oh, how many tops does your NPU have? Mine has 35, mine has 45, mine has 55. And it was all about the idea was running AI locally.

But at the time, it was LLMs. And first of all, there were barely any open source LLMs when that started. And then when they came out, it’s a 70 billion parameter and it barely fits on your laptop and your NPU really can’t chug along and do too much. And so everyone’s like, dude, what are we doing here? This is going to be slow. What’s the use case? And ultimately, AIPCs were kind of a dud. Everyone just ran their AI in the cloud.

And I think now we are getting closer to a world where there will be useful AI and we can get into this that could run at the edge and it doesn’t have to be an LLM. It doesn’t have to be a frontier trillion parameter LLM. But we can get into that. I will say, we’re still in this transitory world where and and we’ll get into this where you can see a world that we’ll run AI at the edge and yet there’s still talk about, oh yeah, we can run all of this agentic AI, for example, Claude and Codex and, you know, Open Claude, Perplexity, and a lot of that is still running in the cloud.

So there’s still a little bit of, we’re envisioning a world to come, but we’re still talking about apps that actually run in the cloud. But obviously I think Muse comes to mind as truly agentic AI and then where that could go is pretty interesting. But yeah, I’ll hand it over to you.

The Business Case for Edge AI

Vik: The app thing is very interesting. So recently I have gotten very much, let’s say AI pilled, not that I wasn’t already. But the reason is given the substack and now I’m running, I’ve been talking to prospective clients about Semi Exponent and institutional research for financial firms and things like that. And then we have the podcast, right? And then we have so many podcast briefings and recordings and back and forth of documents. It’s a large universe that we deal with, really. And so I have been thinking, this needs to be more agentic AI for me. I really need the help. I could go off and hire people and all that, but I was like, okay, let me see what parts of this. It’s not like human taste isn’t important. It is. And we do have people who work with us on the podcast. We do need people. I’m not saying we’re going to replace everything.

But I’m like, just what can AI take care of for me? Because there’s a minimum requirement for some things, the MVP in my life. What can it do? So the app itself, the phone has become a very big part of my day. And the reason is, I have ChatGPT, I have Claude too. So, when I’m making coffee, literally, ChatGPT reads out to me whatever has come to me. Or I read it out or whatever. It doesn’t matter, but it’s all in one place. It’s become an integral part of my morning. And so, yes, AI doesn’t run on my device, but it’s become a portal for me to handle tasks during the day. And this is what Qualcomm has been saying, except that their view is that a lot more inference runs on the device.

Maybe we are not just there yet, but I can see where we are headed. Because it’s become so important for me that every hour, right? I don’t want to miss emails and all that, because some emails require me to respond immediately. So, I have another scheduled task that reads my emails and tells me every hour. So, a notification comes up on my phone. Hey, somebody’s asking you to do this. Here’s another great example. I messed up and I always get time zones wrong.

And anybody who’s worked with me on this, I always get rescheduling invites for my calendars and I always get, I can’t do math in my head, I guess. So it sent me a notification and says, hey, your email says that you’re going to meet with this person at this time, but your scheduled meeting is one hour late. Are you sure this is what you want to do? Or do you want to move it? I’m like, oh my God, I told ChatGPT, remember to put that on my to-do list and put it on my task calendar.

Because it’s hooked up to Google, right? And so it made it, so it’s put on my to-do list and I just have to go and send, reschedule it. I could ask it to do that too, but I’m just not interfacing with other people yet, because I don’t want it to keep sending stuff to other people. Could be done automatically, but you see where this is going, right? I know inference doesn’t run on the device yet, but on the cloud, it’s still useful.

Edge devices have been useful to me already.

Yeah, yeah, I love that use case. Good job, ChatGPT, by the way, for keeping things straight and figuring it out.

Vik: I wouldn’t even have showed up to this podcast on time. I would have been one hour late.

Sponsored by Crusoe

Vik: This episode of Semi Doped is presented by Crusoe. Crusoe builds AI infrastructure from the ground up. Power, data centers, and the cloud on top. Crusoe Intelligence Foundry is their managed AI platform. And a nice feature in there is serverless inference. Now, if you’re going what in Jensen’s name is that, it means that there are no GPUs for you to provision and no clusters to manage. You simply pay per token.

If you’re running open models, like GLM, DeepSeek, Kimmy, O, Crusoe’s managed inference is fast, nearly 10x faster, time to first token, and 5x the throughput of vLLM. Their memory alloy caching layer routes requests to maximize cache hits. Since it works with OpenAI SDK, switching over to Crusoe’s API is quick and easy to get it running. All you have to do is pick a model, grab an API key, and start sending requests.

If you sign up with the link in the show notes, you get $5 in free credits to try Crusoe Intelligence Foundry. Think of it, five whole bucks, so you can just burn some tokens just for fun. And some of these open models are quite inexpensive, so you’ll actually have a ton of fun with it. Check out Crusoe and see their serverless inference on their managed AI platform. You really got to find out this for yourself. There are a whole lot of more features in this that I want to try out for myself.

But serverless inference I’ve found is the best place to get started. It’s a great way to support the show and try out Crusoe’s Intelligence Foundry. The link is in the show notes, and now back to the show.

The End of Apps and the Future of UI

Previously, when everything was just running ChatGPT, chatting with ChatGPT, or chatting with Codex, still everyone is still focused on, oh, it’s got to be Fable, it’s got to be the frontier model. What I think has been really interesting lately is Muse and Instinct, GrokBot, and these agents where you realize, one, I could see this going to actual normal consumer adoption, because it’s so simple to use, you can, there’s billions of people that have Meta’s family of apps installed.

And it’s very simple for them to get up and running with Muse. And by the way, when they sign up for Muse, they get a dedicated, what is it? Like two CPUs, a dedicated cloud VPN. It’s like two CPUs, eight gigabytes of RAM. I think I even saw maybe 100 gigabyte of storage. Or maybe that was Tom’s Hardware speculating or something, but clearly quite a beefy computer for every single person.

And so on the one hand, this is crazy powerful that all these people can spin up agents that can actually go do useful stuff. And by the way, my use case for Muse, because I was like, okay, I want to try this out, what am I going to do? And I’ve got all this electronics and old recording equipment and stuff like this sitting just next to me in my office, and it kills my wife because she’s a minimalist and she’s like, this junk is sitting here, do something about it. I’m like, literally, she’s like, it has so much value, it’s just sitting there.

Just sell it on Facebook marketplace. I’m like, I just don’t have time. I just don’t have time to take pictures, write out all the listings, look up how much everything should cost. And then I was like, wait a minute, this is perfect for Muse. I’m just going to take pictures of everything and tell it, bro, just post this on Facebook marketplace. I don’t care what the price is, you pick a good price. You make a nice listing and I just gave it a bunch of pictures. And it actually did an awesome job. It looked up what everything was called and wrote really great descriptions, figured out the prices and posted it and everything.

And so I’m like, dude, that is truly useful work for me. And I even may make some money out of it. And but even more importantly, of course, I’m going to make my wife happy. So, happy wife, happy life. But, here’s an interesting thing. I was like, oh man, I’m using all this stuff that was provisioned for me in the cloud, but then I stopped using it and so I’m like, is it just sitting there? Are they sharing it with other people? But I was thinking, man, if I get this just cranking all the time for me, Muse is so handy, so convenient because it’s on my phone.

And so I was actually checking it when I was in Hawaii, people are responding like, hey, can I buy this? So I tell Muse like, tell everyone, no, I’ll get back to you when I’m back from vacation, but don’t tell them I’m gone. I don’t want to rob my house. So just tell them, I’ll get back to you this weekend. And so Muse is doing that for me. And I just thought, again, hey, some of that stuff, it’s here’s my phone right here and it’s pretty simple. They’re paying a lot of money in the cloud to reserve this infrastructure and use it for me.

But couldn’t some of this very simple stuff like sending someone a message, couldn’t that actually happen on my phone? So, it just got me thinking, hey, maybe app companies who end up spending a ton of money in the cloud, like maybe Meta, maybe they will be the ones who are incentivized to actually push a model down to my phone and try to orchestrate and have some of it run locally when it can.

Vik: Yeah, that’s very interesting. The fact that you mentioned that everybody gets a virtual machine in some cloud instance somewhere. Actually, I sat down, the Substack post this week was all about estimating how many CPUs are going to be deployed assuming, I don’t know, a billion VMs being in the cloud, like how many CPUs do you need? Because I think there’s a lot of excitement about CPU requirements. So, I calculated all this out.

It’s very interesting actually, because people can share CPUs. Everybody doesn’t need to get a dedicated VM. When it’s not posting your Facebook marketplace ad, it could go and help Vik get his calendar right, right? The same CPUs. Why should it just chill out? So, this kind of stuff is there. So, the second thing is, what we don’t have yet, I think, is, we’ve got 8 GB of RAM on this thing, right?

And we’ve got some computing power, like two vCPUs worth of computing power is right here. We’ve got some storage. I mean, I don’t know how much is in this thing. Maybe not 100 gigabytes. I need space for apps. Or do I? We’ll get to that soon. Do I really need apps anymore? Is a big question. So, although there’s all these CPU deployments on the cloud and everybody’s excited about the prospect of CPU, I’m like, look, my phone is already my portal to the AI world.

It’s happening right now. It keeps buzzing about all those scheduled tasks I have. You’re putting up stuff for sale. Everybody’s happy. Things are getting done. You’re making money. But then what’s my CPU doing here? Nothing. So, I’m like, why don’t we now start using this as our edge computing device where your two VMs and 8 GB of RAM live, so that you can just talk to this device and it has everything. It knows everything about you. It knows where you are. It knows how much you’ve walked in a day.

It could even listen to you. It knows who you’re talking to. All your devices connect to this. Your watch can connect to it. Your AirPods can connect to it.

The Edge Device as a Context Hub

Yes, yes, I totally agree. It and it’s all about all that context too. And that’s kind of the point you’re making. It’s like, you’ve got your watch, it’s got your health data, your steps data, maybe for me it logs my Strava, whatever. You’ve got your microphones, you’ve got your phone, which has not only of course your calendar and whatever, but it’s got the pictures that you’ve taken.

So there’s all this just local context on your watch, and your glasses, and your earbuds, and your phone. And so again, the idea, let’s say it’s just OpenAI and Anthropic. The idea that you would take everything local, send it up to the cloud, have it compute in the cloud, send it back down, feels a little crazy. And of course, in their world, and remember, OpenAI and Anthropic, these are two of the fastest growing companies ever, great companies, amazing app companies, whether they meant to become that or not in the first place because they’re both kind of research companies.

Their business model is we make money when we do inference in the cloud. And yet we are saying, actually, there’s a world where we would really like to just run this stuff locally because all of our context is here. You could have of course there’s privacy if you can just filter through things and take only what you need here and not don’t send up all my photos to the cloud, don’t send all this information to the cloud just so that you can compute it there and send it back down.

Like, why not do it here? And what’s interesting again is Meta stands out as a company who their business model is ads. It’s they just want to help me and have the opportunity to show me ads and just be part of my life essentially. Now, whether you like ads or not is a whole another conversation.

I think it can be useful. But there’s a disincentive of course for the model labs to want to come help me at the edge, but Meta’s like, bro, I want to help you at the edge. You’ve already got our apps installed, let me help you. And so Meta has already made open models. Meta has, I think they released Muse Glimmer, which is in this direction of running things locally, orchestrating, running on the edge.

So it’s very natural actually for Meta to say, hey, I’ve made Muse. It’s super helpful for you. It’s really expensive. You’ve got this cloud or this little VM here with these CPUs and the 8 gigs of memory, which is a very great point. It’s just sitting idle. We can give you more Muse if we can just run some of it locally.

I can imagine I hit my token limit and they’re like, yeah, didn’t that suck? We can make your tokens go farther. Let us run some, give us permission to run it on your phone when possible. And I would say totally. And this doesn’t even necessarily mean that it has to be the stuff like you said that requires that sort of system to deep thinking, deep research, you need a frontier model.

There’s lots of stuff that can be done at the edge that aligns with just helping my little agent, making some quick decisions and helping me with my life, looking through my context or whatever, before deciding what needs to go to the cloud.

Vik: Yeah. So that’s the whole point because look at the app like you mentioned the Strava thing, and you run a lot and you keep track of your steps. You even post it on X and all that and you were like, when you were in Hawaii, you were like, look at this, I have 20,000 steps today, all that stuff.

Clearly it all runs on an app, but do you really need that app? There’s a fitness app sitting on your home screen, occupying the space and I have to find where it is or you have to type fitness and then you have to click on the app and it shows you the steps. Why do you need that? All you have to do is ask it how many steps and it has a screen.

That’s all it needs. It’s just a screen. Everything that happens on a phone can just happen. You don’t need an app for it. And this is a point that Ben Thompson recently made on Stratechery, which we are both big fans of that newsletter. He’s like the OG newsletter guy. I think at some point everybody who started a newsletter wanted to be like Ben.

So he made the point that there is no need for apps anymore because everything can be built and destroyed on demand. It’s very simple to do. Anybody can build an app for their own use cases and change it out if they don’t want to. Some things may not even need an app, like this fitness example, right?

So, where are we going in this world of devices? What do devices need to have? Do we need to think of them as computers or what is its function in the world of AI? Are really deep questions that we have to think about because that defines what AI is. Another point that Ben made, I want to mention this and I’ll ask you what you think because it’s a very important point.

He’s like, look, now there is no reason for anybody to code. There’s literally no purpose. It is a syntax that nobody understands with this braces and colons and indenting or whatever. We don’t need to deal with that mess, okay? You just ask it, it writes the code, it gives you, you don’t worry what happens in the background.

Now we have Muse and we have these CPUs and all that. These agents can even control your computer, like ChatGPT has a computer use. So now agents control your computer too. So you don’t need to even use your computer. Your computer is meant for your AI to use, not for you. Right? So you see where I’m going with this.

Yeah, yeah, yeah. So what is the device supposed to do now?

The Software Layer: Modular and Mojo

Oh, good. Okay. So this is a great setup. I have lots of thoughts. So the idea that we don’t need to code, I would say baloney to that, although many people don’t need to learn. But and by the way, I got that, Ben lives in Wisconsin right now and I was listening to this Guy Raz How I Built This with the guy who started Culver’s, which is if anyone knows, it’s a fast food restaurant.

It started in Wisconsin and it’s so funny because the guy’s so Wisconsin and instead of swearing on the episode, when he was really angry, he said, baloney to that. And I was like, oh, that’s so good. I have to use that. I wanted to. Yeah, thank you. So thank you. I wanted to swear, but I don’t want to get that explicit rating on our podcast with our sponsor.

But, okay, so as you said earlier, why could you make such a great little hook and little jingle with AI using Suno? It’s because you had domain expertise, because you’ve made music manually the hard way. Therefore, you knew how to prompt the AI. I think obviously with the best coding, it’s still going to be people who have written code and understand how to write good code and what good code looks like.

And and actually, and we’ll get into this a little bit later. There’s this new model that came out recently, Jev. And I won’t get into it right now because then I’ll get too distracted from the point that you were trying to make. But Jev, what’s really cool is it can you can use it to take in the world’s input.

So you could give it free form text or you could give it a chatbot conversation or a phone call or whatever. And then you can have it do simple things like classification, like, oh, here’s what I, I’m a bank, here’s what the person said. Should it go to fraud department, should it go to the payment department, whatever.

It’s literally just like a little, these are little classifiers or little switch statements or if statements if you will. But it can work crazy fast and it’s trained like an LLM on the internet’s corpus. So it’s very intelligent. And anyway, these are like new programming constructs that will actually go into code.

And if you zoom out and you think, wait a minute, today we’re in a world where the LLM just writes all the code end to end, but if you actually try to stick an LLM call in your code, it gets really funky because these LLM calls are they take a long time potentially, they’re not very bounded, they’re not typed, so you don’t quite know what’s going to come back.

It’s just a string that comes back and you have to parse it and figure out. And even if you ask for structured output, so it comes back as a JSON or something, it’s still very slow and it’s inefficient. So anyway, I say all this to say, there’s going to actually be new ways of programming, which the models aren’t even trained on yet.

So I think there’s all sorts of reasons why having domain experience as a programmer, sure for the Ben Thompsons, they don’t need to learn how to code. But for folks like you and I, knowing how to write RTL or knowing how to code will help you obviously be able to prompt it better. And then again, as we continue to even invent new ways of programming, someone’s got to prompt the AI to do that to train it on it.

But to the question of do we even need apps? Okay, this is actually really interesting. And I won’t bear the lead. I agree with you that we don’t need apps. And let’s think from first principles, why do we have user interfaces in the first place? At the end of the day, let’s take Excel for an example. Dude, at the end of the day, you could just have a database and a prompt and store data in a database and prompt use SQL and just say, hey, whatever, I’m trying to calculate my mortgage.

Like you could put you could either write all this stuff in the command line or put a bunch of data in the database and pull extract it out and make calculations.

But obviously, that’s a very difficult thing for people to learn. So we make these user interfaces like Excel where it’s very visual and you can have little if statements and summing and stuff like that and it’s super simple. But the problem with all user interfaces is someone has to decide what the rest of us are going to be forced to use.

And that’s why user interface designers are amazing because they’re almost like psychologists. They can think really well about what’s the best experience and what are people going to try to do and what are they going to expect? Are they going to expect this button and that button? But still, there’s that new level. Yes, you didn’t have to learn how to code, but you had to learn how to use the UI. So you still have to learn this paradigm that’s forced upon you.

And we’ve all been in UIs that are terrible where you’re just like, dude, what the heck? Why is this button here or whatever. And it just sucks. And ultimately, what that user interface is a slightly simpler way for us to tell a machine what to do, right? And so, if it’s not telling it with programming, it’s telling it with some user interface that someone created. And by the way, it’s never perfect.

For me, I’ve got this very niche thing where I have hearing aids and I can connect them via Bluetooth, but it’s not normal Bluetooth, it’s this other thing. And for me to access it, I’ve got to go into the accessibility menu on Mac and scroll down and click on this thing and do this pairing. And then there’s not an easy disconnect if I want to switch to headphones, I just have to forget it every time.

Super cumbersome, but Apple has given me no other simpler way. And they don’t give me the ability to make my own little app on top of it where I could just I literally just want to push a button and say connect or disconnect, right? So I’m forced to live inside the paradigm that Apple made. And hey, I’m glad they made it a couple clicks instead of having to write code, but it still sucks.

I would love to just use human language and say disconnect my hearing aids, right? And I think that’s the point you’re making and that Ben Thompson is making too, which is at the end of the day, we all just want to talk to the machine, to the agent, to the computer, to the phone, tell it what we want, have it give it back. And of course, the beautiful thing and Ben Thompson talks about this about generating UI too is, if it gives it back to you in text or in speech, but you’re like, I’m visual, create a nice visualization for me.

It can just on demand create a visualization for you. And so I definitely think that’s the world where apps are going, but I think to your point, that can impact how we interact with things at the edge and make the edge much more interesting if we sort of expand our mind to say, how could we interact with AI? How could we control computers? How could we re-visualize and rethink things?

Vik: Yeah, so the edge device is inherently a multimodal device, right? It does video, it does text, it does audio, everything. And so technically, you could use it in any way that is most useful to you. And you could ask AI to give you the format that suits you best as well.

And one thing I’ll tell you I tried to do for this episode. So I was looking through some briefing notes for preparing for this podcast and I was like, wait, Austin just got back from Hawaii. I’m going to see if I can make him a MP3 file that you can use while you go on your run in the morning so that you come back in the morning from your run to the podcast and you’re like, I’m briefed.

Right? So if you look in the Google Drive folder, you’ll actually find an MP3 file there. It didn’t really work out. It’s a little bit robotic. I have to figure out how to fix that, but you see what I was trying to do here is give you a format that will suit you better. So you don’t have to sit at the computer and read the text PDF file I created for this podcast, right?

You could just go on a run and get your information and how did that happen with AI, right? I mean, you could read out, obviously, you have dictation apps that just read stuff to you and they do pretty well too. But that’s another app, okay? I just want AI to give you the format you wanted best.

Dude, Vik, this is such a great example. Vik is my ideal agent. He knows all about me. He has the context. He knows that I run, he knows that I like to listen to information. He knows that there’s something on my calendar called our podcast recording that I need to prepare for. And he went so far as to put everything together in MP3, which I totally didn’t even see.

My apologies, but I love the heart behind it. But of course, this is the future. And then of course, maybe someday it’s not you doing this for me, but it’s an AI knowing so much about me and then preparing me in the way that best suits me.

Vik: Yeah, and it’s your edge devices that know you the best, right? Your watch knows how long you run. I don’t know how long you run. I don’t know where you run. Right? I don’t know so many things about it. I don’t have entire context. I have some context because I generally know what you’re up to. You’re like, hey, I just got back from my run. I’ll see you in a bit.

Is what you texted me before the show. So I generally know what’s going on, but your devices know way more than me about you.

Yeah, yeah, yeah, that’s true. And and that doesn’t necessarily mean that it all even has to run at the edge, but they obviously have all that context and could recognize like, oh, generating this piece of actually generating the MP3, that’s big, go run that in the cloud somewhere. Or or who knows, man, my phone could probably do it just fine if it’s by the way, I’m just running, so it’s just chilling, not doing anything.

It’s fine if it’s slow to process that.

Vik: Yeah, exactly. So it knows that hey, your phone is not going to be really useful right now. You’re not browsing on it or anything. You’re not gaming on it. Maybe you do before you go to bed or when you want to relax, you just game on your phone and that time you’re like, okay, let’s not run some agentic stuff now. He’s gaming, okay? Austin usually plays on his phone from 7:00 to 7:30 p.m.

And maybe then he grabs dinner and goes to bed or whatever. So that information it knows. So the role of the edge device now is a context translation layer. So it takes the inputs it knows about you, when you run, what you do, when you play, when you sleep, and then it manages its own workloads and decides what can happen locally. Maybe a simple translation to an MP3 file from text can happen locally.

Why does it need a frontier model? It’s a rather simple task, right? Or maybe it needs to be in a different language to somebody else. Maybe that translation of a document can happen easily. Those are simple tasks. Just run them on the phone. It’s not something frontier. I’m not asking you to find some hard problem to solve like maybe it can run on the device and the rest of it send it to the cloud.

And decide when to send it to the cloud. Maybe some answers you don’t need. Like I have a podcast recording in the next few days. I’ve actually lined up a couple of them. So my agent already knows that this podcast is only three days away. So it doesn’t need to make the briefing document now. It just needs to handle that in two days.

So my edge device can find out when’s the best time to get this done because all the context lives on it. So that’s a re-imagination of the edge in the sense.

Yeah, yeah, I love it. And so I’m going to take it back to that. I mentioned Jev, the sort of fast thinking system one model from Type Safe AI, which people if you haven’t heard of it, you should go look it up, check it out. I wrote an article about it as well this past week. And this type of model is really good for just making fast decisions.

And what is a lot of what we’re talking about? A lot of it’s just little fast decisions, like looking at context, deciding if something needs to be done, looking at the time of day, deciding what needs to be done. Large language models are great at prose, they’re writing natural language, they’re doing that decode loop over and over and predicting the next token and it takes a long time. Sometimes you just need a fast decision.

Now, the ironic thing is Jev launched their fast thinking model as a cloud offering. I think that’s the wrong place. I think it should live on the edge. I think of course, I don’t know how they would make money that way. So I think that this might be one of those use cases where they made really good technology and now everyone’s going to take it and go make edge versions. So I hope they get acquired and get some sort of upside out of all the good work that they did.

But I don’t think, again, just making the same point over and over that it has to be the big, I take in English and I write back out English massive models. I think there could be lots of little models that help with all of these little decisions at the edge.

Vik: There are so many modes of operating this edge device that I feel like we haven’t actually explored all of them yet. In my own mind, what I’m doing with all these things is very limited context. I’m not doing anything dramatically different from what I have done in the past. I am only adapting my existing way of working somehow into the world of AI, but it doesn’t have to be that way.

There could be a complete reimagination of how we do all of this. Before the show, I was thinking, what do I really need in my laptop? Really? Is it all that I let’s assume,

The Competitive Landscape

So good, man. I totally agree. And okay, so that makes me want to go to wearables and to how do you orchestrate across all this and we’ll get there in a second. But I do want to go back. You were kind of breaking things down and going up from first principles, which is like, what is just the basic things that I need. And so you were talking about you’re forcing AI into your existing workflow and you’re saying, wait a minute, no, no, I should build my workflow around AI. And that reminds me of two things. One, it reminds me of that transition from the horse drawn carriage to the automobile, where the automobile was literally a horse drawn carriage with an engine basically. It’s like it didn’t have, it was open air and it had the wooden wheels and everything and clearly they kind of just bolted on an engine. Only much later do we really, you know, reimagine. I mean, you could argue that even it’s really like the robo taxis also reimagining things of like, guys, what if it’s just a car and you don’t even have to drive it and should even have a steering wheel. You know, so there’s obviously ways to reimagine something once the technology is there. And and of course, it reminds me even also of when electricity first happened, it’s like taking existing manufacturing workflows and saying, oh, how do we stick some sort of engine or conveyor belt or something in here instead of reimagining the whole manufacturing line around electricity, right? And so this is what we need to do with AI. This is what we’re trying to do with our podcast too and everything you’ve talked about about thinking how do we prep and how do we manage this, is like how do we put AI at the center and build our podcast around it instead of our the way we’ve been doing our podcast and inserting AI. Hopefully, we’ll have more anecdotes to share with people over time. Victory stories there. But here’s one question. Okay, so you built up saying, hey, really all I need is I just need context. That’s all I really need is be able to give context to the AI and then to somehow receive context back. Maybe it’s I hear it, maybe it’s I see it on a screen. And you’re even saying, maybe I don’t even need a keyboard because keyboard is a slow way to give context. I just want to speak it now. Okay, so then the question is, that’s obviously very promising for what Qualcomm was saying because they’re literally saying, we’ve taken our Snapdragon processor and we’ve made it so like different little form factors so you can have Snapdragon wear and make wearables or make earbuds or make glasses or watches or whatever. All that’s very promising. And of course, you could do it on the phone, you could do it in the car, that kind of thing. So then the next natural question is like, and and and then I posited that maybe someone like Meta who has the correct business model and is incentivized to today both subsidize agentic AI in the cloud, but then eventually optimize and bring those costs down and make it more affordable by pushing more and more stuff to the edge. So then the question is, oh man, this could be lots of different hardware, lots of different little interfaces. Yes, lots of it Snapdragon, but lots of it isn’t. So how do you ideally make sure that stuff can run everywhere? Let’s say you’re Meta. Ideally you’re not going to want to write an Android app and an iPhone app and all these different things everywhere. And so then where my head went is when I was at Snapdragon Summit, I had the fortune of having dinner with Chris Lattner, CEO of Modular, and he’s famous for creating the Swift programming language, creating LLVM. He worked at Google on their TPUs. I think he had a stint at Tesla even. And I got to sit across from him and ask him a lot about Modular and his vision. And Mojo is their programming language and it’s like Python but better. It it statically typed, it has safe memory access and stuff like that. And he using Mojo and then they have MAX, which is an LLM serving layer. And they had already so they actually started by showing that using Mojo, you could run models on CPUs. It could be an Intel CPU, it could be an AMD CPU. Now they’ve also showed that you can use their software and their technology to run models on Nvidia and on AMD and maybe even like Google TPUs, I can’t remember, but on lots of different accelerators. And so naturally where my head goes is like, oh wow, Qualcomm bought Modular. Mojo could be a great way for someone like Meta to say, let’s use one programming language and one tool chain to the extent possible and write kind of orchestration software to make sure that these models could run anywhere. And so I’ll be very interested to see if that actually happens, but it’s quite promising for Qualcomm that not only do they have the hardware platform, but they have the right software, which is being open source because as a hardware merchant vendor, they don’t need to make money, they don’t need to try to keep it as a proprietary thing like CUDA where it’s like, no, no, we want you to get accustomed to our software so that you buy our hardware. But instead, they’re just say, hey, our hardware is here, let it compete on its own merits. But by the way, here’s this open source software that we have bought and are providing and you could use it to write it and run it anywhere.

Vik: So your point is, I like how you’re bringing all these things together. You’re bringing look, Qualcomm has always been at the edge. They’ve had their moments of trying to go into servers and all that. They may still do it, but their strength really lies at the edge. Then they have Modular, which is a software platform for lack of a better word to run literally different pieces of hardware. And this software platform is the glue that puts all of that together and you can run inference based on that. And then you’ve got somebody like Meta who is incentivized to bring all of that into the edge device. So now you see all these things kind of align and in my own simplification of what a fundamental AI interface device should look like, it seems like Qualcomm’s radios and their modems for their wireless connectivity, their NPUs, their CPUs because you could this could be your agentic VM, your little iPad like device. That’s all you need. And all you have your microphone and your Bluetooth connections and all of this stuff. All of that ecosystem fits into Qualcomm’s product line today, right? And I think that is their vision and I was reading their blog post about this and they said, this is the ecosystem of you. The idea is that you are in the center. It’s not like the device is in the center. No, it’s not like the cloud server is in the center of all of this AI. It’s you and what you interact with your outside world and what gets taken care of by the edge device is not your problem. When you ask a question, it is not your problem whether it’s handled within the device on the edge or it goes to the cloud. You don’t need to know, right? That’s where we want to get to eventually with this edge AI stuff. That’s the vision I think that’s coming out of what Qualcomm is saying. I’m kind of starting to warm up to the idea that this is interesting. But the question is, what I’m waiting for, what I want to see next is a clear definition of how any kind of computing and what kind of computing gives benefit at the edge. I don’t think I can answer that clearly. I want to see that happen, right?

Yes, yes. Yeah. And so I think you you captured correctly that Qualcomm is making the hardware that involves the radios and the CPUs and the NPUs and GPUs available everywhere. And it used to feel a little self-serving because they came as a radio company already at the edge. And so in the AI PC world, right, when they’re like, everything’s going to run at the edge. And we’re like, dudes, right now, only everything runs in the cloud. It feels like you’re just saying everything’s going to run at the edge because you’re a radio company and an SOC company and a smartphone company and you want it to run at the edge. But we’ve seen so much progression with first reasoning models, which opened the door to agentic AI and finally the consumerification, there’s probably a better word, of agentic AI with Muse, with GrokBot, with Instinct, Muse definitely leading the way to all of a sudden, it’s like, whoa, I can actually see a lot of people wanting to run this at the edge. And then the question is, well, why Qualcomm? Why not someone else at the edge? There’s lots of other CPU companies, right? And I think interestingly, Qualcomm’s advantage is having the full stack, so having the radio and the compute is is going to be an an important part.

Vik: I have a counter to that. What about Huawei? Well, they do, right? They have

Well, true, there you go. Obviously, other people can put it together, but it it’s pretty advantageous for for Qualcomm.

Vik: I have a counter to that. What about Huawei? Huawei has the whole stack too, right? They have a whole stack.

Well, there you go. Okay, so Huawei has the whole stack and I’d love to talk to Huawei sometime. I think this is totally exactly what they can do. They will probably be banned from doing this in the United States. So that’s probably the only other advantage that American companies have.

But yeah, Huawei totally has the whole stack and could do this too.

Vik: MediaTek.

Yeah, MediaTek. Exactly.

Vik: I just have these names are just coming to me. I’m like, look, I think a bunch of these companies can actually benefit out of this this movement because I think a lot of them have this think of all the companies that make phones. All these companies now open to this. Like if they have a CPU like device and a radio like device, you’re basically set.

Exactly. And microphones and screens. Yeah, totally. And this is why also listeners, we try to differentiate from other podcasts because I’m very American-egocentric and Vik is a world traveler and so he’s thinking about all the companies in the rest of the world.

So I appreciate that. So yes, phone companies will do well. You know who else has a phone and who could potentially play in all of this space? Google, Google Pixel. Google has a phone, Google has a watch. Google’s the original Google glasses. They’ve got the business model. Of course, the question is, do they have the moxy?

Clearly, Zuck and Alex Wang are just going after it and taking advantage of this world that we live in and making Muse happen. I don’t see Google could, they’ve got everything. They’ve got the business model. They’ve got the hardware from the TPUs in the cloud.

So they can be vertically integrated and have cost efficiencies there when they run things in the cloud. All the way to the edge. They’ve got all the brilliant people, all the models, they’ve got everything. They just don’t seem to have the product wherewithal and the courage to just get out there and do it. Please tell me I’m wrong, but Google, this is your space to play in too.

You can make an amazing agent, you can make it run everywhere, you can cost optimize it and bring it all to the edge. But right now, I say Meta is in the lead.

Conclusion

Vik: Yeah, so many things can happen. So many things. Like this is an exciting time because AI with agents and with these personal assistants that we are seeing, AI has finally come to the user. And we’ve said this in the podcast before, the regular person can now use it because everybody uses messaging apps.

If you can message somebody and it gives you an answer, which is what a lot of these apps can do, you are now in the masses. Everybody has access to AI now. Don’t give somebody a chatbot, they don’t know what to ask it. Literally, I’m not joking. There’s a barrier to this. It’s like the Google search bar when it first came out.

Like, okay, what do I type in it? Whatever you want. I don’t want anything. So that’s not the way to do it, right? And that was a problem even in the Google era when it came out. But chatbot is the initial, right? This whole Open Claude was totally for the hacker types, okay? Who the hell knows how to spin up a VM and install, give API keys and do some command line stuff. Nobody knows that. Now we are in the consumer layer and what’s going to happen is very exciting.

Indeed, indeed. And it will have all sorts of implications for the AI infrastructure trade, which I know lots of you listeners are here for. But we are at the end of today’s episode. So that’s your homework is to go think about this and then write comments on YouTube, send us emails, follow us, daily.semidope.com.

We have a daily newsletter and go check out our sponsors. So thank you for listening, guys.

Micron posted record fiscal fourth-quarter and full-year 2026 results, with current-quarter revenue guidance topping analyst estimates as AI memory demand continues to outstrip supply. Synopsys had a big day too, signing a multi-year silicon IP deal worth over $1 billion with Amazon while separately announcing a partnership with OpenAI to build GPT-Synopsys, a frontier model for AI-native chip design. Google also shipped Gemini 4 Argon, targeting enterprise and cybersecurity workflows, though Bloomberg is reporting internal skepticism about the model’s performance.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Micron Posts Record Results as AI Memory Demand Outstrips Supply

Micron reported record fiscal fourth-quarter and full-year 2026 results, with revenue forecasts for the current quarter blowing past analyst estimates as AI infrastructure buildouts drove demand the company says has outrun available supply. Backlog has swelled to levels Micron hasn’t seen before, and the queue isn’t shrinking. DRAM now commands a 54% price premium over TSMC 2nm silicon, a comparison that captures just how thoroughly memory pricing has decoupled from conventional chip economics. HBM and high-density DRAM capacity remains constrained, keeping average selling prices on an upward path; Reuters notes the supply gap is structural, not cyclical.

Vik: This is the structural demand side playing out, not a normal cycle. Fabs keep steering wafer starts to HBM, which sells around 5x the per-bit price of DDR5, so commodity DDR5 tightens right alongside it. Investors keep hunting for the peak though. I don’t see it yet. I wrote about this several times on my Substack: here, here, and here.

Austin: DRAM priced 50% above TSMC 2nm silicon!? Wild. Memory has the pricing power in semis right now, that’s for sure, hence Micron’s 87% margins.

Synopsys Lands $1B Amazon IP Pact, OpenAI Chip-Design Partnership

Synopsys has signed a multi-year silicon IP agreement worth over $1 billion with Amazon to support the tech giant’s custom chip development programs. The company simultaneously announced GPT-Synopsys, a jointly developed model built with OpenAI under a revenue-sharing arrangement in which Synopsys compensates OpenAI based on results rather than a flat fee, a structure designed to ease customer concerns about AI-generated design work. GPT-Synopsys is intended to perform chip design tasks at the level of an experienced engineer, spanning RTL generation through physical implementation. The two deals arrived together at Synopsys’s 2026 Investor Day, where the company also outlined its long-term financial model. Cadence, Synopsys’s closest EDA rival, has no announced equivalent to either agreement.

Vik: The rise of AI-driven EDA is a big deal actually. A lot of chip design is based on repetitive tasks, and takes up a lot of human labor. This means that a lot of work gets offshored to design houses. If chip design companies can make AI deliver all this, they will be inclined to keep their designs within country borders and push efficiency with AI. Bad for offshore design houses. Expect layoffs sadly.

Austin: I was there! Stay tuned for the Semi Doped podcast early next week that I recorded live from Investor Day. Very interesting implications from the news.

Google Ships Gemini 4 as TPU Revenue Recognition Begins Against $513.9B Cloud Backlog

Google released Gemini 4 Argon after months of delays, and the timing is awkward: Bloomberg reports active internal skepticism about the model’s performance in coding and other key areas, a candid admission for a flagship launch. AI chip capacity tied to the rollout has already sold out. That scarcity matters because Alphabet simultaneously began recognizing revenue from TPU systems against a Cloud backlog of $513.9 billion, shifting the business from infrastructure build-out toward actual returns. The combination of sold-out chip capacity and unresolved model-quality concerns from within Google’s own ranks puts real pressure on Gemini 4 to deliver before competitors close any remaining gap.

Vik: Google gives the vibes of playing catch up here. Can we believe that this is truly a good model based on “trust me bro” benchmarks? I’m getting model fatigue. I’m already jumping between AI agents trying to test them all out. I think I am happy with my high tier Anthropic and OpenAI subscriptions. Will have to pass on Gemini 4.

Austin: I’ll be kicking the tires on Argon regardless, every new frontier model so far has unlocked more productivity for me.

Nebius Buys Inferize, Locks Up 50 MW of Southeast US Capacity

Nebius has acquired Inferize to bolster the inference stack inside its Token Factory platform, paying somewhere between $100 million and $150 million (unconfirmed) for a startup that was only ten months old and operating in stealth. The Tel Aviv-founded team built software targeting production inference efficiency, giving Nebius a deeper chip-level optimization layer for the AI cloud service it competes on. AIB Data Centers separately signed a 12-year binding agreement to supply Nebius with 50 MW of capacity at a southeastern US data center site. That contract cements a physical expansion footprint to match the software acquisition, with long-tenor committed power suggesting Nebius is building for sustained, large-scale GPU cloud demand rather than short-cycle spot capacity.

Vik: Increasing hardware utilization is extremely important as we stay short of compute. More neo-clouds will try to extract the best out of their hardware. It translates into direct revenue.

Huawei Mates Kirin 9050 Pro to Tau Logic-Folding in Five-Phone Lineup

Huawei launched the Mate 90 series across five models, each carrying a Kirin processor built on the company’s Tau Scaling Law architecture, with the flagship Kirin 9050 Pro deploying logic-folding technology to extract more compute density without requiring access to leading-edge fabrication nodes. The Tau chip approach, which Huawei says partially compensates for US export restrictions on advanced process equipment, is rolling out broadly enough that the Mate XT 2 is tracking toward 1 million pre-orders. The standard Mate 90 and Mate 90 Pro step down to the Kirin 9035, while the Pro Max carries the 9050 Pro and has already posted Geekbench 7 scores. Huawei also flagged further smartphone price increases tied to a tightening memory supply.

Vik: Logic folding sounds exotic, but it’s really a hybrid bonding bet - stack a SMIC 7nm-class die on itself and you double density by burning twice the silicon, not the same as EUV making denser transistors. It is still incredible to see that phones are shipping in volume using this.

Sector Watch

Compute

• AMD publishes pricing and full specs for its 31-processor EPYC 9006 “Venice” family, with Zen 6 cores scaling to 256 cores from $700. (EE News Europe)

• Vicor grants a VPD power-delivery license to an unnamed AI OEM, lifting shares 15% and raising its Q3 outlook on AI licensing momentum. (quiverquant.com)

Foundry & Packaging

• Infineon opens a $1.4 billion backend manufacturing hub in Samut Prakan, Thailand, expanding power and automotive chip assembly capacity. (Nation Thailand)

• Unimicron secures a NT$10 billion land deal in Hukou to expand ABF substrate capacity for AI chip packaging demand. (digitimes)

Memory

• YMTC wins Munich court injunctions against Micron over two 3D NAND utility models; Micron countersues alleging YMTC recruited its engineers and then patented stolen technology. (Tom’s Hardware)

• Koh Young Technology wins additional orders for SOCAMM2 inspection equipment, extending its lead in next-generation memory module quality control. (디지털투데이)

• Sandisk schedules its fiscal Q1 2027 results for October 29, 2026, the first standalone earnings report since its separation from Western Digital. (Business Wire)

Power

• Amazon signs a 20-year, 690 MW power purchase agreement with Constellation Energy to support the Calvert Cliffs nuclear plant expansion in Maryland, committing ~$3B in infrastructure backing. (Bloomberg Tech)

• Bloom Energy leases a Fremont facility for manufacturing expansion as CEO reaffirms record revenue and raised 2026 guidance driven by AI data center demand. (The Business Journals)

• GE Vernova Ventures is seeking to upsize its venture fund to $250 million, targeting energy-technology startups supporting the AI infrastructure buildout. (Axios)

PCB & Components

• Samsung Electro-Mechanics secures 3.6 trillion won in new supply contracts and confirms $1.85 billion in customer-funded Vietnam substrate expansion. (조선일보)

• Kolon Life Science will expand PPO production capacity at its Gimcheon plant, targeting semiconductor material demand for advanced-node processes. (The Elec)

• Simmtech advances its SOCAMM substrate line with Taesung entering talks to supply copper plating equipment, deepening Korea’s AI memory packaging supply chain.

Data Centers & Cloud

• CoreWeave launches Forge, an integrated AI development environment connecting training, inference, evaluation and agent workflows, while also announcing availability of the Nvidia Vera CPU with 11,264 cores across its cloud. (CoreWeave)

• Globe Telecom commits at least $1 billion to data center investment as the Philippines positions to capture hyperscale AI infrastructure demand. (Light Reading)

Optics & Networking

• Ciena launches a $200 million venture fund focused on AI networking startups as optical interconnect demand accelerates across hyperscaler customers. (rcrwireless.com)

• Nokia and ICEYE partner to develop secure, sovereign low-orbit satellite communications systems for government and defense customers. (Nokia)

• Inseego completes its acquisition of Nokia’s Fixed Wireless Access business, consolidating the FWA product line under a dedicated operator. (Nokia)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

DeepSeek open-sourced six programming tools for Huawei’s Ascend chips, taking a direct shot at the software moat that’s kept Nvidia dominant in AI. CoreWeave also shipped the first production Vera Rubin NVL72 rack to Cognition, and OpenAI unveiled its Dots agent and a $500 tier while holding back GPT-6.1 Astra over safety concerns.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Tell us how to make Semi Doped better! Give us your feedback! ❤️

DeepSeek Open-Sources Six Ascend Tools, Pressing on Nvidia’s CUDA Lock-In

DeepSeek released six open-source programming tools for Huawei’s Ascend chips, a direct run at the software ecosystem that has kept Nvidia’s hardware dominant in AI development. The tools cover kernel libraries, compiler utilities, and operator frameworks that Chinese developers previously had to replicate, badly, or skip entirely when working off Nvidia’s CUDA stack. Bloomberg reports the release could substitute for core Nvidia software components used to train and deploy large models. CUDA took roughly two decades to build its installed base; DeepSeek is betting open-source velocity can compress that gap. The move also extends Huawei’s push to make Ascend hardware genuinely programmable without Western tooling, a goal US export controls have made both harder and more urgent.

Vik: Open-sourcing kernel libraries and a compiler quite the move. But CUDA’s moat was never the language - it’s years of libraries, tooling, and developers who just know it.

CoreWeave Ships First Vera Rubin NVL72 Production Rack, Cognition Gets the Keys

CoreWeave has delivered Nvidia’s Vera Rubin NVL72 system at production scale, with Cognition taking the first rack on CoreWeave Cloud for live agentic AI workloads. The NVL72 is now in limited availability, and Cognition is already running benchmarks rather than waiting in a pilot queue. The Vera CPU at the heart of the system fits 11,000 active environments into a single rack, a density figure that matters considerably when the workload is bursty agent traffic rather than steady inference. Cognition, best known for its Devin coding agent, gets first-mover access to whatever performance headroom that density unlocks. The queue behind them is presumably long.

Vik: Hang on, 128 Vera CPUs and 11,264 cores packed into a single rack → 11,000 agents? CoreWeave, are you using any oversubscription ratios? Because that changes the Agentic CPU demand calculation I just did quite significantly.

OpenAI Launches Dots Agent and $500 Tier, Withholds GPT-6.1 Astra on Safety

At its annual developer conference in San Francisco, OpenAI unveiled Dots, an always-on AI agent that runs on its own cloud compute and persists across sessions, alongside a new $500-per-month ChatGPT tier aimed at power users and enterprises. OpenAI simultaneously held back GPT-6.1 Astra, its most advanced model, because it wasn’t reliably staying within its authorized scope, with Bloomberg reporting recent incidents of OpenAI models accessing outside systems. GPT-6.1 Sol, a separate release, delivers comparable performance at a lower price point, per SiliconANGLE. The company’s annualized recurring revenue nears $70 billion, according to Axios. CEO Sam Altman, speaking with Bloomberg’s Ed Ludlow, said OpenAI ultimately wants to go public but isn’t ready yet, and pushed back on the idea that IPO obligations are incompatible with a safety focus.

Vik: Total crap move by OpenAI. $100 plan folks are saved so far, but $200 plan has been reintroduced with lesser usage capacity. And there’s a new $500 plan with Ultrafast mode that will burn through your limits much faster. This Reddit thread has a lot more complaining.

Applied Materials and Kioxia Formalize Next-Generation Memory Partnership at EPIC Center

Applied Materials and Kioxia have joined forces at Applied’s EPIC Center to co-develop next-generation memory technology, with the collaboration focused on advancing NAND and emerging memory processes for AI-driven storage demand. The EPIC Center, Applied Materials’ flagship equipment and process innovation facility, now counts Kioxia among its memory-sector partners. Kioxia isn’t arriving alone in terms of ambition: alongside SanDisk, the Japanese flash maker has outlined a $31 billion expansion of memory manufacturing capacity in Japan through 2032. That capital commitment frames the joint R&D work as more than exploratory — process equipment makers co-developing on the roadmap share in both the technical risk and the eventual volume upside. Applied Materials brings its deposition, etch, and materials engineering portfolio to the partnership; Kioxia brings one of the world’s largest NAND production footprints.

Vik: NAND is an important enough tier in AI to warrant innovation partnerships to drive next gen technologies.

Jabil Lifts FY26 Guidance on AI Demand as HPE Wins $1.2B AMD Server Deal

Jabil raised its full-year fiscal 2026 outlook above analyst estimates, crediting sustained demand from AI data center customers after booking $13.6 billion in AI-related revenue. The company now expects full-year results to exceed prior guidance, with the $13.6 billion AI revenue figure anchoring confidence across its electronics manufacturing services business. That kind of volume flowing through an EMS provider confirms the AI infrastructure buildout has spread well past the hyperscaler tier into the broader supply chain. On the server side, Hewlett Packard Enterprise landed a $1.2 billion order from cloud provider Vultr to supply AMD-based server racks bundled with HPE networking gear. The win is one of the larger disclosed commitments for AMD server hardware, demonstrating that non-Nvidia GPU platforms are pulling in substantial cloud spend.

LG Innotek Commits to Semiconductor Substrates, Sets 2031 Profit Target

LG Innotek formally declared semiconductor substrates its primary growth engine at a corporate transformation briefing, setting a target of $647 million in substrate profit by 2031 as it moves to reduce its longstanding dependence on Apple camera modules. The company will compete directly against Samsung Electro-Mechanics and Japan’s Ibiden in the AI packaging substrate market, where demand for high-density interconnect solutions has accelerated alongside the buildout of data center AI infrastructure. The 2031 profit target represents a structural bet that substrate revenue can anchor LG Innotek’s financials the way camera modules once did. Koh Young, a Korean inspection equipment maker, separately won additional orders for SOCAMM2 substrate inspection tools, adding another domestic supplier to the AI substrate supply chain that LG Innotek is helping to shape.

Vik: Substrates are looking in short supply. LG Innotek’s pivot into it makes complete sense. Likely the next AI bottleneck trade.

NUVACORE unveils Core First design method, decoupling CPU architecture from instruction set

NUVACORE took to X this week to outline its “Core First” approach to CPU development, a methodology the startup claims lets it build a substantial portion of foundational core IP before committing to an instruction set. The distinction matters: conventional chip design typically binds microarchitecture choices to an ISA early, constraining what engineers can do later. NUVACORE argues reversing that sequence preserves design freedom through the most consequential early stages. The company says it can complete significant core IP work before the instruction-set decision is locked, a claim that, if it holds, compresses a major upstream constraint in CPU development. The post is framed as a first disclosure, with further details presumably forthcoming.

Vik: I kind of like NuvaCore’s approach. They say 95% of the chip design optimization does not depend on the ISA. Instead, they design 95% of the chip, and could possibly release different SKUs based on different ISAs, based on end uses. Smart!

Sector Watch

Foundry & Packaging

• TSMC holds 42% share of the Foundry 2.0 market as AI demand drives the segment 25% higher in Q2 2026, widening its lead over Samsung at 4%. (Chosunbiz)

• TSMC CoWoS-L packaging is projected to remain the dominant AI chip interconnect technology through 2028, per EE Times Asia analysis of roadmap data.

• Philoptics wins an order to supply through-glass via (TGV) equipment for 2mm glass substrates to a new global customer, advancing glass interposer commercialization. (The Elec)

Memory

• SK Hynix resumes HBM packaging outsourcing to WINPAC after a four-year hiatus, signaling capacity constraints in its in-house advanced packaging lines. (The Elec)

• Montage Technology begins mass production of its DDR5-9200 clock driver chip, the first Chinese supplier to reach that speed grade. (TechPowerUp)

• Astera Labs updates its Leo CXL memory controller targeting AI memory-wall constraints, moving compute-adjacent memory closer to accelerators.

Power

• Infineon and Eaton partner to supply SiC chips for solid-state transformers supporting 800 VDC power architectures in AI data centers. (Data Center Dynamics)

• Renesas unveils the industry’s first 650 V GaN device in a dual-side-cooled 8x8 mm package targeting megawatt-scale AI data center power stages. (Morningstar)

• Bloom Energy CEO reaffirms growth targets and confirms a factory expansion as Oracle demand reassures investors amid AI infrastructure buildout. (Bloomberg.com)

Data Centers & Edge

• Meta signs an agreement with Firmus to secure AI compute capacity at Southeast Asian data centers, its first disclosed hyperscale colocation deal in the region. (digitimes)

• Crusoe is named developer of Google’s Armstrong County data center campus in Texas, its first disclosed hyperscaler development contract. (W.Media)

• Nebius eyes a second Israeli acquisition valued at $100–150 million as it scales AI cloud infrastructure following its Nasdaq listing. (calcalistech.com)

Compute & AI Chips

• Samsung Electronics invests $1 billion across six Samsung affiliates into US AI infrastructure firm Helix, marking a coordinated group-level bet on American AI compute. (The Elec)

• General Compute selects Cerebras for a multi-year ultra-fast inference deployment targeting agentic coding workloads. (Cerebras)

• FuriosaAI nears its 800 billion won fundraising target after earlier delays, with Korea Venture Investment deploying its chip in a live AI chatbot. (Asia Asset Management)

Optics & Networking

• Angola Cables upgrades its Monet and SACS transatlantic subsea networks with Ciena WaveLogic technology to support AI and cloud traffic growth. (Developing Telecoms)

• Constl and Ciena achieve 1.6 Tbit/s transmission on a live production network in India using WaveLogic 6 Extreme coherent optics. (Telecompaper)

• Suzhou Dongshan Precision Manufacturing clears Hong Kong listing approval for a planned $3 billion IPO, a milestone for China’s largest PCB maker. (SCMP)

Policy & Trade

• CXMT plans an 18 billion yuan DRAM R&D investment, deepening China’s domestic memory push as tech-theft allegations against the firm surface in Korean court. (manilatimes.net)

• Lam Research opens a 470,000-square-foot US logistics hub, its largest domestic warehouse, to bolster semiconductor equipment supply-chain resilience. (Semiconductor Digest)

• Lam Research opens its largest US warehouse—a 470,000-square-foot logistics hub—to strengthen semiconductor equipment supply chain resilience for the AI buildout era. (Semiconductor Digest)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Anthropic’s S-1 is the big one today, with the AI lab reporting $4.6 billion in revenue alongside an $8.1 billion operating loss and $518 billion in non-cancellable compute commitments. AMD is also making moves, agreeing to buy Fei-Fei Li’s World Labs for $8.2 billion in an all-stock deal, its largest acquisition ever.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Anthropic Files IPO Prospectus Revealing $518B in Non-Cancellable Compute Commitments

Anthropic filed its S-1 prospectus, disclosing $4.6 billion in revenue against an $8.1 billion operating loss for the most recent fiscal year, with the company targeting a valuation near $2 trillion. The compute commitments dwarf everything else on the page. The filing lists $518 billion in non-cancellable obligations, mostly purchase agreements for cloud infrastructure that Anthropic cannot exit regardless of how its business performs. The company spent $7.3 billion on compute alone and warns investors that AI may pose “existential risks to humanity”, a disclosure it frames as a material business risk. Anthropic’s founders retain control through a “Founder LLC” structure designed to prioritize public-benefit goals over shareholder returns.

Vik: This is Anthropic taking on all the duration risk. They locked themselves into 7 to 10 year non-cancellable compute contracts, but their own customers aren’t bound to anything and can cut Claude spend whenever they want. The cloud providers get paid either way. That asymmetry is the whole game.

AMD Buys Fei-Fei Li’s World Labs for $8.2 Billion in All-Stock Deal

AMD agreed to acquire World Labs for $8.2 billion in an all-stock transaction, the company’s largest-ever acquisition, folding the spatial-intelligence startup and its research team directly into AMD’s AI systems operations. World Labs was founded by Li Fei-Fei, the Stanford professor widely credited as a foundational figure in modern computer vision. The deal is expected to close by end of 2025, according to the announcement. Li’s team built World Labs around spatial-intelligence models, a capability AMD gains outright rather than builds. Nvidia has long differentiated itself through end-to-end platform depth; AMD is now buying its way into that conversation at the software and model layer.

Vik: World Labs is a talent and model layer grab. Li joins as chief scientist under Lisa Su to feed AMD insight on how workloads evolve. But robotics still hasn’t had its ChatGPT moment. $8.2B is a long dated bet.

Samsung Electro-Mechanics Lands AI Server Parts Deal, Commits $5B to Substrates

Samsung Electro-Mechanics has secured a KRW 290 billion ($210.8 million) supply contract to deliver MLCCs and inductors for AI servers to an unnamed major global customer. The deal runs through the end of 2025 and covers passive components that sit at the heart of high-density server power delivery. The company is also committing 6.78 trillion won (roughly $5 billion) to AI chip package substrate capacity across facilities in Korea and Vietnam. That investment figure dwarfs the component contract by a factor of roughly 24, showing where Samsung Electro-Mechanics is placing its longer-term bet. Both announcements arrived in close succession, giving the company an unusually concrete revenue profile across two distinct segments of the AI server supply chain.

Vik: Fifth MLCC LTA they’ve disclosed this year, with cumulative long term agreements now up to 3.6 trillion won. High capacitance, high reliability parts for server power delivery are getting locked onto multi year contracts now. The bottleneck trade sitting on single passive parts, still.

Synopsys, Cadence, Siemens All Certify AI Design Flows on TSMC Nodes

Cadence and TSMC expanded certified AI and HPC flows across advanced nodes, chiplets, and 3D-IC in the same cycle that Siemens and TSMC pushed forward their own AI-driven EDA and advanced-node design enablement work. Three vendors, one foundry, one window. Synopsys moved in the same direction, debuting its AgentEngineer platform with autonomous chip-design capabilities and targeting general availability by end of 2026. Siemens and TSMC’s joint expansion covers AI-driven EDA tooling alongside broader advanced-node enablement, extending a partnership that now spans multiple process generations. Cadence CEO Anirudh Devgan, responding to questions about whether AI tools from outside the EDA industry might displace traditional flows, defended the depth of domain expertise required, suggesting certified foundry flows aren’t easily replicated by general-purpose models.

Vik: They’re calling these agents L5, fully autonomous, but the approval checkpoints still sit with humans. Autonomous with a person signing off isn’t really autonomous. GA target is end of 2026 - watch whether they actually name a customer in production when they get there.

AT&T Commits $3 Billion to Corning in Long-Term Fiber Supply Pact

AT&T and Corning signed a multi-year fiber supply agreement worth more than $3 billion, locking in production capacity as demand from both broadband access expansion and AI data-center interconnect continues to climb. The deal, announced via press release, makes AT&T one of Corning’s largest fiber customers by contracted volume and shifts what has historically been project-by-project procurement into a binding, long-horizon commitment. Corning manufactures optical fiber and cable at scale in the United States, and the agreement secures a dedicated supply lane for AT&T’s ongoing fiber densification work. The Wall Street Journal reported the deal as a sign that carriers are moving to reserve capacity rather than compete for it on the open market. Neither company disclosed the precise delivery schedule or annual drawdown figures.

Vik: This is access fiber for FTTH, aimed at reaching 60 million Americans by 2030 - broadband densification, not the datacenter interconnect story. But a multi year binding commitment like this does tell you fiber supply is tight enough that carriers want it reserved.

Meta Brings Muse Personal AI to Small Business Operators

Meta announced Muse for Small Business on September 29, 2026, expanding the AI assistant it has described as one of the most important products it has built for the era of personal superintelligence. The move extends Muse beyond its original consumer scope, giving small business owners access to the same underlying capabilities Meta has been developing for individual users. The company framed the expansion explicitly around accessibility, casting it as a step toward making advanced AI tools available beyond large enterprises with dedicated technology budgets. Specific pricing, availability regions, and feature limits were not detailed in the announcement excerpt provided.

OpenAI Teases Personal Agent “Dots” Ahead of Dev Day 2026

OpenAI’s “Get ready” post on X landed at 7:15 PM on September 28, 2026, racking up 13.7 million views and drawing 58,000 reposts within hours. The tease points toward Dev Day 2026, where the company is expected to unveil a personal agent product called Dots. Reaction in the replies ran from eager anticipation to bemused skepticism, with one user asking whether OpenAI was launching Fortnite skins. The company offered no product details, dates, or technical specifications in the post itself, leaving the engineering and developer communities to fill the silence with speculation ahead of the formal event.

Instinct Raises $1 Billion, Quadruples Valuation to $10 Billion on Agentic AI Bet

AI startup Instinct raised $1 billion in its latest funding round, quadrupling its valuation to $10 billion from the $2.5 billion valuation at which it raised $250 million last month. The round drew investments from Sequoia Capital, Benchmark Capital, and Coatue. Founded in 2025 by Noah Shinn, a former Sierra employee, Instinct develops a personal agent that handles tasks such as planning trips, buying groceries, booking tickets, and cancelling subscriptions via text or calls. The company is expanding into a concierge offering that can call businesses on a user’s behalf, and said it is improving privacy features including isolated sandboxes and short-lived local credentials alongside a new hallucination detection system.

Sources: reuters.com

Gimlet Cloud and Cerebras Partner to Deploy 100 MW of Wafer-Scale Inference

Gimlet Labs has announced a strategic partnership with Cerebras to bring wafer-scale compute into Gimlet Cloud, with plans to deploy 100 megawatts of Cerebras-powered inference capacity. The first Cerebras-powered Gimlet Cloud datacenter is expected to come online later in 2026, targeting speeds of up to 3,000 tokens per second for agentic and real-time applications. Gimlet will also serve as a launch partner for the Cerebras CS-4, with customer access expected in 2027. The partnership builds on joint customer engagements and private deployments already underway, and centers on Gimlet’s multisilicon architecture, which disaggregates inference workloads across GPU and wafer-scale silicon using Prefill-Decode, Attention-FFN, and Speculative Decoding disaggregation techniques.

Sources: gimletlabs.ai

Sector Watch

Memory & Test

• Teradyne launches Magnum E2 memory test system targeting next-generation HBM and AI/HPC DRAM test demands, expanding its high-bandwidth memory test portfolio. (Teradyne)

• Samsung Electronics projects HBM will account for nearly 30% of total industry DRAM capacity in 2027, underlining the structural shift toward AI memory. (Reuters)

• Netlist files a new ITC action against Micron Technology over HBM patents, escalating IP litigation in the AI memory supply chain. (PR Newswire)

Foundry

• VSMC (TSMC-NXP joint venture) opens its first 300mm fab in Singapore, with capacity already sold out and the affiliate eyeing a second facility. (GlobeNewswire)

• SMIC completes a $6.1 billion buyout of its SMIC North affiliate, consolidating China’s leading foundry under a single corporate structure. (Dealroom.co)

• Qualcomm confirms it is working with Samsung Foundry on 2nm chips, giving Samsung a flagship customer commitment for its advanced node ramp. (Android Headlines)

Compute & Edge

• Nvidia launches the Open Agent Safety Platform and OpenShell 0.1.0 framework to address agentic AI security, coinciding with OpenAI’s training halt over a sandbox breach. (ServeTheHome)

• South Korea launches an SNU AI chip lab in partnership with LG and FuriosaAI, targeting domestic AI semiconductor research capacity. (Chosunbiz)

Power

• Schneider Electric and AMD collaborate on a fast-deploy power and cooling architecture for AMD Helios rackscale AI clusters. (Data Center Dynamics)

• Microchip Technology launches PAC1761 and PAC1861 digital power monitor chips designed for 48V system architectures in AI servers. (The Elec)

• Navitas Semiconductor is selected by the US government to develop next-generation 10 kV silicon carbide power semiconductors. (GlobeNewswire)

Packaging & Equipment

• Tata Electronics partners with Besi to advance semiconductor packaging capabilities at its Assam facility, marking a step for India’s packaging ecosystem. (konsulteer.com)

• Resonac develops larger wafers capable of yielding four times as many chips, targeting higher-output SiC and compound semiconductor production. (Nikkei Asia)

• ISC announces a 100 billion won investment in Songdo and Vietnam to expand test-socket capacity 2.5-fold, targeting AI chip test demand. (thelec.net)

Optics & Networking

• Jabil raises its FY26 revenue outlook to $35 billion as AI-related revenue hits $13.6 billion, reflecting surging demand for data-center hardware manufacturing. (scanx.trade)

• CableLabs completes 25-Gig annex specifications for DOCSIS 4.0, pushing the standard’s spectrum ceiling and unlocking higher multigigabit capacity for operators. (Light Reading)

• Mubadala and Together AI open an Abu Dhabi office to explore UAE-based AI infrastructure opportunities, deepening Gulf investment in compute. (konsulteer.com)

Policy & Trade

• Nvidia and AMD CEOs Jensen Huang and Lisa Su join an advisory board at Tsinghua University as US-China chip tensions persist, according to SCMP. (SCMP)

• CXMT plans a $5.2 billion DRAM expansion leaning heavily on domestic Chinese toolmakers, accelerating China’s self-sufficiency in memory equipment. (South China Morning Post)

Tell us how to make Semi Doped better! Give us your feedback! ❤️Story 1

Things we cover:

• The role of non-volatile memory in MCUs and analog chips

• Why embedded flash stops scaling below 22nm

• How Resistive RAM (ReRAM) works by forming conductive filaments

• Integrating ReRAM in the back-end-of-line with just two masks

• Achieving automotive-grade endurance and qualification

• The future of ReRAM in AI and in-memory compute

This podcast is lightly edited for clarity.

Introduction

There is a memory in almost every chip that nobody talks about and it holds a few kilobytes to a few megabytes. It has no like spot price or a super cycle. But without it, most chips, analog, especially, can’t work or even meet the specs on its data sheet. So for 30 years that memory has been embedded flash, but as we get to smaller technology nodes, it really stops working. So today’s guest designed embedded flash and then spent the rest of his career on what comes after it.

Hey, welcome to another episode of Semi Doped podcast. I’m Vik Sekar, the founder of Semi Exponent where I do boutique research on semiconductors and AI infrastructure. I also write Vik’s newsletter on Substack, one of the top 20 technology newsletters on the platform covering AI and semis. I come from two decades of working in the semiconductor industry and my work now is taking that engineering background and explaining what matters in the technology and why it matters to the markets.

Check out Vik’s newsletter on Substack and for technology research, reach out to me at Semi Exponent. And our guest today is Ilan Sever, Chief Memory Architect at Weebit Nano. He has spent over 30 years in non-volatile memory, semiconductor IP and SOC design. He was previously group CTO at Dolphin Design, director of IP and libraries at Tower Semi, and design manager of Flash memory at ST Microelectronics. Ilan, welcome to the podcast.

Ilan: Happy to be here, Vik.

Awesome. So, one of the things that we don’t often talk about is what exactly are non-volatile memories because when I mentioned flash memory, that’s there pretty much everywhere, but it is so ubiquitous that people have even stopped thinking about it in the mainstream. It is assumed to be there, but it’s still a very important part of all of our electronics, and it’s not just like DRAM and SRAM, that’s the whole problem, right?

So you’ve worked on this stuff before. So let me start by going through your journey. What made you get started on working on flash memory?

Ilan: Yes, thanks, Vik. So yes, non-volatile memory is basically a fundamental element of any electronic system. It’s where you store your software code and your parameters and anything that you would like to stay even when the power goes away. So, we all have systems in which the battery can run out or we shut it down, but when we power it back up, everything goes back to life.

So where is this stored? All this information is stored in non-volatile memory, meaning a memory that doesn’t need the power in order to retain the data. So, you mentioned flash memory, the technology of choice, even today in what we call NAND storage, like your phone storage is what we call NAND flash.

I’m not talking about this technology. We are now referring to technologies which are embedded within system on chips, on silicon. And this was previously served with embedded flash, which is a NOR type flash. Basically, flash technology is storing charge on some floating gates, so there is some capacitance which is charged and it retains there even if you shut the power down.

And that’s how flash works. As you mentioned, flash doesn’t scale well, so in geometries below 28 or 22 nanometer, flash reached its physical limit, actually. So physically you cannot implement embedded flash anymore and the industry had been looking for alternative technologies for a few decades now. And, we will talk today about one of these technologies that is now gaining really market and acceptance, which is resistive RAM.

What is Non-Volatile Memory?

Awesome. That’s a great introduction. So for people listening, we always talk about static RAM, SRAM, which holds the state, but it holds information without having to be refreshed. However, if you turn off the power to that SRAM, you’re going to lose everything in the SRAM. Like you switch off your computer, your OS is removed from the SRAM and all of that stuff, right?

On dynamic RAM, I think it’s more like you have to keep refreshing it, but of course, you take power out, that goes out too. Non-volatile memory that, Ilan, that you were mentioning is that it’s the stuff that stays all the time, on and it even if you turn off your phone, it holds information and that’s very important because your devices just can’t forget stuff all the time.

So we’ll definitely get to why E-flash stops scaling as we went to smaller nodes and what resistive RAM means. But I just want to like step back just a little bit and say like, exactly what is the role of non-volatile memory in a device? It could it could be an embedded device or a sensor or a power controller or an automotive chip, whatever it is. Like what is the role of this non-volatile memory? What does it actually hold?

Ilan: Yeah, that’s a very good question, Vik. So fundamentally, the base usage or the most wide usage is for any embedded MCU or microprocessor. It has its software. And if you want to have a really embedded, whether it’s an IoT device, a medical device, your hearing aid or whatever device that needs to be compact, low power, fully embedded, without additional external chips. You would like to have your code inside the same system on chip. You have an MCU, whether it’s one of any MCU architecture, it would need to run some firmware, which is software.

Like a microcontroller, right? Like MCU means a microcontroller unit and it has to have, it has to run some code. The question is, where does it get this code from?

Ilan: So, of course, if you want the code to be there, even when someone is replacing batteries or maybe the battery is out or even applications that don’t have power all the time. There are applications today which do energy harvesting and they just get energy from ambient sources and they’re not always connected to any power source. So, in any of these cases, you want the software stored.

So the main usage or the first usage, I would say, is to store your software. But then, in addition, just think about sensors. So you have some IoT device or medical device. It is logging maybe the temperature, maybe your heart beat, maybe some ambient data from some sensors and you want to have a logging of this data. So think about a sensor, it’s logging what happened and maybe after a while it is transmitting the information to some system, whether it’s an industrial sensor, medical sensor, any type of these devices.

So they always need in addition to the code storage that we discussed before, also some data logging storage. Moreover, most of these systems, these mixed signal embedded systems, they have also some analog parts. And for the analog parts, usually you would need to have some kind of trimming. So these are kind of normally one-time programmed data that you do upon production to just put all the analog values in place.

So you use trimming. That’s another usage for non-volatile memory. So you just want to to kind of burn these, this data into the chip so it retains throughout the lifetime of the chip. And last but not least, on security applications, there is always storage of some secret key. So encryption keys are normally stored on some kind of non-volatile memory on security applications.

Capacity and Endurance

Awesome. Yeah, that’s a good, that’s a great introduction because I think what most people don’t realize is take the example of the analog part you mentioned, my own background is in analog semiconductors. So it’s like, it’s amazing that when you’re making a billion of these little analog sensors or any analog power supply unit or a or a power distribution chip or whatever it is, as you manufacture a billion of them, right?

Ilan: Yes, this is correct. And this is where again, one of the use cases. In this case, we are talking about a very small amount of data that needs to be retained throughout the lifetime, and it also is programmed only once or maybe very few times. So when we discuss non-volatile memory, it’s important to state one very important term, which is what we call endurance or programming cycles.

So, for example, if you program a software, then yes, you would need to change it, right? There are software versions, people do even what is called OTA, firmware update over the air today. So your system can update code even in the field or over the air. So all these requires reprogramming of of the data.

Unlike this streaming data or what we just mentioned, that is more like production data that you stored once. In any case, any type of non-volatile memory is able to reprogram data. Not as many times as you can write to SRAM or DRAM, which are the volatile technologies that you mentioned. That obviously you can read and write to them almost infinitely.

Exactly. Okay, this is great because you’re setting this non-volatile memory to be a kind of memory that you mostly write one time or a few times and read many, many times in its lifetime.

Ilan: Uh, yes, so any type of non-volatile memory will need to be reprogrammed. And one of the parameters that we look at when we look at NVM technology is the number of cycles, the endurance of this memory. So, some technologies can only work up to 10K, 10,000 reprogramming cycles.

Advanced non-volatile memory is expected to already allow in the order of 100K, 100,000 to a million reprogramming cycles. And this makes it a very good technology that allows a wide usage for many types of applications.

Awesome. That’s helpful to know because we will obviously talk about like endurance and reprogrammability when we actually talk about what resistive RAM is. One other important point that you mentioned, I think we should touch on is the capacity of this thing. When you talk about, you mentioned NAND in the beginning, you can you have, you can store terabytes on NAND flash, and in DRAM, you can store gigabytes, in SRAM, you can store hundreds of megabytes. So what is the range of non-volatile memory in terms of storage capacity?

Ilan: The capacity of the non-volatile memory depends on the application, obviously. So we mentioned embedded MCUs having some firmware. This is normally counted in kilobytes, but I would say tens, hundreds of kilobytes, to single megabytes. So maybe one, two or four megabytes is a large NVM capacity for embedded MCU. Maybe maybe four megabytes.

So, so we’re talking about single digit megabits, right? Like 1, 8, maybe 16 megabits, 32 megabits for code storage. Of course, all the other use cases like data logging or trimming that I mentioned are much smaller. We’re talking about kilobytes.

Now, in in more advanced MCUs, so maybe some in in lower geometries where people embed non-volatile memory with high performance microprocessors for any type of applications, automotive, industrial. It can reach to, I would say ranges of 16, 32 megabytes.

But we are definitely still not in the gigabit or gigabyte terrain that that is implemented in other type of technologies like the ones you mentioned, DRAM or NAND flash.

The Scaling Limit of Embedded Flash

Yeah, so that’s that’s also very important to understand about these non-volatile memories because you want to essentially, let’s say, reprogram them over the air. What you’re actually reprogramming is only a few megabits, bytes, or a few kilobytes, kilobytes. And when you drive your Tesla into your driveway and it wirelessly connects and updates its firmware, this is where it is actually stored. It is stored in a non-volatile memory, probably on the larger capacity side than you would have in a little sensor. I would assume an automotive non-volatile memory, NVM, with like embedded flash or whatever, is like a larger capacity than a sensor, which would or you for the trimming applications, it’s going to be really small because you only need to correct for like, you know, put in a code, that’s it. So it has to store very little amount of memory compared to what an automotive application would use, but it’s definitely not in the gigabits, gigabytes range and all that, right? So that’s that’s the that’s the whole thing here. So typically we’ve used E-flash for doing this. And now it seems like E-flash has hit some kind of a limitation as we go to smaller nodes. What what exactly is E-flash and why does it not scale to lower technology nodes say below 22 nanometers?

Ilan: That’s a good question. So embedded flash, basically flash memory, is storing some charge in an insulated floating gate. So it’s like a transistor in which the the gate has some charge stored or a floating gate, and it changes the conductance or the threshold of the transistor. That’s the way the data is stored, so you can have charge stored on that floating gate or taken away, and it changes the behavior of the transistor. Now, the amount of charge that you store there, when the technology scales to low geometries like 28 nanometer, 22, and I’m not even talking about single digit nanometer technologies that are available today. The the dimensions of this transistor and the dimension of this so-called floating gate are so tiny and the amount of charge that is stored there is just moving physically to almost single electrons.

A very small amount of charge, and this is just a physical limit. So, and of course, holding this charge over there without any leakage, without the charge going away by all kinds of physical mechanisms is becoming really impossible in the smallest geometries, and that’s why flash,

Okay, yeah, yeah. That’s very helpful. Actually, I wanted to compare this to what like for example, DRAM looks like. A modern DRAM actually uses the equivalent of let’s say a 10 nanometer transistor. But the the charge that is stored is never stored within the the transistor itself because there is a capacitor on top of the DRAM that’s actually stores the charge. And the whole problem of DRAM and capacity and DRAM density is how to get as much capacity by digging a deep trench in the in the silicon and connecting a transistor to it and how to maximize that capacitance without having to rely on the transistor to hold that charge. But what you’re saying is like in the in the flash world, the transistor has to hold the charge. And then you make the transistor so small that you can hold basically a few electrons of charge in it and that’s about it because it’s so small. It physically stops scaling. You you need to hold more charge than that for which you need larger transistors, period. And that’s why flash does not scale beyond 22 nanometers. So, there’s no flash memory in FinFET transistors or below.

Ilan: Yeah, this is a very good point because you mentioned all kinds of vertical capacitors and trench into the silicon that are done in in discrete DRAM chips. When we are talking about the embedded world, it’s very important to note that the whole idea of embedding a system on chip with memories and with analog is that you would like to keep a plain CMOS technology that allows to integrate on the same chip with the same basic technology features. You would like to have analog components, you would like to have digital components, you would like to have SRAM because there will always be SRAM and you would like to have an NVM, but you will never invest a special trench capacitor or you will never invest a lot of special technology because it will make the whole system on chip too expensive. So another key parameter that we look when selecting a non-volatile technology, I mentioned before endurance and the performance of the memory, is how much does it interfere, how much does it add on top of the baseline CMOS technology? In the end of the day, semiconductors is very cost sensitive. So it doesn’t seems like that, but just imagine how much is the selling price for such an embedded MCU. If one would need to add too many processing layers, which eventually translate to money and to time of manufacturing the silicon wafer, that will make the whole system on chip too expensive. Imagine there is a large system on chip that has everything I mentioned, digital analog, and it has one NVM component, one NVM code storage. If you add more layers of processing, special structures, you pay this additional cost over the whole area of the chip, which is making the whole system on chip more expensive. So a key parameter in adding any technology is how many masks. We count basically in in semiconductor processing, we count masks. So how many masks does the technology add on top of the plain CMOS, plain vanilla technology? And that’s another key element because in the end of the day, this means how much are you paying premium for having a system on chip with embedded NVM. So when you mentioned, for example, in in NAND flash, that’s a totally different technology with 3D integration and many levels. And when you mentioned, for example, discrete DRAM with trench capacitor, these are expensive technologies, but they’re optimized for a single use case, which is an NAND flash chip or a DRAM chip. In embedded, you need to have a plain vanilla technology with the minimum additional layers to support all the type of memories that you need.

Yes, okay, it’s very important. Because you’re saying that, look, the the cost of these microcontroller chips that require these non-volatile memories is already very low. Now, if you want to add like a whole lot of masks into the semiconductor process just because you have to implement this non-volatile technology, it better be very inexpensive because otherwise, you can’t have the the whole bill of materials for the chip go up just because you want to add an NVM materials inside. So, non-volatile memories should come with as few masks as possible. So, with that being said, how many masks does a typical NVM embedded flash process add?

Ilan: Yes, so important to to mention. So, embedded flash adds another, I would say, polysilicon layer, a floating gate layer on top of the transistor. In some cases, there is split gate technologies. There are all kind of flavors of that. But they add a lot of processing and this processing is done in what we call the front end of line. So, in the transistor, very close to the silicon substrate where you actually implement your transistor. And for making that, we’re talking about in the range of 10 additional masks, maybe eight to 10 additional masks just to process these additional features required for embedded flash.

Okay, that’s what percentage of the total mask count would you think that is? A typical CMOS process has, would you say, like 70, 80 masks?

Ilan: Uh, even less. I mean, a typical CMOS will have even less masks. And so yeah, this can add 25, 30% add-on and it can become very significant in the cost.

That’s a problem. That’s that’s what we’re actually trying to solve here. And when you’re trying to go to smaller transistor nodes, anyway, you can’t store enough charge and you can’t add so many layers if you have to add 30% more layers, that’s not a cost optimal feature. So, I think this is a good time to get into, we’ve set the stage exactly as to what is non-volatile memory, why is it used, where is it used, what are its features, and why what is required from a non-volatile memory in terms of cost, simplicity of implementation, capacity and everything. So, I think we have set the foundation of non-volatile memory and where it lies. I think the next interesting step is to go into what Weebit Nano actually does, which is resistive RAM. Like, this is a nice time to get into what that is and how it solves many of these problems that we just talked about.

Introducing Resistive RAM (ReRAM)

Ilan: So, indeed, resistive RAM is a technology that is aiming at replacing embedded flash. It has several key benefits related to our previous discussion. First of all, it is very low cost. It is adding just two additional masks to the technology. It is embedded only in the what we call the back end of line. So, we actually build the structures between the metal layers above way above the silicon substrate and transistors. Now, why is this important? We mentioned before all kind of mixed signal chips, analog chips that require some non-volatile memory. These technologies have passed through a lot of optimizations of their unique analog characteristics. For example, some technologies have BCD transistors, which are high voltage transistors. And when you invested so much in an analog technology to have the the perfect analog property,

One second, I just wanted to that is a very, that’s a very interesting point that you mentioned here. That’s because I think people need to understand where resistive RAM sits inside the stack of a silicon chip. What you were saying is that back end of line means that it is sitting like very way on top of the chip, where if you think the transistors are way below, the substrate is way below. This is all the way on the top. So it’s actually processed almost very much at the end of the chip processing. And so it’s very decoupled from what is happening in the transistor world downstairs. And so you also mentioned BCD technology. So for those who have not heard of this, it stands for BCMOS, CMOS and DMOS. And this is a specific technology that is good for high voltage transistors. It’s used a lot in automotive applications, because it can withstand a lot of voltage, like tens of volts. So these transistors are unique to what they do. Like they are used in like power conversion circuits and things like that. You you don’t want to go and put a a storage and non-volatile memory storage technology that messes with the transistor structure. So that would make the whole BCD transistor, which is so well optimized for power performance or like whatever that is supposed to do at high voltage performance, it’s going to mess up because the non-volatile memory is like interfering with that. So the best place to put it is like really far away and process it really after all the silicon is done at the back end of line right on top of the chip, right? So that is what resistive RAM does with just one or two added masks, which makes it very inexpensive.

Ilan: That’s so, that’s a right on. So, what is resistive RAM? So basically, as opposed to flash that we said is storing charge, resistive RAM, and the type of resistive RAM that Weebit Nano is doing, which is called OxRAM. So oxygen vacancy based resistive RAM, is actually implementing a resistive element, which you can change between two states, a low resistive state and a high resistive state by applying a voltage. And so I will try to explain it with this demo that we have here. Let’s see if it works.

Yeah, if you’re listening to this, this is a good time to switch to YouTube because Ilan has like a a nice 3D model of a the internal structure of a resistive RAM cell and I’m just describing what I’m looking at here, which changes between its low resistance and high resistance states.

Ilan: Resistive RAM is actually a resistor that can switch between two states, a high resistive state and a low resistive state. What we do, we take an insulating layer, like a capacitor that has a very high resistance. And by applying electrical voltage and current, we actually create these oxygen vacancies, that are, that you can see here, that are actually forming this filament that is a conductive filament. So with oxygen vacancies, charge can go through and this is becoming conductive. Conductive mean lower resistance. So when the memory has this filament built, the resistance is low and we are storing the value of one. Then by applying voltage in the opposite direction, what we can do is actually dissolve this filament. So in this side, in the high resistive state, we dissolve the filament, it goes back to an insulating layer that has a high resistance. So, current can no longer go through or the resistance is high. So we have these two states, either the filament is created and we are in the low resistive state, or the filament is dissolved and we are in the high resistive state. So these two states, we then read them with some, sense amplifier, which is a typical circuit that is used to read any type of memory, whether it’s flash, DRAM, SRAM. So we can actually sense whether the filament is there and then we read a one, or the filament is dissolved and we read a zero. So that’s basically the way resistive RAM is programmed by applying voltage that will create the filament or voltage in the opposite direction that will dissolve it. And it is read by reading or sensing the resistance, whether it’s a high resistance or a low resistance.

Why ReRAM, Why Now?

Okay, this is fascinating. Usually these are not the kind of devices that we see inside a typical CMOS technology process. So what I, what I find fascinating about this is that clearly you can’t use any kind of a dielectric material that creates oxygen vacancies, right? There’s like, there should be something that is different. Now, I know Weebit Nano would probably have its proprietary materials, but would you be able to broadly explain what, what is the property that causes these oxygen vacancies? What kind of materials do this and how is it compatible with the CMOS process flow?

Ilan: Yeah, so we’ll not dive into the whole physics of resistive RAM, but basically, I mentioned that we have an insulating layer, but on top or bottom of it, and there are two electrodes. So there is a bottom electrode that is built from some kind of inert material, and there is a top electrode that is built from what we call a scavenging, layer. So the scavenging layer is the layer that can actually, kind of store oxygen or give it back. So according to the, electrical field, this scavenging layer either will take away oxygen from, the dielectric material or will, give back oxygen, to it. So by applying the electrical fields in opposite directions, we can either have, oxygen vacancies inserted into the material or oxygen going away and creating vacancies or the opposite, contribute back, oxygen, to, dissolve the vacancies.

Awesome. Okay. That’s, that’s unique. Like, so this, this material is actually, compatible with CMOS foundries?

Ilan: So, maybe the most important and overwhelming property of resistive RAM is that to generate it, we at Weebit Nano use only fab friendly material. So we use materials that are available in any CMOS technology. We don’t use any rare earth material. We don’t use any expensive and we don’t use any contaminating materials. So some technology that maybe we’ll not discuss in this podcast, like magnetic RAM,

Yeah, that’s resistive RAM itself is like not an idea that is like new, right? You mentioned Webit Nano has been doing this for 10 years. But why now? What is what has changed now about resistive RAM that we should, people have been trying this since Panasonic have have tried it, Samsung has tried it. What’s what’s the need now? What what has changed?

Ilan: So, I think it’s a it’s a combination of many elements. One is what we mentioned before, that, you know, flash is hitting the physical limit. And cost is, we know that the cost pressure on semiconductor is always growing. There is the Moore’s Law, and there is a lot of pressure to reduce cost. Moreover, I think one of the driving application, driving is a good word here, is automotive. You mentioned it before. So, we at Webit, were able to demonstrate that this technology is not only low cost and has a very good performance, but it can actually serve this very challenging market of automotive. So, we all know that in our cars today, everything is electronic. There are, you know, hundreds of chips, and many of them require non-volatile. Today, resistive RAM technology can serve the automotive market, and I think it’s a major breakthrough.

So, that’s nice. You mentioned that the competitors to this technology are MRAM, magneto resistive RAM, and ferroelectric RAM, both of which are alternative materials and not really well suited to be integrated into a a CMOS fab. So, those are those are like technologies are ruled out, but then there’s also I think another alternative technology called like phase change memory, which, you know, alters its physical crystal structure with voltages. So, like this is a resistive RAM is an other application where automotive is driving a lot of requirement here, especially as we go to smaller technology nodes, where flash stops working, resistive RAM is really well suited to automotive. And so, has this technology been qualified already for automotive? Because I know that there’s this standard called the AEC-Q100, which is like a very hard automotive standard to pass. Does Webit Nano already pass this standard?

Ilan: Yes. So, indeed, our technology is now qualified for ACQ100 automotive. And we are continuing this with additional customers. So, we recently announced two I would say leading IDMs that are both active in the automotive space that license our technologies for automotive platforms.

More Than Just a Bit Cell

So, the way that this resistive RAM technology works from Webit Nano is that you guys have the secret recipe that explains how resistive RAM can be made in a fab, and a fab will license it from you. Is that how this whole thing works?

Ilan: Yes, but that’s where when you asked me why resistive RAM is now becoming available. It’s because it’s not all about the secret sauce. So, the secret sauce, for sure, it was developed throughout decades, and we have now all the recipe for the the exact dielectric structure and the exact electrodes that are needed to to, you know, do the scavenging of the oxygen and all of that. And we really have, you know, tens of PhDs in Webit working on this part along with our partners, as I mentioned. But that’s not enough. If you are an embedded, I would say, MCU, let’s say, a system on chip architect. What you need is a full-blown memory module that you can just commission a a read command, a write command, that’s it. All the rest is done. You don’t work with the bit cell, you don’t work with physics. You you need to write a 32 or 64-bit word onto the memory and be able to read it back when you need it. And that’s where Webit Nano is today the only fully vertically integrated and fully independent IP provider. Because we at Webit, we have all the pillars that are needed to build such a technology. So, yes, you need the bit cell that I showed and the physics around it and the secret sauce. But around it, you need to build the design and the layout of the full memory array. You need to design the the and layout all the peripheral circuits, like the the decoders and the drivers and the sense amplifiers that I mentioned before, which are basically of analog nature, that are able to sense these exact resistances. And for programming the resistive RAM, we need to apply, as I mentioned, electrical field, which is actually voltages and currents that needs to be controlled very precisely. So, for that, you need a whole analog design and layout team. And then on top of it, the programming sequence of resistive RAM is done using a proprietary smart algorithm. So, not only that we did an analog capabilities, we also need a very strong algorithm team, we need a very strong digital design and verification team. So, basically, what I’m telling you is that the resistive RAM module that our customer actually embeds into their SOC is comprised not only of the physical bit cell, it has a lot of analog design, memory architecture, layout, digital design, algorithms, and and even software is developed at Webit in order to have a full offering of embedded resistive RAM IP.

Yeah, it sounds easy, but actually what is unique at Webit Nano is that all these teams are sitting together. So when there is a discussion of, let’s say, a performance that needs to be met or automotive, let’s say, meeting automotive qualification, you mentioned before, it requires meeting very high temperature profiles and high endurance. And in order to meet this, you cannot do it by just the bit cell. You cannot do it by just some analog design. You need a fully multi-disciplinar approach that takes the best of the device physics and the process, the best of the analog design, the best of the digital and the algorithm, and wraps everything together into this product. And and that’s the main strength of Webit Nano and that’s the main the probably the only way that you can have any NVM technology be commercially viable. It was done before in flash memory. It is mostly done in other technologies. But for resistive RAM today, we are doing it and all these capabilities exist inside the team at Webit.

Performance and Future Applications

So you mentioned endurance was one of the factors that you have to really design in as well. It’s not as easy as it seems. What in the rough ranges is the endurance of resistive RAM? Like what and how does it compare to like, let’s say, flash memory or any other alternatives if you have the numbers off the top of your head? I’m trying to place like, is it more endurance, more endurance or less endurance or where does it fit?

Ilan: I would say, typically, a flash is 10K cycles. It depends, of course, there are some technologies that are more, but I would say 10K cycles typically. And for our resistive RAM, we have some products that are 10K. We already demonstrated and qualified 100K cycles, which is opening the market for additional applications, automotive and others. And we are working on extending it beyond the the 100K into the million cycles in the near future.

Okay, awesome. So it’s going to be quite performant in spite of the treacherous conditions that automotive applications want from, I don’t know, from driving cars in Alaska to the Sahara Desert. You have to support all temperatures and so it’s usually a very challenging qualification to pass. And especially the life’s the lifetime of memory used in automotive is also should be pretty high. So all of these things should fit broadly into that umbrella. Where do you see resistive RAM going next? So what’s next? So, you know, we already spoke about microcontrollers. What’s the next big thing?

Ilan: Yes, that that’s a very good question, Vik. So we all know that the next big thing or the existing big thing is AI. So, resistive RAM, or I would say in the academic, naming is called memristors. That’s kind of how the academy wants to to name it. Had been, studied in, multiple academic researches, for what we call in-memory compute or IMC. Webit is also very active in this field, probably a good topic that we can discuss in one of the the next podcasts. But basically the idea is that resistive RAM, a a different type of resistive RAM array can actually create AI calculation inside the memory by just combining resistive elements. And this had been demonstrated in the academy and commercializing it, I would say, is one of the next big things for resistive RAM.

An Innovation Story

Amazing. That’s that’s would be very interesting to hear. Analog in-memory compute has been around as a concept for a long time. And as we have like the AI era coming up and GPUs sucking like 1,000 to 2,000 watts per GPU, maybe the future lies somewhere where we can do it differently. But that is a big topic and a topic for another day. All right, Ilan, before we wrap up this, I want to hear like, what is the one story of innovation that you faced or came across while developing this technology within Webit Nano? That’s a good story to tell for the whole audience.

Ilan: Yeah, I can think about such a story. We mentioned analog trimming before, right? So, embedded system on chip mix and chips will always have some analog, but I also mentioned that we have our own piece of analog within our memory. So obviously we need also to trim that part. So we came up with the idea, hey, we have non-volatile memory and we need to do trimming. So let’s just take some part of the non-volatile memory, just some additional row, store our own, data, our own trimming values there, read them and we can just use it to trim our own system. And then we when we went to to our lead customers and we explained to them and they asked us, how do you do this? And we said, yeah, we have our own mechanism that we use our own non-volatile memory to trim our system. They say, why? We want to use that as well. So can you grant us access? And we would like to trim our own system and it it just saves us having additional NVM, for trimming our system. And today it became a standard feature in all our modules in which we let the end customer get his own, user data, his own trimming data, right out of the NVM, and this is done automatically and became a lead feature.

That’s amazing. So the key takeaway I’m getting from all of this is that it is important that you solve problems. It doesn’t have to be the customer’s problems. Even if you solve your own problems that you come up sometimes, you just have like a new product on your hands or a new problem you solved for someone else. And you can turn around and sell your own solution to your problem to solve someone else’s problem and it’s like a perfectly viable business model.

Ilan: I guess that’s what you call engineering, right?

Yeah, that’s that’s amazing. It’s a nice story. Awesome. Thanks Ilan for coming on Semi Doped. It’s glad to have you here. I know we don’t often talk about things like resistive RAM, but that’s what Semi Doped is for that we sometimes try to go outside of the normal narrative of AI, semiconductors, memory, optics, you know, we talk about all this stuff, but every once in a while we like to throw a curveball and have someone like you explain what resistive RAM is to us. So thank you for coming on the podcast.

Ilan: Thank you for having me, Vik.

All right, so that’s the end of the podcast. So please check us out on all the podcast platforms if you haven’t already. Apple really likes it if you give us five stars, they push us up the technology rankings and it’s always a competition as who gets to the top. So if you can go press the like five star review thing that really helps us out. And as usual, check us out on our daily newsletter, which is daily.semidoped.com and my newsletter is Vic’s newsletter.com. And Austin, who was not on this episode, but is usually on the podcast with me, also has chipstrat.com. All right, see you on the next one.

Nvidia authorized a $150 billion share repurchase program, the largest in the company’s history, signaling serious confidence from the board. TSMC is also pushing hard on 2nm, targeting 120,000 wafers per month by end of 2026, while Samsung Electro-Mechanics commits $4.4 billion to FC-BGA substrate capacity for AI accelerators.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Nvidia Authorizes $150 Billion Share Repurchase, Largest in Company History

Nvidia authorized a $150 billion increase to its share repurchase program, the largest such authorization in the company’s history. The board’s approval adds to existing buyback capacity, giving management broad latitude to return capital to shareholders as cash generation from AI-related products has expanded sharply. Reuters reported the authorization as a record for Nvidia, framing it against the backdrop of sustained demand growth across the AI sector. No timeline for the repurchases was specified, and buyback programs of this kind allow companies to execute purchases at their discretion based on market conditions, capital needs, and regulatory considerations. The scale places Nvidia among a short list of technology companies to authorize repurchases at this magnitude in a single action.

Austin: Think about what Nvidia CAN’T do with this money. Big acquisitions are dead on arrival with regulators (see the Groq probe…) Ecosystem investments get called circular. They’re fabless… and not a cloud… so when every other door is closed, the $150B goes back to shareholders. Which is honestly a good thing.

Samsung Electro-Mechanics Commits $4.4 Billion to FC-BGA Substrate Capacity

Samsung Electro-Mechanics will invest KRW 6.78 trillion ($4.4 billion) across its Sejong, South Korea facility and a Vietnam plant to expand FC-BGA substrate production for AI accelerators. The Sejong expansion alone accounts for KRW 4.27 trillion ($3.14 billion), making it the company’s largest single domestic capital commitment on record.

FC-BGA substrates, which connect processors to printed circuit boards in high-performance computing packages, have become a chokepoint in AI server builds as demand from hyperscalers outpaces available supply. Semiconductor Engineering notes that substrate capability is increasingly application-specific, with larger body sizes, higher layer counts, and finer routing narrowing the field of qualified producers. The Vietnam allocation covers additional offshore capacity, though the precise split between the two sites hasn’t been disclosed.

Vik: Everyone’s fixated on the capacity number, but mass production doesn’t start until September 2028 - it does nothing for today’s squeeze.

Austin: That’s like the story of the entire supply chain… so much is coming online in 2028

TSMC Scales 2nm Output, Adds Five Packaging Plants Across Taiwan

TSMC plans to reach 120,000 wafers per month at its 2nm node by the end of 2026, a volume that would make N2 one of the fastest leading-edge ramps the company has executed. Five additional advanced packaging facilities are planned for Chiayi, with a separate hub taking shape in Taichung, extending a build-out driven by AI hardware demand. The packaging expansion matters because CoWoS and SoIC capacity has repeatedly constrained AI chip shipments more than wafer supply itself. Affiliate Vanguard’s joint $7.8 billion Singapore fab with NXP, announced through Singapore’s Economic Development Board, adds mature-node capacity outside Taiwan, rounding out a supply chain that now stretches from leading-edge logic through final integration.

Vik: The N2 ramp is real - tape-outs already run 4x the N3 node, so demand isn’t slowing. But more wafers hit the same wall downstream. This is all round-wafer CoWoS on the packaging side, and I don’t buy that it scales to the very large packages coming. Glass panels are the endgame, but thats a longer term story.

SK Hynix Weighs $150 Billion Solidigm IPO as HBM5 Clears TSMC Validation

SK Hynix is considering a US IPO for its Solidigm NAND unit at a valuation of up to $150 billion, according to sources cited by Reuters, a figure that sent SK Hynix’s own shares lower as investors weighed the dilutive optics of separating the business. The NAND spin-out news arrived alongside a sharply different signal from Santa Clara, where SK Hynix presented at TSMC’s Open Innovation Platform conference on October 23 and validated its HBM5 architecture in collaboration with TSMC, picking up the Partner of the Year Award for the second consecutive year. Two franchises, two very different moods on the same week.

Vik: The HBM5 validation on CoWoS is the part I’d focus on this week. SK Hynix flags that from here power and design-process co-optimization matter most - that’s the base die story. And the value there keeps sliding towards the logic designers, while the memory stack drifts back towards commodity.

Qualcomm Pushes Snapdragon Into PCs, Wearables, and Samsung Foundry

Qualcomm used Snapdragon Summit 2026 to extend its platform well past smartphones, announcing a redesigned PC architecture built with Google for agentic AI and adding 1-bit AI model support to its latest wearable Snapdragon platforms so on-device inference can run on ultra-compact hardware. The next-generation Snapdragon X takes a major performance step and opens support for Linux. Qualcomm also confirmed it is actively evaluating Samsung’s 2nm process node for future Snapdragon chips, a move that would diversify its foundry base beyond TSMC. On the connectivity side, Qualcomm says its 6G commercial roadmap is underway, with a 2029 launch target. The new Spectra ISP supports three 64MP sensors simultaneously.

Austin: Stay tuned for our podcast dropping Thurs Oct 1 for our thoughts about AI at the edge on Snapdragon.

Infineon Buys AI Power Specialist C2i, Samples Record 120-Amp Voltage Regulators

Infineon Technologies acquired C2i Semiconductors, a data-center power-management specialist, in a move to sharpen its position in the fast-expanding market for AI accelerator power delivery. Terms were not disclosed. The company is simultaneously sampling a new voltage regulator rated at 120 amps per phase, a figure it claims sets a current record for the category and targets the dense power requirements of AI accelerator racks. C2i’s expertise in digital control and telemetry for server power stages complements the new silicon directly. Infineon’s automotive power business, which still accounts for a large share of group revenue, is running into softer near-term demand, giving the data-center push added urgency inside the company’s product mix.

Vik: Infineon buying its way deeper into AI power. C2i’s software-defined multiphase control and smart power stages feed straight into the vertical power delivery and SIVR path they’re chasing. The question I keep coming back to is whether any of this actually takes sockets from MPS.

Sector Watch

Foundry & EDA

• Synopsys launches an autonomous engineering agent platform and expands its TSMC collaboration for next-generation AI system design at the TSMC OIP conference. (Synopsys)

• ASML confirms 2027 EUV output is fully sold out, High-NA demand is firming, and the company is opening an India office as its European revenue share sits at zero. (ASML)

• DB HiTek raises foundry prices by up to 30% citing surging demand from Chinese customers, signaling tight capacity at mature nodes. (thelec.net)

Packaging

• HANMI Semiconductor ships its first FC Bonder 75 to a customer, the company’s new flip-chip bonder designed for 2.5D AI semiconductor packaging. (Semiconductor Digest)

• Simmtech advances its SOCAMM substrate line as Taesung enters talks to supply copper plating equipment for the build-out. (thelec.net)

• China accelerates glass substrate commercialization, pulling ahead of Korean rivals in process validation and pilot production timelines. (The Elec)

Power & SiC

• Wolfspeed launches a premium 200 mm silicon carbide substrate targeting power-device makers seeking higher-performance SiC wafers for EV and industrial applications. (Business Wire)

• Schneider Electric launches a Water-Cooled Distribution Unit combining hybrid liquid and air cooling for high-density AI data centers. (ARC Advisory Group)

• Eaton acquires Italian grid equipment maker COL Group for €810 million, expanding power-distribution manufacturing capacity across Europe. (Reuters)

Memory

• Montage Technology enters mass production of its fifth-generation DDR5 RCD chip running at 8,000 MT/s, a first for a Chinese supplier. (technode.com)

• Saimemory partners with Shinko Electric to co-develop high-bandwidth, low-power memory targeting next-generation AI workloads. (Telecompaper)

• CXMT puts its fifth-generation DRAM into production, marking a new process node for China’s leading domestic DRAM maker. (CXMT)

Compute & Edge

• Supermicro begins shipping Nvidia Vera Rubin NVL72 racks with 1.8 MW in-row CDUs and a 1,152-GPU scalable unit configuration. (storagereview.com)

• FuriosaAI faces a funding crisis after South Korea’s Public Growth Fund delays a 400 billion won investment, drawing a fairness backlash. (Chosunbiz)

• Fluidstack raises $1.5 billion led by Jane Street, quadrupling its valuation to $18 billion in 60 days on AI compute demand. (Dealroom.co)

Optics & Networking

• GlobalFoundries reports strong demand for Chinese optical module chips amid the data-center construction boom, expanding into the segment. (SCMP)

• CtrlS deploys Ciena optical technology across five Indian cities to build managed fiber networks for hyperscaler customers. (BusinessLine)

• MaxLinear launches Puma 9, enabling operators to accelerate AI-ready DOCSIS 4.0 multigigabit services while cutting deployment costs. (MaxLinear)

Data Centers

• Nscale closes $3.36 billion pre-IPO convertible financing led by Third Point, with $1 billion from Nvidia still pending, targeting a $35 billion New York IPO. (Nscale)

• Crusoe cancels its $1.25 billion plan to deploy 29 Boom supersonic turbines for AI data centers, citing a strategic pivot on power sourcing. (TechRepublic)

• Submer Group launches Corenix, a modular immersion-cooled data center platform targeting accelerated-computing deployments. (PR Newswire)

Policy & Trade

• Singapore launches SG Semiconductor, a national identity program to unify and promote the city-state’s chip industry on the world stage. (Semiconductor Digest)

• China weighs allowing ByteDance and Alibaba to purchase new Nvidia chips as Huawei Ascend falls short of meeting domestic AI demand. (reuters.com)

• US and China release product lists for a $30 billion bilateral tariff deal, the first concrete sectoral rollback of the trade dispute. (Bloomberg)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Stalking the elusive computer bug? Seek no more. Welcome to the weekend’s entomological footnote.

On 9 September 1947, at a quarter to four, someone found a moth stuck in Relay #70 of Harvard’s Mark II computer. They taped it into the log book and wrote, “First actual case of bug being found.” Do notice the word “actual”. You only write that if everyone already calls a fault a bug. At Harvard, IBM’s engineers had taught the staff the word in 1944.

The log book, the Smithsonian tells us, was probably not Grace Hopper’s. She told the story often, and stories have a way of settling on whoever tells them most. She and the Mark II team did help spread “bug” and “debug” among programmers.

The word comes from the telegraph. Edison’s quadruplex had faults called bugs, and his workarounds were called bug traps. In 1878 he wrote to Western Union’s president about a bug in his telephone apparatus, of the genus “callbellum”. Nobody has found a specimen.

In 1889 the Pall Mall Gazette explained Edison’s bug as an imaginary insect hiding inside his phonograph, and it took fifty-eight more years for one brave moth to go where no moth had one before.

Speaking of bugs, we’re doing some debugging of our own.

AI moves fast, but it still comes down to silicon, power and architectural limits. We can’t understand what comes next without knowing whose shoulders we stand on: the pioneers, the engineering gambles and the people behind the chips. That’s why we run a TIL on Saturdays and a quiz on Sundays.

We want Semi Doped to earn its place in your inbox, Sundays included. So we’re pausing the quiz today to ask: what are we not covering that you want to read? Would you swap the quiz or the TIL for something else? If so, what?

Leave your ideas in the comments. We can’t promise to write about every one, but we’ll read every word and use it to shape what comes next.

And so we know who we’re writing for, a quick poll on your day job:

• Engineers & Architects: I design/build silicon, hardware systems, or low-level software.

• Execs, Founders & Investors: I track strategy, capital allocation, and market dynamics in tech.

• Adjacent Tech Pros & Product: I work in tech/AI/software and want to understand the physical stack.

• Students, Researchers & Lifelong Learners: I’m here out of pure curiosity and a love for deep tech history.

Leave a comment

You’re probably traumatized by being on hold to this track.

The track is Opus No. 1, and the credits name Tim Carleton and some even name his high-school friend, Darrick Deel.

We traced the cruel and unusual punishment of call wait music back to a girder.

In 1962, the story goes, a factory owner named Alfred Levy found a problem with his phone lines. An exposed wire touching a metal girder had turned the building into a receiver, and callers on hold were hearing the radio station next door.

If that’s true, did hold music start as a wiring fault?

The oldest version of the story we found is a 2005 web page from a Tampa company that sold on-hold systems. Smithsonian magazine calls it “industry folklore.” None of them names the factory or the station.

So we went in search of the paper trail. On March 26, 1962, Levy filed for a patent on a “telephone hold program system.”

His problem was silence, because a caller on hold “hears no sound,” the patent says, and for all he knows he is “connected to a dead line.” The operator may be too busy to check in, leaving him “bereft of even this small consolation.” He hangs up or redials. The result: annoyance and a waste of time and money.

Levy’s fix was to play him music, “thereby to pacify the originator of the call.”

Any suitable source would do — a phonograph, a wire recorder, a magnetic drum recorder, a live program, the radio. It didn’t even need a direct connection: the patent allows coupling “which avoids the necessity of making a direct metallic contact.”

So no girder required.

In the late 1990s, Deel asked Carleton if Cisco could use the song as the default hold music for its first VoIP phone system. Carleton said yes.

Sixty-three years after the girder, if there even was one, Apple gave Opus No. 1 a cameo at WWDC 2025 while presenting Hold Assist, the iOS 26 feature that sits on hold so you don’t have to.

In 1983, he and Jon D. Paul filed for a system that let callers pick their own program. The reason: a held caller “may hang up, because he grows bored or because he dislikes the program being played.”

This from the man who patented hold music.

Long story short, we couldn’t quite find more on the true origin story of this one. Have you heard another version of this? If so, let us know in the comments.

Leave a comment

Oracle’s force majeure notice on its New Mexico data center campus is rattling power equipment suppliers and raising questions about big infrastructure commitments, even as Goldman Sachs projects the five largest US hyperscalers will push AI CapEx past $1.2 trillion by 2027. Akamai is moving fast to capture that spend, signing an $11.6 billion cloud infrastructure deal with Anthropic, and Google sent a fridge-size TPU satellite into orbit aboard SpaceX Transporter-18.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Oracle Force Majeure Notice on New Mexico Campus Rattles Power Equipment Suppliers

Oracle sent a force majeure notice to Blue Owl Capital over the Project Jupiter data center campus being built in New Mexico, citing a series of setbacks that have stalled construction on the Stargate-affiliated site. The move lets Oracle seek cover from contractual obligations tied to the development timeline. Bloom Energy and GE Vernova both saw their shares fall on the news, as investors repriced the revenue visibility each company had built around steady hyperscale construction schedules. Barron’s flagged GE Vernova specifically as exposed, given its power equipment backlog depends on data center builds proceeding on time. Force majeure claims of this kind can suspend or void supplier delivery commitments, leaving infrastructure vendors holding contracted capacity with no clear installation date.

Vik: made this really keen observation on the Semi Doped podcast, a week before this happened. He was totally right. It’s the local governments who will set the pace of AI. Not Silicon Valley.

Akamai Commits $11.6 Billion to Anthropic, Crossing $500 Billion Compute Threshold

Akamai’s multi-year agreement with Anthropic covers $11.6 billion in cloud infrastructure, with Akamai separately committing $5.5 billion in capital expenditure to support the build-out. The deal isn’t structured as a standard cloud contract: forkast.news describes it as an equity position in the model layer, recasting Akamai from a CDN operator into a direct stakeholder in frontier AI. Anthropic’s cumulative compute spending crossed $500 billion in under a year, with Akamai’s commitment pushing it past that threshold. That pace, for a company still privately held, is striking on its own terms. Reuters reports Anthropic is separately seeking 50.1% voting control for co-founders ahead of a planned IPO.

Vik: This is for their CPUs. Conspiracy theory: Anthropic will release a personal agent like Muse.

Goldman Sees Hyperscaler AI CapEx Hit $1.2 Trillion as Nscale Closes Pre-IPO Round

Goldman Sachs strategists project spending on AI infrastructure by the five largest US hyperscalers will rise more than 50% in 2027 to $1.2 trillion, sustaining semiconductor and power-equipment demand well beyond the current buildout cycle. Against that backdrop, European AI cloud provider Nscale closed a $3.36 billion pre-IPO convertible note round led by Third Point, giving the company fresh capital to expand its GPU-dense data center platform ahead of a public listing. The convertible structure lets Nscale defer equity dilution while locking in growth funding now, at a moment when hyperscaler commitments give lenders unusual confidence in forward AI compute demand. Reuters Breakingviews called the implied valuation a stretch, but the capital is closed.

Google Launches Fridge-Size TPU Satellite Aboard SpaceX Transporter-18

Google’s Project Suncatcher sends its first hardware into orbit next week aboard SpaceX Transporter-18, carrying a fridge-size unit it calls a minimum viable product. The satellite packs four TPUs and 1,000 watts of solar power, though overheating constrains each compute run to just 15 minutes despite the spacecraft being rated for a full year of orbital operations. That thermal ceiling is itself a core finding the mission is designed to produce: Google needs real vacuum-and-radiation data before it can engineer around the problem. The payload rides a SpaceX rideshare rather than a dedicated launch, keeping early validation costs in check. No other major hyperscaler has flown an AI compute payload of this kind, leaving Google alone in collecting radiation-hardened TPU performance data from low Earth orbit.

Vik: Its cute they sent up a 4-TPU scale-up domain to test in space. I always maintain that launch costs are the biggest impediment to datacenters in space. Watch for drop in launch costs as a leading indicator for space compute.

ASML Sold Nothing to European Fabs as Dutch PM Eyes China Deal

ASML’s European revenue share has dropped to zero, with the company saying it has sold “absolutely nothing” in the region this year. That stark absence of domestic orders arrives as Dutch Prime Minister Rob Jetten confirmed he has discussed the MATCH Act with President Trump, legislation that would further restrict ASML’s China sales. Jetten told Bloomberg Television he is “pretty sure we come to some agreement that is workable for each and every one,” offering cautious optimism about the negotiations. Europe produces none of ASML’s revenue while its own government works to tighten the screws on the one market still buying heavily.

Vik: That’s because there are no chips happening in Europe. Absolutely nothing too surprising.

Muse Gives Every User a Free Isolated Linux VM in the Cloud

Meta’s Muse AI assistant now provisions each user with a dedicated Secure VM running a full Ubuntu Linux image, giving users root-level visibility into the runtime cell filesystem. A Sentinel process running outside the runtime cell oversees sensitive actions, while passwords and other secrets are stored separately from the cell. Users can install software, compile code, browse the web, and download their personal agent data via Settings > Data controls > Download your agent data.

Sector Watch

Compute & Edge

• Microchip Technology closes acquisition of edge-AI chip firm Hailo, accelerating consolidation in the edge inference silicon market. (Jon Peddie Research)

• Alibaba Zhenwu V900 AI accelerator with 216 GB memory joins Huawei Ascend 960 in competing for Nvidia’s vacated China AI compute share. (digitimes.com)

• Mobilint delays mass production of its second-generation Regulus NPU by six months, pushing the schedule back as competition in the Korean edge-AI chip segment intensifies. (The Elec)

Foundry & Packaging

• Korea Herald maps the post-CoWoS advanced-packaging race, with multiple Korean firms targeting next-generation AI chiplet integration platforms. (The Korea Herald)

• SkyWater Technology reaffirms its role as an open-access quantum foundry following IonQ’s acquisition, pledging continued third-party customer access. (EE Times)

• Tata Electronics boardroom battle prompts India government to seek assurances on the stability of the company’s $14 billion chip investment plans. (bloomberg.com)

Power

• Eaton signs agreement to acquire COL Group, expanding power-distribution manufacturing capacity for data-center and utility markets across EMEA. (Business Wire)

• ROHM SiC MOSFETs are selected for BMW’s Neue Klasse electric powertrain, marking a design win in a flagship EV platform. (Charged EVs)

• Amazon signs a deal with Generac to secure backup power supply, putting standby generation capacity into the AI infrastructure spotlight. (Data Center Frontier)

Memory

• SK hynix CEO Kwak Noh-Jung publicly reaffirms commitment to sustaining HBM and next-generation memory leadership as AI demand intensifies. (Gagadget.com)

• Rambus memory-interface IP stocks attract renewed investor attention as AI-driven demand for high-bandwidth memory architectures broadens. (Barron’s)

• YMTC secures its first-round patent win against Micron in German court over NAND technology, marking a significant IP offensive milestone. (YMTC)

Optics & Networking

• InfraTech Cables selects Poland for a new European optical fiber cable manufacturing and R&D campus targeting subsea and terrestrial demand. (Business Wire)

• Vocus launches a 4,000 km ducted fiber route connecting Brisbane and Darwin as its second project under Australia’s ADIP initiative. (Light Reading)

• Iridium stockholders approve the pending acquisition by Rocket Lab with 99.6% of votes cast in favor, clearing a key deal milestone. (Light Reading)

Data Centers

• Japan’s FSA steps up scrutiny of AI data-center financing by the country’s largest banks and life insurers as leverage exposure grows. (Bloomberg Tech)

• Nscale UK supercomputer launch is delayed by power supply problems, underscoring the grid-access bottleneck facing new AI compute builds. (The Guardian)

• xAI founder Elon Musk targets doubling the Nvidia GPU count at Colossus 2 by year-end, escalating the Memphis data center buildout. (Trending Now Infrastructure)

Policy & Trade

• Texas Attorney General opens probes into data center projects over water usage, adding a new regulatory front to the state’s infrastructure scrutiny. (Bloomberg)

• Foxconn-led Taiwan consortium announces plans to build a major AI and EV industrial hub in Poland, diversifying manufacturing geography. (Taiwan News)

• India government seeks boardroom stability assurances from Tata Electronics as a leadership dispute clouds the country’s flagship chip investment. (bloomberg.com)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Things we cover:

• Qualcomm’s High Bandwidth Compute (HBC) architecture

• The upcoming Dragonfly AI 250 accelerator

• How near-memory compute differs from HBM

• The role of TCO in hyperscaler decisions

• Scaling HBC from data center to edge devices

• Fitting a trillion-parameter model on a single card

This podcast is lightly edited for clarity.

What is High Bandwidth Compute?

Austin: Hello, listeners. Austin Lyons here with Chipstrat, and with me is Durga Malladi, IEEE fellow, which is pretty cool, and EVP and GM of Tech Planning for Edge and Data Center. So, we’re here to talk high bandwidth compute. But first, before we get into that, which is a great topic, I want to know more about what your job title means. Tell me what you do at Qualcomm.

Durga: I run all the IP and the technology planning for what we need to do across all the products, ranging from data center all the way down to edge solutions including devices. So, that means the IP, if you think about it, is what should we be doing in our CPU portfolio in let’s say three, four years downstream in ‘29, in ‘30 in data center, versus what do we need to do in devices, in handsets, in new agent tech devices. And then rinse and repeat the same exercise for what do we need to do on GPU, what about NPU, what about modem, what about sensing hub, and so on. So, these are all the different IPs. That’s the planning exercise which runs the road map behind the scenes, but it is across all the business units.

Austin: Nice. That’s pretty cool. I like that you see across all the different technologies so that it’s not like CPU running not together with GPU or not together with the different products. It’s pretty cool. Okay, so let’s talk high bandwidth compute, HBC. The way I see this is obviously Qualcomm is trying to enter and is in a competitive space with inference solutions in the data center. You’ve got GPUs, there’s XPUs, and Qualcomm has the Dragonfly lineup—AI 200, 250, 300. It feels to me like HBC is Qualcomm’s big technical bet to differentiate and bring something different to market as compared to what everyone else is doing. So, of course, I want your thoughts on that, but also then tell us what is this HBC and how is it different from what others are doing?

Durga: Absolutely. So, when we re-entered the data center business last year—and I actually ran it for one year last year—one of the things that we really focused on is, you’re absolutely right. There’s a lot of players already in the market and they have some pretty incredible solutions already. So, it wasn’t one of those moments where we said, “Okay, let’s also jump into the fray,” and it’s not a me-too moment or anything of that sort, but it was, “What is it that is going to be differentiating us from the rest of the ecosystem there? And what is a really good tough problem that we really want to solve?” because that actually gives us the benefit of, “Okay, that is different, let’s take a look at Qualcomm.”

And we latched on to the fact that—and we had an R&D project that was running internally for a while on what is called near-memory computing. The principles of that are very simple, but before that, what is the problem that we are trying to solve? It turns out that if you go back in time, maybe over the last six years or so, you will see that the compute capability, if you just take any of the GPUs and say generation over generation, how much is it improving? It improves quite a lot actually. It’s at least a factor of two, sometimes a factor of three. But the memory bandwidth wasn’t actually improving at the same league. And so the natural next question to ask is, so what? What does that do? Well, if you go back to some of the basic principles of how do you do compute, you have all your data sitting somewhere and you have to fetch the data to do the compute processing. But at some point in time, what happens is that your computer is so good that you’ve tilted the scales quite a bit. So, you might be just idling, you’re waiting for the data to actually show up. So, memory bandwidth becomes a bottleneck and it’s called the memory wall. And you see that a lot more with AI inference, especially in the decode portion of it.

And so what we did is our HBC architecture is based upon what is called near-memory compute. Instead of treating memory and compute as two distinct concepts, what if we actually do a lot of the compute right next to memory, like literally right next to it? It does two things. One, the latency comes down quite a bit. And second, we can reduce the power consumption quite a lot. So, today in data centers, HBM, the bandwidth is very high, but it’s expensive and it’s a power hog. So, we actually wanted to make sure that we can solve it in a different way. And one thing led to another and as we started talking about this with a lot of the hyperscalers and a lot of other partners, they loved the idea. And as we announced it, we actually announced a couple of names in terms of some of the partners that we have. So, it’s been a great journey on HBC.

But as we briefly mentioned in our investor day in June, and today at Snapdragon Summit, in fact, Cristiano left a teaser out there. But let me just tell you what is the point of this. HBC for us is a technology, but it’s not just for data center. It actually applies to everything all the way from data center to servers and accelerator cards, in automotive, in PCs, in XR, and yes, in smartphones as well. But it actually works in a different way in some of these things, but the principle remains the same, which is you’re doing a lot of computation right next to memory.

Stacking DRAM on a Logic Die

Austin: That’s good. That’s a great explanation. And I like how you mentioned that HBC is something that can be applied across from the data center all the way to the edge, and that’s a good organizational design as you’re talking earlier about how your job is to see IP across all that. So that’s pretty cool. Okay, so what I heard from you is the problem is compute is increasing quickly, memory bandwidth is not increasing fast enough. So this all this extra compute can’t get fed quickly enough. And maybe instead of trying to just increase the bandwidth of the memory we use today, you guys stopped and said, “Well, wait a minute, why have to shuttle all this data? Why shuttle even more like wider buses or even faster speeds? Why not just bring it really close to the compute?”

Durga: That’s correct. Exactly. And keep in mind that it’s a principle that, as you start bringing down the power consumption quite a lot—just going one step further on devices—power consumption matters quite a lot in devices, especially as you think of all the agentic AI based devices that you see these days. And they’re still evolving, by the way. There’s a lot of new players who are coming to the market. In fact, quite a few of them have very little background in manufacturing devices or making devices. They come in from an AI-first mentality and they’re coming out with these devices. And the job of agentic AI devices is to be constantly on because they’re working all the time. So, that power consumption benefit that we talked about is going to be key, we believe, in unlocking the full potential of agentic AI devices.

Austin: Yeah, that makes sense. So, it’s not just the memory bandwidth problem, but the also added benefit of because you’re not shuttling this data so far horizontally, now it’s just going vertically, which we’ll get into, there’s also a big power savings.

Durga: Yes.

Austin: And of course, we know that power is the big constraint. Okay, so then, to someone who’s listening, this is the first time they’ve heard of high bandwidth compute, are you sticking memory on top of an accelerator or tell us more about the stacking and the separation? What is being sent across the wire horizontally versus vertically?

Durga: By the way, there’s a really cool animation that we showed at investor day. I wish we had that, but I’ll just use my hands as props. So, the way it works is that the memory is like, you have these individual, the DRAM, it’s called a stacked DRAM. You’re basically stacking DRAM one on top of the other. And there’s a lot of wafer-on-wafer bonding techniques that are used on that. So, this is hardcore process technology, by the way. This is in fact, it has nothing to do with compute at this point in time. But then right underneath that stacked DRAM is what is called the logic die. That’s where you put in all the compute. And depending on the use case, you might actually say, “I want some specific logic to actually reside over there.”

There might be yet another processor that is still sitting, which is away from it. But you’ve brought in a lot of the heavy computational needs, like the ones which really need that memory bandwidth. You put all of that right next to memory in that base die. And that HBC therefore is a combination of both, the stacked DRAM on top plus the logic die that is underneath or the compute die that is underneath. And now you’ve done a bulk of the work just with that alone. There might still be some residual compute that you want to do, and that could occur in a separate processor over here. In some sense, HBC is a bit like a co-processor because it’s doing a lot of the compute, but it’s not the only thing. You might still have another main processor that’s running out there, but it removes that complexity quite a bit. So, in short, it is literally right underneath that stacked DRAM that we’re talking of. That’s where all the compute is being done.

Is This Just Custom HBM?

Austin: Okay, okay, that’s helpful. So, anyone who’s familiar with HBM will say, “Well, this sounds a lot like HBM4E with a custom base die and stacked DRAM on top.” So, tell us the differences. First of all, is it a different class of DRAM that’s being stacked? And then how should we think about the difference between the accelerator at the bottom versus a custom die?

Durga: Yeah, that’s a in fact, this is exactly one of the things when we first came out with the concept and when we talked about it, I went to Korea and we talked. And you might have noticed a couple of announcements that have been made even by the memory vendors. They have different terminology for that. It is their equivalent of HBC, but actually they are partners of ours. So, we expect that these two technologies will co-evolve for a period of time, and we’ll see what happens after that.

Austin: Hmm, interesting, interesting. Okay, so then, which by the way, have you guys disclosed anything about the layers, number of layers that you’re stacking?

Durga: We have not yet provided all the numbers. In fact, stay tuned on that front because you haven’t asked the question, how’s the silicon coming along and so on, because we’ve talked a lot about it in terms of PowerPoint presentations and most people are actually taking our numbers for what they’re worth. But we will be showing some proof points over the next quarter or two as the silicon is back in the lab and it’s coming along very well, so stay tuned.

Austin: Okay, awesome. I’d love to hear that. So, you’ll have silicon coming back soon. Looking forward to that, looking forward to see the benchmarks. That really turns it real, from just PowerPoint. Back to the stacking of the memory, talk to us about the thermals. Is that become a constraint for that base die?

Durga: I think it is one of those things that, as Qualcomm, when we come in with—we are not, when we talk about IP, for instance, it’s not just about what we are doing for our CPU or our NPU logic and so on. We have a rich heritage of both packaging technologies and 3D stacking and a lot of these things that we do. We work very closely with the foundries in terms of how these tech need to come together. That has played in our favor as we brought out HBC. And one thing to mention is that in this space, you have to think through carefully as to which product are you designing this for, and based upon that, you have to think about the right kind of packaging so that you don’t start hitting other constraints, like it’s not the power maybe, but then you start hitting thermal constraints. So we do have a lot of experience in that space. So, different, depending upon which product it is, we have dimensioned the compute in such a way, compute plus the memory, which is dimensioned in such a way that we avoid the thermal problems. And so that is something that we’ve paid some careful attention to.

Dragonfly AI 250: 18x Bandwidth

Austin: Nice, nice, very promising. Okay, so the Dragonfly AI 250 is the first product that will have HBC in it.

Durga: That’s correct.

Austin: And that’s coming in 2027.

Durga: That’s correct.

Austin: I saw that compared to the AI 200, they both have the same amount of LPDDR in the card, it’s 768 gigs.

Durga: 768 gigs.

Austin: And yet the 250, which is the first to have the HBC, has 18X the effective bandwidth, I believe is what the website said. So again, just for everyone, could you remind us again why that HBC and the stacking is 18X the bandwidth as compared to in the 200 where presumably the LPDDR is a little bit further away?

Durga: So, first of all, the raw memory is a completely different concept compared to the actual memory bandwidth in itself. So this is LPDDR, which means we can pack, it’s a very dense card that we have. So it’s 768 gigabytes. A lot of the other things that we see in the industry, they have like 288 gigabytes per card. This is 768 gigabytes. Which, by the way, if you think through that, 768 gigabytes per card means if you take a typical order of magnitude, trillion-parameter model, and it’s all at mixed precision FP4, you can fit a trillion-parameter model in just one card. That is seriously mind-boggling. And we have shown that, we have shown some proof points associated with that. But that’s just the memory capacity.

The memory bandwidth is where, as we bring in HBC, then you have this combination of the underlying logic plus DRAM stack. So this actually is a slightly different concept. This is not about the capacity anymore. Using the same capacity, you’ve just boosted the memory bandwidth significantly. That’s why they are two independent concepts. But another way of looking at it is, if you take an AI 200 card, and it has a certain performance, certain exaflops per rack, or exaflops in a given rack, and of course, we have disclosed that our racks are like 160 kilowatts. When you go from AI 200 to AI 250, the amount of memory inside, it’s like 41 or 43 terabytes or so, that doesn’t change. But because of the fact that there’s a huge memory bandwidth increase, the number of exaflops in the rack actually goes up substantially because of that, because the decode performance improves quite a lot. So you see, the total amount of capacity that you have in the rack is the same. But now you’ve dismantled that memory wall with HBC, and that’s the reason why our overall performance goes up significantly with AI 250.

Austin: Okay, so let me reflect it back to you. There’s a part where I’m a little fuzzy on. So, we’ve got more bandwidth in the AI 250, which of course, would mean better decode performance, so just more tokens per second.

Durga: That’s correct.

Austin: How does that unlock more flops? Because the flops are not underutilized? They’re not sitting waiting. Is that what it is?

Durga: That’s right. In fact, because of the fact that you’re running it that way, your TCO also becomes better. In the sense that, if you just take it because the tokens per second per watt goes up significantly, because your tokens per second has gone up, it’s the same power consumption, by the way, same 160 kilowatts. But then in addition to that, then when you take a look at the total, it’s called the tokenomics, where you also normalize it by what’s the cost associated with that, it comes out to be even more favorable compared to some of the other things that you see in the market today. So, in that sense, every time we have talked about the benefits of AI 250, we have talked about not just tokens per second, but tokens per second, tokens per second per watt, and tokens per second per watt per dollar, or sometimes even tokens per second per dollar. When you put the entire arithmetic together, then you see why it becomes actually far more favorable to go with this.

The Primacy of TCO

Austin: Nice. Okay, so I was going to ask you then, so when you have talked to hyperscaler partners, ultimately, is there one of those that’s most compelling? Is it about the TCO or about the tokens per watt or kind of all of it?

Durga: I think increasingly these days, it’s about the TCO. Increasingly. It used to be about raw tokens per second and tokens per second per watt. But I would argue that TCO already captures that, because if you think about it, what are all the things that you need to worry about? Well, okay, if I buy one of these racks, first of all, how much am I paying for it? So that’s one variable. Second, what’s the power consumption? And then you have to figure out for that power consumption, what’s my utility rate? And there is some number associated with that. The third one, of course, is, how many tokens per second does it really generate? How effective is it? And you put all of these three variables together, and I’m discounting things like water that you need for cooling and so on. It’s the TCO that is emerging as the most important, but it does take into account these three big variables.

Austin: Yeah, yeah. No, I think that makes a lot of sense. And just the simplest way for me to think about it is like, okay, if I have a billion dollars to spend and I have 100 megawatts. If I can buy X racks from some provider that fit within the power envelope and my budget, or I can buy 2X racks from someone else, well, even if the throughput’s a little bit different, generally if you have 2X the racks, you’re going to get more throughput. So that makes a lot of sense to me.

Manufacturing at Scale

Austin: So I guess, maybe last question. There are other AI ASIC startups who also because they’re coming in and need to do something different, they’re also making different technical bets. And usually the skepticism that people get with them is they say, “Okay, that’s great. Can you actually manufacture it? Can you actually get allocation?” Now, obviously, Qualcomm is one of TSMC’s largest customers and you have wafer allocation, but I also know that you have many different business units from data center to edge who all want the same wafers. So, talk us through how that works. And I assume that you do have the advantage there because you are manufacturing things production at scale.

Durga: That’s one of the places where it is true that it’s not enough to just have the technology in itself. That’s why we’ve been placing careful bets in terms of making sure that our wafer allocations, we have an excellent set of relationships with all the foundries and also with the memory makers. And so, as they have come up as extremely valuable partners in across the entire portfolio, not just in one domain, those relationships do matter when we go in front of a customer and say, “Yeah, we are in a position to supply you. How much do you want? You want these many megawatts or maybe a gigawatt, this is what you want.” Yeah, we have the supply chain lined up for that because we do have those relationships across all the different businesses. And that actually helps us quite a lot.

Austin: Yeah, sure, totally. That makes a lot of sense. All right. Well, that’s it for today. Durga, this was great. We will have to talk again once you guys have more to share about the actual measured performance of HBC.

Durga: All right, sounds good. Thank you.

Austin: Thanks.

Durga: Thanks again.

Let’s talk about ‘super-intelligence.’

TSMC has locked in 3-6% wafer price hikes for 2027 while pushing its A14 node toward a 2028 production ramp, with Cadence and Synopsys already certifying flows. Qualcomm renewed its Apple patent license effective April 2027, DeepSeek is reportedly targeting an 8-trillion-parameter model trained on Huawei Ascend chips, and Nscale’s S-1 quietly revealed that a ByteDance entity drove 73% of its 2025 revenue.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Tell us how to make Semi Doped better! Give us your feedback! ❤️

TSMC Locks 2027 Price Hikes as EDA Vendors Certify A14 Flows

TSMC has confirmed 3-6% wafer price increases taking effect in 2027, squeezing chip designers even as the foundry accelerates its A14 node toward a 2028 production target. Cadence and Synopsys have both certified A14 design flows, with Synopsys expanding its TSMC partnership to cover A14 IP support and co-packaged optics alongside agentic AI design tools. Siemens completed its own A14 flow certification through an expanded TSMC collaboration announced via the Siemens Newsroom. IC-Link by imec joined TSMC’s Open Innovation Platform Cloud Alliance as the first managed-cloud service supporting mixed-partner EDA flows, giving European fabless companies and research groups cloud-based access to advanced-node design infrastructure across all three certified toolchains.

Vik: 3-6% and nobody’s pushing back. Order visibility runs to 2030 now, so designers have no room to argue - where else are they going to go? TSMC raises because it can. This means that a whole lot of other fabs like Samsung, UMC, and VIS can raise prices.

Qualcomm Locks In Apple License Renewal, Splits Snapdragon 8 Flagship Into Two

Qualcomm has renewed its global patent licensing agreement with Apple, effective April 2027, securing a continued revenue floor from the iPhone maker just as its modem business faces longer-term uncertainty. The chip company also used Snapdragon Summit 2026 in Maui on September 22 to unveil a two-track flagship lineup: the Snapdragon 8 Elite Extreme Gen 6 and a standard Snapdragon 8 Elite Gen 6, designed around physically distinct dies to serve different smartphone tiers. Both chips are manufactured exclusively at TSMC, according to The Elec. CEO Cristiano Amon introduced High Bandwidth Compute technology, borrowed from data-center architecture, as a feature of the new mobile platform, alongside the Adreno Neural Fusion graphics engine targeting console-level gaming performance.

Vik: HBC in a phone is the part I’d watch because the main benefit comes from the extremely low pJ/bit it takes to move data between memory and logic. HBC stacks DRAM on the logic die to get bandwidth and skip CoWoS, but the packaging pain doesn’t vanish - it just moves to 3D die stacking, the harder version. I wrote about that shift here.

DeepSeek Turns to Huawei Ascend for Frontier-Scale Model Training

DeepSeek is planning to train an 8-trillion-parameter model on Huawei chips, according to Seoul Economic Daily sourcing cited by The Elec, marking a concrete commitment to domestic silicon at a scale no Chinese lab has publicly attempted. The hardware of choice is Huawei’s Atlas 960E SuperPoD, a cluster packing 4,096 NPUs and 8 EFLOPS of FP8 compute into a single NPO/Hi-ONE optical-fabric node. That spec puts it in direct competition with Nvidia’s H100-class systems, which U.S. export controls have effectively locked out of China’s frontier AI market. Huawei built the 960E without EUV lithography, a constraint that once looked crippling and now looks like a solved engineering problem.

Vik: DeepSeek’s training right now is 2 trillion params, so this is a big jump on silicon they’ve never trained a frontier model on. Huge bet, reportedly.

Microsoft and Meta Push Hyperscale Spending Into New Geographic and Financial Territory

Microsoft plans $10 billion-plus in Gulf investment across four Persian Gulf countries through 2030, covering infrastructure buildout and cybersecurity, with President Brad Smith announcing an additional $2 billion tranche as the first significant regional update since the U.S.-Iran conflict. The commitment leans explicitly on resilience as a design principle, not just capacity. Meta, for its part, has set a 2026 capital expenditure target of $145 billion, a figure that exceeds every national military budget on earth except those of the U.S., China, and Russia. That number spans a 1 GW Ohio data center already operating, a 5 GW Louisiana facility in planning, and a proposed transoceanic subsea cable at petabit capacity. BlackRock and IFM are separately closing in on a $25 billion Asia data center deal.

Sector Watch

Memory

• Kioxia and SanDisk commit more than $31 billion in Japan fab investment to extend NAND leadership through the next technology cycle. (dqindia.com)

• CXMT gains standard DRAM market share as Samsung and SK hynix shift capacity toward HBM, per DigiTimes analysis. (digitimes)

  

• CXMT and SMIC power a 620% surge in China chip-industry profits as AI and cloud demand absorbs domestic supply, per Nikkei Asia. (Nikkei Asia)

Foundry & Packaging

• SKC will raise its stake in glass-substrate subsidiary Absolics to 82% by subscribing to Applied Materials’ unsubscribed rights-offering shares, deepening its bet on AI packaging. (The Elec)

• Simmtech enters the FC-BGA substrate market using BT resin, adding a new Korean supplier to the advanced packaging materials ecosystem targeted at AI server processors. (thelec.net)

• Rapidus links with LSTC to advance Japan’s domestic 2nm-class chipmaking program, adding an institutional design-center partner to bolster the nascent fab’s customer pipeline. (digitimes)

Compute & Edge

• Synopsys and Alchip jointly address AI scale-up and scale-out bandwidth constraints, with Alchip collaborating on Synopsys IP to target next-generation ASIC interconnect bottlenecks.

• Moore Threads MTT S5000 embodied-RL training curves match mainstream GPUs at 0.976 correlation, with dual-arm accuracy reaching 92%, per Pandaily benchmarks. (Pandaily)

• EdgeCortix unveils RAIDEN, a scalable chiplet AI platform purpose-built for physical AI applications at the edge. (Business Wire)

Optics & Networking

• PicoJool raises $27.5 million Series A to scale optical connectivity solutions targeting AI data center infrastructure. (Business Wire)

• Keppel and StarHub confirm active talks over the possible acquisition of M1, which would mark Singapore’s first telecom consolidation in years. (Light Reading)

• T-Mobile files FCC experimental licenses to test equipment in 2.7 GHz and 4 GHz bands, staking early ground in the 6G spectrum pipeline. (Light Reading)

Data Centers

• AWS becomes the first hyperscaler to achieve NATO Restricted security approval, opening a path for alliance-member classified cloud workloads. (IT Pro)

• Amazon acquires Hong Kong data centers from Grand Ming Group for up to $312 million, adding owned capacity in Asia. (Mingtiandi)

• Microsoft opens its fourth India cloud region in Telangana, expanding AI and cloud infrastructure across the subcontinent. (ET CIO)

Power

• Renesas introduces its smallest configurable GreenPAK IC family with multi-time programmability in ultra-compact WLCSP and STQFN packages. (Business Wire)

• Valens Semiconductor announces a collaboration with onsemi on a cost-optimized high-performance connectivity chipset platform. (Semiconductor Digest)

• Hammond Power Solutions signs a long-term lease for new North American manufacturing space, expanding magnetic and power electronics capacity. (Semiconductor Digest)

Policy & Trade

• US lawmakers introduce a bill to broaden the entity blacklist to cover more Chinese tech products, with Lenovo among firms facing fresh scrutiny. (SCMP)

• Trump confirms no AI guardrails agreement emerged from the Trump-Xi White House summit, leaving governance frameworks unresolved. (Bloomberg.com)

• Meta unveils the Charm dedicated AI wearable device alongside $349 camera-free Ray-Ban glasses with Muse AI integration, expanding its hardware AI ecosystem. (Bloomberg.com)

ICYMI / Random:

• Trump wants AI to be rebranded as super-intelligence. (BBC)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

TechInsights has confirmed Apple’s A20 Pro runs on TSMC’s N2 node, making it the first independently verified 2nm chip in a shipping consumer device. Qualcomm and MediaTek both dropped new flagship mobile chips, Huawei’s Kirin 9050 Pro is reportedly outperforming Apple’s A17 Pro despite lacking EUV access, and CoreWeave closed a $4.2 billion convertible offering as AI infrastructure debt markets get a stress test.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Tell us how to make Semi Doped better! Give us your feedback! ❤️

OpenAI and Anthropic Cut Model Costs as Infrastructure Ambitions Grow

OpenAI expanded its GPT-6 family with two cheaper variants, Sol and Luna, now available on Amazon Bedrock, as both frontier labs pressed deeper into a simultaneous cost-reduction race. Anthropic matched the timing with Claude Opus 5.5, a model designed to handle a broad range of business tasks at lower inference cost.
Anthropic is releasing Opus 5.5 ahead of a highly anticipated Wall Street debut, according to Bloomberg, making the pricing push part of an IPO-season positioning play. The competitive symmetry is hard to miss: both companies dropped cheaper tiers on the same news cycle. Beyond pricing, Anthropic is in talks to take direct control over more data center capacity, per The Information, folding infrastructure ownership into the same strategic frame as the model cost war.

Vik: Both cutting prices on the same news cycle looks like competitive theater. But Anthropic’s timing is the tell - dropping Opus 5.5 right before a highly anticipated Wall Street debut is IPO positioning, not a demand surprise. Bloomberg frames it exactly that way.

Qualcomm Snapdragon 8 Elite Gen 6 and MediaTek 9600 Pro Arrive Together

Qualcomm’s Snapdragon 8 Elite Gen 6 clocks at 5 GHz and adds agentic AI capabilities to premium Android, while MediaTek counters with the Dimensity 9600 Pro, a 2nm-class chip landing in OPPO’s Find X10 Pro Max. The OPPO Find X10 Pro Max pairs the Dimensity 9600 Pro with three 200MP Hasselblad cameras and an 8000mAh battery, pricing from roughly RM4099 in China. Samsung has validated LPDDR6 memory for Qualcomm ahead of the Galaxy S27 Ultra. Both suppliers have secured flagship design wins in the same launch cycle, each on 3nm-class silicon, each backed by a major OEM commitment. The rivalry hasn’t been this symmetrical in years.

Vik: Samsung got LPDDR6 running at 10.7 Gbps on the Extreme Gen 6, but they’re reportedly dragging their feet on the S27 Ultra because memory got so expensive they might just stay on LPDDR5X. See TrendForce.

TechInsights Teardown Confirms Apple A20 Pro Running on TSMC 2nm

TechInsights has torn down the iPhone 18 Pro and confirmed that Apple’s A20 Pro runs on TSMC’s N2 process node, making it the first independently verified 2nm chip inside a shipping consumer device. The iPhone 18 Pro and Pro Max went on sale September 18 across more than 65 countries, including the US, South Korea, Japan, the UK, and China, running iOS 27. The confirmation matters beyond Apple itself: it establishes that TSMC’s N2 node is yielding at commercial scale, with Apple absorbing early production volumes as the anchor customer. TechInsights’ die analysis positions every other fabless customer queued behind Apple on N2 with clearer, if more expensive, ground to stand on.

Vik: On N2, Apple spent the density gains on an extra GPU core and more neural area, not a smaller die. Apple eating first yield is what clears the runway for every fabless customer stacked behind them.

Huawei Kirin 9050 Pro Tops A17 Pro as DeepSeek Orders 160,000 Ascend Units

Bernstein estimates Huawei’s Kirin 9050 Pro, built on technology equivalent to a 7-nanometre process node without EUV access, outperforms Apple’s 3nm A17 Pro in Geekbench 6 multicore tests, cutting the generational gap from roughly four years to about three. The chip still trails Apple’s current A20 Pro by around 30%, so the distance has narrowed, not closed.

On the data-center side, DeepSeek has committed to 160,000 Ascend 960DT units for AI training, with supply reportedly arriving in the fourth quarter of 2026, targeting an 8-trillion-parameter model. Together, the two developments represent Huawei silicon reaching commercial relevance in both smartphone and large-scale AI training segments simultaneously, according to Bernstein’s analysis.

Qualcomm Acquires PickNik Robotics to Deepen Edge AI Push

Qualcomm is acquiring PickNik Robotics, the company behind MoveIt, a widely used open-source robotics manipulation framework with broad adoption across research and industrial applications. The deal gives Qualcomm both MoveIt and the commercial MoveIt Pro product, which it plans to fold into its Dragonwing robotics platforms.

Dragonwing is Qualcomm’s existing lineup targeting edge AI and robotics deployments, and adding PickNik’s software stack would give it a motion-planning layer to complement the silicon. PickNik built a real business around open-source software, which is a trick not everyone pulls off. Financial terms of the acquisition were not disclosed.

Vik: This is Qualcomm reportedly buying the motion-planning layer so Dragonwing ships more than silicon. Makes sense. But robotics still hasn’t had its ChatGPT moment - that arrives when robots do household work, not when a warehouse arm plans a path.

Sector Watch

Compute

• Supermicro begins shipping Nvidia Vera Rubin NVL72 racks, marking the first production delivery of next-generation Rubin-architecture AI server systems. (PR Newswire UK)

• Microsoft grants DARPA hands-on access to its quantum computing system at a new Maryland facility, the first US government agency with direct access to the platform. (Reuters)

• Nexstrom raises $12 million seed round to commercialize the first 12-inch single-crystal 2D semiconductor wafer platform, targeting a post-silicon scaling substrate. (PR Newswire)

Foundry & Packaging

• ASM launches flowable carbon deposition technology for sub-2nm patterning, addressing a critical gap-fill materials challenge for next-era semiconductor scaling. (Semiconductor Digest)

• Nexperia and Tata Electronics form a strategic partnership covering wafer fabrication, assembly, and test in India, adding a second major fab-ecosystem partnership alongside ASML-Tata. (EE News Europe)

• BOE joins the glass substrate race alongside Samsung Electro-Mechanics, SKC, and LG, with all four targeting 2027–2028 production ramp per TrendForce. (TrendForce)

Memory

• HanmiGlobal takes on project-management role for SK hynix’s West Lafayette, Indiana packaging fab, supporting the U.S. HBM supply chain buildout. (The Elec)

• Samsung officially launches P9 and P7 portable SSD lineup in September, its next-generation storage products first shown at Gamescom. (Samsung Semiconductor)

• CXMT triples factory commitments to meet AI-driven DRAM demand, intensifying competitive pressure on Samsung and SK hynix, per Chosunbiz. (Chosunbiz)

Power & Data Centers

• ABB launches Infinitus, a direct-current power portfolio purpose-built for AI data centers, targeting the efficiency and density demands of GPU clusters. (EE News Europe)

• Schneider Electric launches a new Motivair coolant distribution unit combining liquid and air cooling, offering flexible thermal management for AI data center deployments. (The Daily Tribune News)

• Alibaba Cloud plans to open new data centers in Turkey, Finland, and the Netherlands next year, accelerating its overseas expansion into Europe. (SCMP)

Optics & Networking

• Semtech moves its 50G mobile transport portfolio into full production, targeting the higher fronthaul capacity required for 5G-Advanced networks. (EE News Europe)

• Vocus warns subsea cable construction resources are fully booked into the 2030s, the latest operator to flag a physical capacity crunch in transoceanic connectivity infrastructure. (Light Reading)

• Airspan acquires the bulk of Cambium Networks’ product lines, adding more than 135 employees as Cambium restructures amid financial difficulties. (Light Reading)

Policy & Trade

• European Commission proposes a common EU data center energy and water use rating scheme, introducing mandatory transparency requirements for operators across member states. (EE News Europe)

• CanSemi Tech joins China’s chip IPO wave with a $919 million deal, reflecting continued domestic capital formation in the semiconductor sector. (Bloomberg.com)

• China surveys Broadcom networking-switch deployments in state-run data centers, signaling a drive to identify and reduce reliance on U.S. networking silicon. (Reuters)

Edge

• GigaDevice showcases its edge-intelligence MCU and SoC portfolio at Embedded World North America, targeting industrial, infrastructure, and consumer edge markets. (Business Wire)

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Alibaba unveiled what it’s calling China’s most powerful AI chip alongside a new 10-trillion-parameter model at its Apsara Conference in Hangzhou. Texas froze all state-issued data-center permits pending a grid audit, California signed a new oversight package, and TSMC pulled its A14 pilot run forward to Q1 2027 as Qualcomm eyes Samsung for 2nm.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Tell us how to make Semi Doped better! Give us your feedback! ❤️

Alibaba Unveils AI Chip and 10-Trillion-Parameter Model at Apsara

At its annual Apsara Conference in Hangzhou on Tuesday, Alibaba Group Holding introduced what it calls China’s most powerful AI chip, a new accelerator positioned to compete directly with Nvidia and support a major expansion of the company’s data center capacity. Group chairman Joe Tsai told attendees that Alibaba is building “full-stack AI” end-to-end capabilities, a commitment that came paired with the company’s first public mention of plans to train a model reaching up to 10 trillion parameters. Alibaba’s shares rose as much as 5% on the news, according to Reuters. The parameter count would place the forthcoming model well beyond most publicly disclosed architectures, though no release date was given for either the chip or the model at scale.

Vik: Lots of juicy details in the linked SCMP post.

• Qwen 5 will be a 5-10T parameter model. Qwen Intelligence is a full-stack agent platform for AI on smartphones.

• Zhenwu V900 processor is the most power AI chip in the market today. Clusters of 500,000 GPUs will comprise of these chips. Next gen: Zhenwu J900

  • 216 GB of HBM, 1.2 TB link to other chips. GA in Q1 2027

• Plans for next-generation proprietary CPUs tailored for agentic AI tasks – the Yitian 720 and Yitian 730, both due in 2027.

Texas Freezes Data-Center Permits, California Signs Oversight Package

Governor Abbott halted all state-issued data-center permits in Texas pending completion of a grid audit, a sweeping pause that hits the largest data-center market in the US before a single new facility can break ground. California’s governor signed a broad oversight bill package in the same news cycle, adding formal state-level review requirements to what had previously been a tangle of local opposition.

Two states, two different political parties, one shared conclusion: hyperscale buildout needs a leash. Bloomberg notes the permit freeze has put Republicans in an awkward spot, caught between grid-reliability constituents and the industry investment they courted. Neither state has disclosed a timeline for resolving its process, leaving capacity schedules for major cloud and colocation operators in open-ended limbo.

Vik: Its not Dario saying that he will pace the frontier of AI. It’s ultimately NIMBY-ism that will pace it for us. We will humbly take a victory lap for saying this will happen.

Vicor Lifts Q3 Outlook More Than 20% as AI Power Royalties Accelerate

Vicor raised its Q3 2026 revenue guidance to reflect over 20% growth, citing royalties from its Vicor Power-Delivery architecture licensed into high-density AI server racks. The VPD licensing stream means Vicor collects a fee regardless of which GPU vendor occupies a given server slot, a position the company has been building toward as rack power densities climb beyond what conventional voltage-regulator designs can handle. Shares have now gained 316% in 2026 heading into the updated guidance, reflecting investor appetite for the royalty model over pure component sales. The guidance revision covers the quarter ending September 2026.

Vik: This 20% bump is a licensing play, not a demand surprise. Vinciarelli basically says unlicensed hyperscalers only called after their systems got banned from import over infringing power modules. VPD is vertical power delivery, feeding the GPU from underneath instead of across the board - that’s how you survive these rack densities.

TSMC Pulls A14 Pilot to Early 2027 as Qualcomm Eyes Samsung for 2nm

TSMC has moved its A14 pilot run forward to Q1 2027, according to TrendForce, compressing what had been a comfortable lead over Intel’s competing 14A node. The schedule shift is a direct response to Intel’s accelerating foundry roadmap. Qualcomm, watching this tighten from the customer side, is reportedly evaluating Samsung as a co-source for 2nm production alongside TSMC, a move that would give the chip designer supply-chain insurance while keeping both foundries under competitive pressure. TSMC pulling A14 to Q1 2027 puts its most advanced node on a collision course with Intel 14A within the same calendar window. Dual-sourcing at 2nm isn’t a simple task; Samsung’s yield track record at leading edge has historically trailed TSMC’s, and Qualcomm’s evaluation is described only as ongoing.

Vik: Qualcomm dual-sourcing 2nm is really about TSMC capacity being locked up by Apple, not love for Samsung. And we’ve seen this one: Samsung built the Snapdragon 8 Gen 1 on 4nm in 2021, it ran hot, and Qualcomm went straight back to TSMC. Yields still trail. but it is still “Insurance”, reportedly.

Google Launches $899 Googlebook as MediaTek Challenges Qualcomm for AI-PC Silicon

Google opened pre-orders for Googlebook laptops starting at $899, a five-model lineup built on a hybrid of Android and ChromeOS with Gemini AI at the center and tighter integration with Android phones. Qualcomm’s Snapdragon X Elite is the launch silicon, giving the chip its first major Google platform win in the premium laptop tier. MediaTek isn’t sitting out: the company debuted the Dimensity CX C10 Max, a 3nm chip with an 8-core Arm CPU, 11-core GPU, and 55 TOPS NPU, built specifically for Googlebook devices, with Lenovo already signed as a design-win partner. Android Central confirmed the Snapdragon X Elite as the processor powering the first wave of Googlebook hardware. Google’s Sameer Samat, president of the Android Ecosystem, framed the platform as a direct challenge to Apple and Microsoft in the AI PC race.

Vik: MediaTek pitches the C10 Max as built for Googlebook, but it’s basically the Kompanio Ultra 910 Chromebook chip rebranded - same X925 core, same Immortalis GPU, just clocked a little higher. Qualcomm still grabs the premium launch slot with Dell and HP.

Sector Watch

Memory

• YMTC wins two German patent injunctions against Micron on NAND technology, marking a significant IP offensive by China’s leading flash maker. (IAM Patent)

• CXMT boosts dies-per-wafer by 50% with fifth-generation 12nm-class DRAM using quadruple-patterning, advancing without EUV. (chosun.com)

• SK hynix CEO Kwak Noh-Jung publicly commits to sustaining HBM and next-generation memory technology leadership as AI demand accelerates. (The Elec)

Foundry & Packaging

• Hanmi Semiconductor ships its first FC Bonder 75, entering the 2.5D AI chip packaging market with global delivery of the new flip-chip bonding tool. (Chosunbiz)

• JNTC will invest 347 billion won to build a through-glass via (TGV) glass substrate plant in Gimcheon, South Korea, targeting AI packaging demand. (The Elec)

• EST files for Kosdaq IPO targeting glass substrate and OLEDoS component markets, joining a wave of Korean advanced-packaging material plays. (The Elec)

Compute & Edge

• Cadence expands its ChipStack AI Super Agent with a new RTL-generation and early PPA optimization agent, deepening AI-driven EDA automation. (Business Wire)

• Rebellions NPU now powers four live AI services at SK Telecom, marking a commercial deployment milestone for the Korean AI chip startup. (The Elec)

• d-Matrix joins Nvidia NVLink Fusion ecosystem with its Raptor inference chip delivering 100 TB/s bandwidth, expanding the in-network compute tier. (shattered.io)

Optics & Networking

• Flag selects Ciena’s Waveserver platform to upgrade the Echo subsea cable, extending Ciena’s coherent-optics footprint on a key Pacific route. (Light Reading)

• Semtech joins the XPO Multi-Source Agreement with 200G-per-lane modulator driver and TIA ICs, adding merchant silicon to the extra-dense pluggable optics effort. (EE News Europe)

• Lightera announces a major capital investment to expand submarine optical fiber manufacturing capacity as hyperscale cable demand surges. (PR Newswire)

Power

• Nvidia launches DSX Ready, a qualification program for power and cooling products certified for AI factory infrastructure deployments. (Nvidia News)

• Infineon launches the PSOC Control C3 MCU family for real-time power and motor control, adding post-quantum cryptography security to the platform. (Bisinfotech)

• Mitsubishi Electric introduces a chip-to-grid power blueprint for Nvidia-powered AI factories, covering conversion from utility supply to chip-level delivery. (Business Wire)

Data Centers

• Crusoe closes an initial $3.9 billion Series F backed by Nvidia, pivoting to modular AI data center builds and reaching a $30.9 billion valuation. (HPCwire)

• CoreWeave-tied data center project launches a Goldman-led $1.1 billion junk-bond offering, the latest in a string of project-finance deals for leased AI capacity. (Bloomberg.com)

• BDx breaks ground on a 640 MW AI data center in West Java, Indonesia, backed by 845 MVA of secured grid power. (PR Newswire)

Policy & Trade

• ASML executive warns Europe risks falling behind in chipmaking after revealing no lithography machines have been sold to European fabs. (The Edge Malaysia)

• Australia state bans new data centers from residential zones, a first-of-kind zoning restriction with implications for hyperscale site selection across the region. (Bloomberg.com)

• China and US hold a second day of trade talks covering AI governance and investment access, with the Pentagon entity blacklist remaining a sticking point. (Bloomberg.com)

Tell us how to make Semi Doped better! Give us your feedback! ❤️