Hello world, it’s Monday, October 26th.
Anthropic files S1. Andy Jassy says AWS capacity is sold out into 2028, while a reported $5 billion of Nvidia GB300 racks sits in warehouses waiting on power. Elsewhere, SK Hynix is reportedly planning a $12 billion special dividend from gains on its Kioxia stake, Infineon lands a GaN design win in Nvidia’s 800 VDC rack architecture, and leaked benchmarks have Samsung’s Exynos 2700 ahead of Qualcomm’s Snapdragon 8 Elite Extreme Gen 6 by up to 20 percent.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
Presented by Crusoe
Crusoe builds the full stack for AI, from power and data centers to the cloud. It can also help you tune a small model on your own data for just a few bucks.
For repeatable work like reports, an open model trained on your terminology and formats can match a frontier model, and it’s cheaper and faster to run.
Crusoe’s Serverless Fine-Tuning handles the training: upload your dataset, pick an open model like Llama, DeepSeek, or GLM, and pay per million tokens used, not per GPU hour.
About $2 buys roughly 5M training tokens on smaller models.
Deploy on Crusoe Cloud in one click or download the weights and deploy on your own.
Anthropic Files Public S-1; Broadcom TPU Leases Are the Largest Compute Commitment
Anthropic publicly filed its S-1 to list on Nasdaq under “ANT,” with Goldman Sachs and Morgan Stanley leading and the price range still blank. First-half revenue was $16.3 billion against an operating loss of $693 million; compute and infrastructure spend of $8.9 billion fell to 55% of revenue from 160% in 2025. The largest noncancelable commitment is $161.2 billion to Broadcom, in five-year leases on TPU-based hardware, ahead of Google at $114.5 billion, Amazon at $110 billion and Microsoft at $44.9 billion. Another roughly $207 billion with SpaceX’s AI unit, Nscale and Lambda can be cancelled. Net revenue retention is about 900%, and the seven co-founders hold 50.1% of the vote through a single Class F share.
Vik: Apart from the broad numbers above, haven’t had the time to dig into the S1. I’m sure it has some really juicy details. Anthropic is the fastest rising company in history, and it going public is serious news.
Amazon Books AWS Compute Through 2028 While $5 Billion in Nvidia Servers Sit Dark
Amazon CEO Andy Jassy told investors that AWS capacity is sold out through 2028, a booking horizon that stretches well past the 2027 figure he cited to Reuters just weeks earlier. The demand is real and contracted. The catch is that a reported $5 billion worth of Nvidia AI servers sits idle in Amazon warehouses, powered off not for lack of buyers but for lack of electricity to run them. Amazon is, in effect, simultaneously oversold on compute and unable to deploy the hardware it has already purchased. Power infrastructure, not chip supply, is what prevents those servers from going live and generating revenue. Jassy has acknowledged that computing supply cannot keep pace with demand, a constraint that Wall Street has flagged but that booked capacity figures alone obscure.
Vik: Dark GPUs. The binding constraint is power, not silicon - $5B of GB300 racks parked in warehouses because the sites can’t turn them on. Compute’s short, electrons are shorter. Feel the GPU glut coming? Ken Griffin from Citadel explains what happens when GPU glut shows up.
Austin: Read the Business Insider report carefully though. The $5B is what AWS expected to take delivery of in August under its early rack receipt process, with another $3B of GB300 racks a month after that. So it’s a delivery forecast, and nobody has said how much is sitting dark today. And Jassy’s actual words were “a good chunk of 2028 is sold out.” Power problem is real. Numbers are softer than the headline.
SK Hynix Plans $12 Billion Dividend Backed by Kioxia Stake Gains
SK Hynix shares bucked a broad AI-stock selloff this week after Bloomberg reported the South Korean memory maker plans to distribute roughly $12 billion to shareholders from gains on its Kioxia Holdings stake. The Japanese flash-storage company’s rising valuation has handed SK Hynix a windfall large enough to fund what would rank among the largest dividends in the memory sector’s history. Timing is tight: the company is scheduled to announce specific shareholder-return measures on Wednesday, according to Reuters. SK Hynix has held a stake in Kioxia since 2018, when it joined a Bain Capital-led consortium that acquired the business from Toshiba.
Austin: Remember the original check here was about $3.5B (395B yen) into the Bain consortium for Toshiba Memory, and now the dividend alone is $12B. That’s quite a trade on a NAND competitor. Open question is whether SK Hynix keeps any of Kioxia after the selldown or exits entirely.
Ajinomoto Reveals Next-Generation ABF Films as AI Substrate Crunch Deepens
At IMAPCT, Ajinomoto unveiled advanced ABF laminate technologies targeting the tighter dimensional tolerances and thermal demands of AI accelerator packaging, arriving at a moment when the materials feeding that packaging stack are drawing fresh scrutiny across the supply chain. Korean substrate suppliers are splitting on AI order momentum, with some capturing outsized share while others stagnate, according to Chosunbiz. Nvidia is actively courting Samsung to supply FC-BGA substrates, a move that directly threatens Taiwanese incumbents Unimicron and Kinsus, which have dominated that business. ABF film sits upstream of every FC-BGA panel those suppliers produce, so Ajinomoto’s roadmap disclosures land with particular weight for manufacturers scrambling to qualify next-generation laminates before capacity decisions get locked in. The geometry is tightening fast, and the film is not the easy part.
Vik: the substrate crunch is arguably the biggest constraint right now in the AI data center buildout. The UniMicron CEO recently pointed out that we are expected to face a 20% shortfall through 2027. Everything from T-glass to ABF and CCL is backlogged
Infineon Wins GaN Power Design in Nvidia 800 VDC Rack Architecture
Infineon Technologies has secured a GaN power semiconductor design win inside Nvidia’s 800 VDC data-center rack architecture, confirming the German chipmaker as a key supplier for the platform widely expected to succeed today’s 400 VDC installations. Infineon will supply GaN power chips for Nvidia’s 800 V data centers, a designation that locks the company into the next generation of AI infrastructure power delivery before the Vera Rubin platform ramps. The shift from 400 VDC to 800 VDC enables higher power density per rack and suits the efficiency demands of GPU-heavy AI clusters. Competing power-semiconductor suppliers now face the qualification clock: winning a spot on a successive Nvidia platform typically requires demonstrated silicon before volume production begins, and Vera Rubin’s ramp timeline compresses that window considerably.
Vik: Infineon’s turning into a real AI power play. The design win sits at the intermediate bus converter, where GaN switching near 1 MHz shrinks the magnetics and lowers board height - exactly what you need as racks push towards MW.
Austin: This is the GaN Systems deal paying off. Infineon bought them for $830M in 2023 and that portfolio now sits in the Vera Rubin power chain. Design win here means a seat at a long table though. Nvidia’s own 800 VDC silicon list has Infineon, MPS, Navitas, onsemi, ROHM, ST, and TI on it, plus a few more.
Samsung Exynos 2700 Clears Qualcomm Snapdragon in Leaked Benchmark Tests
Early benchmark results for Samsung’s Exynos 2700 show the chip outpacing Qualcomm’s Snapdragon 8 Elite Extreme Gen 6 by up to 20 percent, according to leaked internal test data reported across multiple trade outlets. The Exynos 2700 is built on a 2nm process node, giving it a fabrication-generation edge over the Qualcomm part it is being measured against. Samsung’s chip is expected to power the Galaxy S27 flagship, per 파이낸셜뉴스, which would mark a meaningful shift after Qualcomm supplied the Snapdragon variant across much of Samsung’s recent premium lineup. The Korea Herald and Notebookcheck both peg the performance gap at nearly 20 percent, though all figures come from pre-production tests and no final silicon has been announced.
Vik: Leaked internal Samsung numbers nobody’s verified, so grain of salt. And Geekbench peaks tell you nothing about sustained thermals or power draw - which the leak conveniently leaves out. Not in volume yet anyway.
NVIDIA Lowers Its Official HBM4 Pin Speed
But why? 11.7 → 9.6 Gbps (Jukan issued a correction in the comments)
Sector Watch
Foundry & Packaging
Intel accelerates equipment installation at Arizona Fab 62, targeting late-2028 operations as its sixth high-volume node on the Ocotillo campus. (The Elec)
HCLTech signs partnership with Rapidus and opens two dedicated semiconductor engineering centres to support Japan’s 2nm program. (Business Standard)
Memory
imec publishes research on 3D volumetric DRAM-on-GPU stacking, expanding the thermal design envelope beyond conventional HBM-on-GPU and pointing toward a post-HBM memory integration roadmap. (Semiconductor Engineering)
YMTC targets 20% global NAND market share while pushing into DRAM with a 4F² cell structure aimed at 9,600 MT/s LPDDR5X mass production. (Businesskorea)
PCB & Components
TLB begins development of SOCAMM3 substrates for AI servers, targeting the next-generation small-outline memory module standard as demand for advanced memory packaging accelerates. (The Elec)
Samsung Electro-Mechanics develops new technology to suppress moisture ingress and crack formation in high-capacity MLCCs. (머니투데이)
Suzhou Dongshan Precision raises HK$9.9 billion in Hong Kong’s biggest IPO of the year, funding PCB capacity expansion. (Bloomberg Tech)
Compute & Edge
Qualcomm announces global job cuts starting November as CEO Cristiano Amon moves to streamline operations amid shifting mobile and PC market dynamics. (LatestLY)
Supermicro launches a 100% liquid-cooled Nvidia Vera CPU rack system targeting next-generation AI inference deployments. (Data Center Dynamics)
Optics & Networking
Tata Communications commits up to $450 million to build a long-haul fibre optic AI Superhighway linking India’s key technology hubs. (Bloomberg Tech)
Gigaphoton extends its DUV excimer laser business beyond scanners, targeting new precision light-source applications across the fab ecosystem. (Semiconductor Digest)
Power
Schneider Electric launches a 2.5 MW modular power module engineered specifically for Nvidia Vera Rubin AI data center infrastructure. (THE ELEC, Korea Electronics Industry Media)
GE Vernova targets a $200 billion order backlog by early 2027, driven by surging AI data center power demand for gas turbines. (scanx.trade)
Hyosung Heavy Industries targets domestic development of 2 GW VSC-HVDC technology by 2027, reducing Korea’s dependence on foreign power-transmission systems. (The Elec)
Data Centers
TikTok U.S. is sourcing up to 4 gigawatts of computing capacity by the early 2030s, emerging as one of the largest single AI infrastructure buyers outside the established hyperscalers. (CNBC)
Google is ordered to halt construction at two Finnish AI data center sites after illegally clearing an estimated 250,000 trees across 740 acres. (Fortune)
SK Hyper, SK Telecom’s AI data center unit, targets its first large-scale external funding round by year-end as it scales capacity. (The Elec)
Policy & Trade
Nvidia CEO Jensen Huang pitches sovereign AI infrastructure to Kazakhstan, announcing plans for tens of thousands of GPUs with partner Firebird. (SCMP)
Vietnam’s first domestic chip plant is unlikely to meet its 2027 target, with Reuters reporting no deal yet signed with ASML for lithography tools. (Reuters)
Cambricon Q3 profit surges 148% year-on-year to outpace estimates, but revenue of 3.53 billion yuan misses on output constraints, underscoring China’s AI chip production bottleneck. (Bloomberg.com)
Tell us how to make Semi Doped better! Give us your feedback! ❤️