Hello world, it’s Wednesday, October 21st.
Anthropic’s draft S-1 puts up to $750 billion in future compute obligations on the record for the first time, alongside an $80.7 billion first-half net loss against $16.3 billion in revenue. Elsewhere, Elon Musk says ASML and other toolmakers are expediting equipment for Terafab’s 2028 push, Super Micro lifted its preliminary fiscal Q1 revenue to $18.5 billion to $19.5 billion on AI server demand, and Vicor posted Q3 revenue of $197.1 million, up 78.5% year over year.
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’s Draft S-1 Turns Revenue Story Into $750 Billion Compute Liability
Anthropic’s IPO filing, reported by the Wall Street Journal and confirmed across filings coverage, puts up to $750 billion in future compute obligations on the record for the first time. The company posted an $80.7 billion net loss in the first half of 2026 alone, against $16.3 billion in revenue. Anthropic depends on AWS, Google, and Microsoft for compute supply, and the filing discloses difficulty securing adequate capacity from all three. The $750 billion figure isn’t a budget line but a ceiling on potential obligations, yet its scale now puts cloud providers’ capacity commitments on the public record. Hedgeye’s Felix Wang flags Chinese AI competition as an additional investor concern heading into the listing.
Vik: For us the signal is noncancelable - much of that compute bill is payable regardless of usage. That’s a floor under accelerator and HBM orders even if the IPO wobbles. And they’re getting squeezed on capacity because the clouds treat Anthropic as a competitor now, not a customer.
Austin: Keep the $80.7B in context. Roughly $72B of it is non-cash, accounting charges on the convertible notes and other instruments Anthropic raised money with. I’d guess that’s how yesterday’s profitable Q3 and today’s first-half loss can both be true.
Terafab Locks In ASML Gear and Korean Suppliers for 2028 Chip Push
Elon Musk confirmed that ASML and other chipmaking equipment vendors have expedited deliveries to Terafab, the joint Tesla and SpaceX semiconductor manufacturing venture, clearing one of the harder early hurdles for a project that targets volume production by 2028. The supply chain is taking shape on the materials side too: roughly ten South Korean suppliers of semiconductor materials, components, and equipment are in active talks with Terafab, and several have already signed NDAs and moved into discussions over specific supply terms. HPSP, the Korean high-pressure hydrogen annealing specialist, has already contracted to supply Terafab’s research line. Alongside TSMC Arizona and Intel’s revived domestic ambitions, Terafab is shaping up as a third major advanced-fab program on U.S. soil.
Vik: One real contract so far on the Korean side - HPSP, research line only, not the production fab. Everything else is NDAs and sample swaps. Hanmi on bonding equipment is interesting since packaging is where a project like this pays out first. Early days.
Austin: Expedited is doing a lot of work here. ASML said last week it’s sold out on EUV through 2028, and when an analyst asked on the call whether SpaceX had actually ordered EUV tools, Musk didn’t answer. So which tools are being expedited, and whose slot did they come out of?
Super Micro’s Raised AI Server Outlook Lifts Memory Stocks in Seoul
Super Micro raised its fiscal first-quarter revenue outlook to between $6.6 billion and $6.8 billion, up sharply from its prior guidance of $5.6 billion to $6.4 billion, citing surging demand for AI servers built around Nvidia’s latest GPU platforms. The revision landed as a preliminary figure ahead of the company’s full earnings release, giving the market an unusually clean read on accelerating hyperscaler deployment. Samsung Electronics and SK Hynix each rebounded on the Seoul exchange in response, as investors read the AI server build rate as a direct proxy for high-bandwidth memory consumption. The Taiex also recovered, led by AI-adjacent names in Taiwan’s supply chain. Super Micro shares climbed on the news, with the company also flagging improved gross margin expectations alongside the higher revenue figure.
Vik: The bounce makes sense - AI server build rate is the cleanest proxy we have for HBM pull. But watch the divergence: SK Hynix popped on the news while foreigners kept net selling, 18 sessions straight. Skittishness.
Samsung M2800 Modem Targets 2nm as Qualcomm Eyes Same Node
Samsung Electronics is advancing its M2800 mobile modem to a 2nm process, stepping up from the 4nm node used for its current external modems, after the M2700 cleared MX division validation. The move comes alongside a reorganization of Samsung’s 4nm capacity toward base dies for high-bandwidth memory, freeing the leading-edge node for new priorities. Qualcomm is now evaluating Samsung’s 2nm process for future Snapdragon chips, putting Samsung Foundry back in contention for high-volume flagship mobile logic. Two of the industry’s largest fabless and IDM players converging on the same nascent node at roughly the same time is either a coincidence or a vote of confidence, and probably both. TSMC currently holds dominant share of flagship mobile SoC production at the leading edge.
Vik: Modem to 2nm is fine. The more interesting point is Samsung reorganizing 4nm capacity onto HBM base dies - their in-house 4nm logic is exactly the edge in base die design, while TSMC’s 12nm node held their HBM4 lane speeds back (base die note).
Vicor Q3 Revenue Hits $197 Million as AI Power Backlog Builds
Vicor posted third-quarter revenue of $197.1 million, a 78.5% jump from the year-ago period, driven by surging demand for its high-density 48V power delivery architecture in AI accelerator racks. EPS came in at $1.33, topping consensus by four cents. The company’s backlog rose sharply alongside revenue, a combination that shifts the near-term constraint from customer demand to Vicor’s own manufacturing throughput. Vicor’s power components sit close to the silicon in AI servers, where conventional power architectures struggle to feed the current densities that modern accelerators require. The numbers confirm that rack-level power delivery is growing fast enough to stress the supply chain.
Austin: Vertical power delivery is the most interesting topic here. Vertical power delivery is one of the things pushing advanced packaging right now, feeding current straight up into the die. Gross margin at 68% is impressive.
Sector Watch
Memory
SK Hynix will host its Q3 2026 earnings call on October 28, 2026, with a press release before 9:00 a.m. and a live Korean-English interpreted conference call to follow. (SK Hynix)
Micron invests in Indium Corporation’s New York critical-minerals processing facility to secure domestic indium supply for chip manufacturing. (Micron)
Compute & Edge
Qualcomm plans global performance-driven layoffs beginning November, confirmed by an internal CEO memo, as it faces simultaneous headwinds from Arm royalty litigation and Galaxy S27 modem thermal issues. (Reuters)
Altera targets physical AI markets including robotics and autonomous vehicles with FPGAs designed for Nvidia Sensor Bridge integration. (The Elec)
Google open-sources TorchTPU, enabling native PyTorch workloads on its TPU chips in a direct software-layer challenge to Nvidia’s CUDA ecosystem. (Unite.AI)
Foundry & Packaging
TSMC suspends some Kaohsiung construction work after a scaffolding collapse kills four workers at its fab site in southern Taiwan. (Bloomberg Tech)
Kinsus commits NT$23.5 billion to expand AI chip substrate capacity, betting on sustained ABF and glass-core substrate demand through 2028. (Kinsus)
Unimicron chairman forecasts substrate panel sizes will reach 14 times current scale by 2030, outlining the physical roadmap for next-generation AI packaging. (digitimes)
Policy & Trade
ByteDance, Alibaba, and Tencent have placed over $35 billion in Nvidia chip orders, according to a Reuters exclusive, marking one of the largest disclosed Chinese hyperscaler procurement rounds. (Reuters)
Samsung and SK Hynix are reportedly upgrading their China NAND fabs to more advanced nodes to defend market share as YMTC overtakes Samsung in China’s domestic NAND flash market. (digitimes)
Optics & Networking
NTT-led consortium launches program to develop mass-production technology for optical chiplets, targeting high-volume co-packaged optics manufacturing. (Telecompaper)
Advanced Photonics Coalition publishes a portfolio of specifications and best-practice documents to accelerate high-volume manufacturing of optical networking gear. (PR Newswire)
Singtel and Ericsson sign a five-year partnership to scale Singapore’s 5G Advanced capabilities and deploy AI-native network infrastructure. (Light Reading)
Data Centers
Nextdc slashes full-year earnings guidance by one-third and halts development plans after public backlash delays power supply to an Australian data center site. (Bloomberg Tech)
Nvidia introduces DSX POD AI Factory reference architecture in Berlin, providing a turnkey blueprint for European operators building AI compute infrastructure. (Nvidia News)
Power
Infineon, Volkswagen CARIAD, and VCTC establish a joint automotive chip innovation center in Hefei targeting next-generation vehicle power and control silicon. (Gasgoo)
Vinatech secures over 200 billion won in data-center supercapacitor orders in the second half of 2026 alone, signaling rapid adoption of ultracapacitor-based UPS solutions. (The Elec)
Tell us how to make Semi Doped better! Give us your feedback! ❤️