Hello world, it’s Thursday, October 22nd.
Meta narrowed its 2026 capital expenditure guidance to $135 billion to $145 billion and spent $34.9 billion in the third quarter alone, a sign that big tech AI spending isn’t slowing down anytime soon. Elsewhere, Lam Research and ASML are flagging a tighter equipment supply outlook, Huawei closed a $2.3 billion offshore loan as DeepSeek raised $15 billion on the same day, and Besi saw orders surge 96% on hybrid bonding demand.
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.
Meta Lifts AI CapEx to $145 Billion as IBM and Nokia Beat on AI Demand
Meta raised its 2026 capital expenditure guidance to $145 billion, with CEO Mark Zuckerberg telling investors the company’s AI spending is delivering results, and announced a new class of nonvoting stock that could be used to fund acquisitions and other expenses. IBM exceeded third-quarter expectations on accelerated software and infrastructure revenue growth and raised its full-year free cash flow outlook, with Reuters reporting the software revenue guidance lift came as deal delays eased. Nokia’s quarter was blunt about where the money is coming from: group net sales to AI and cloud customers surged 123% year-over-year, IP Networks grew 43%, and third-quarter profit beat analyst estimates as data-center demand drove Network Infrastructure sales up 15% on a constant currency basis.
Austin: The $145B top end hasn’t moved since April. This just lifted the low end to $135B. Still an enormous number.
Vik: What’s up with the new nonvoting stock class?
Lam Research Earnings and ASML Shortage Warning Tighten Equipment Supply Outlook
Lam Research posted its quarterly results for the period ended September 27, 2026, adding fresh evidence that wafer fabrication equipment demand is holding firm even as the industry absorbs a blunter warning from ASML, which told customers to expect equipment shortages through 2028 as AI-driven orders outpace production capacity. SEMI projects global silicon wafer shipments will hit 14,420 million square inches in 2026, an 11.8% annual increase, with successive record highs forecast from 2027 through 2029. The convergence of strong Lam revenues, a multi-year ASML supply constraint, and rising wafer demand creates a queue that existing backlog holders sit at the front of, while fabs chasing faster build-outs have fewer good options.
Huawei Secures $2.3 Billion Offshore Loan as DeepSeek Raises $15 Billion
Huawei closed a $2.3 billion offshore loan as chip spending at the company climbs, drawing on international credit markets even as US export controls squeeze its access to advanced semiconductors. DeepSeek, the Chinese AI lab whose low-cost models rattled US competitors earlier this year, separately raised $15 billion to take on the top American AI labs, according to Bloomberg. SemiAnalysis, covering what it calls the Huawei-DeepSeek symbiosis, reports the two organizations are already co-optimizing across hardware and models, with Huawei’s Ascend chips and DeepSeek’s training work converging into something closer to a vertically integrated stack. The combined capital raise exceeds $17 billion, all of it aimed at building AI infrastructure that doesn’t depend on US suppliers.
OpenAI Brings Top Models to Google Cloud as AMD and Intel Unite on x86 AI
OpenAI will sell its most advanced models directly through Google Cloud, a deal that puts Microsoft’s longtime Azure exclusivity arrangement under genuine pressure. The companies announced the partnership without disclosing financial terms, but the move gives enterprise buyers a second major hyperscaler route to OpenAI’s frontier models. On the same day, AMD and Intel announced they are setting aside decades of rivalry to collaborate on x86 AI development, consolidating their CPU optimization work into the oneAPI Deep Neural Network Library within PyTorch as a direct answer to Nvidia’s CUDA grip on the developer ecosystem. Two chip competitors agreeing to share infrastructure code, just to fight a common lock-in threat, is a rare thing in this industry.
Austin: Microsoft will always be remembered for betting on OpenAI first, but seems all that first-mover juice was squeezed.
Besi Orders Surge 96% as Hybrid Bonding Demand Outpaces Supply Concerns
Besi reported a 96% surge in quarterly orders, driven by AI data center customers snapping up advanced-packaging equipment faster than the Dutch company can comfortably track. CEO Richard Blickman credited momentum in hybrid bonding, the chip-stacking technique that lets processors and memory sit closer together without traditional solder bumps, as a primary growth engine. ASML has signaled ambitions in the bonding equipment space, but Blickman flatly dismissed the lithography giant as a direct hybrid-bonding rival, arguing the two companies occupy fundamentally different positions in the packaging workflow. Orders beat analyst expectations, and Besi said backlog conversion into revenue remains on track for the quarters ahead.
Sector Watch
Foundry & Packaging
Nvidia locks in capacity deals with six packaging and substrate partners including Ibiden, targeting a 2.5x expansion of substrate supply. (Unite.AI)
Intel completes first major structural truss installation at Ohio One fab and reports Arizona Fab 62 tooling progress, signaling steady 18A ramp. (ServeTheHome)
Memory
Micron rules out 4F² DRAM architecture despite CXMT’s advance, with CTO confirming the company will not pursue that design path. (Nikkei Asia)
Groq is evaluating a shift from SRAM to 3D-stacked DRAM for its next-generation LPU, a move that would reshape its memory-bandwidth architecture. (TrendForce)
Hanmi Semiconductor wins Micron’s top supplier award for a second consecutive year, deepening its TC-bonding equipment role in HBM production. (Chosunbiz)
Power
Infineon and Nvidia jointly advance GaN-based power conversion for 800 VDC AI data-center infrastructure, publishing a joint reference design. (digitimes)
Navitas unveils a next-generation ultra-high-voltage SiC power module targeting high-density AI and industrial power-delivery applications. (Semiconductor Digest)
STMicroelectronics expands its MasterGaN7 integrated GaN transistor-driver family with new 650V devices for high-efficiency power stages. (Semiconductor Digest)
Optics & Networking
CoolIT Systems and Stäubli extend direct liquid cooling to high-power pluggable optical transceivers used in AI data-center racks. (EE News Europe)
Nokia advances AI-RAN toward commercial deployment and launches a sovereign agentic AI platform for telco autonomy built on Nvidia infrastructure. (Nokia)
PCB & Components
LG Innotek expands its AI component partnership with Nvidia, joining Samsung Electro-Mechanics in supplying parts for AI chips and robotics platforms. (The Korea Times)
Unimicron warns of double-booking risk as AI PCB capacity demand outstrips supply, with panel-scale expansion timelines under pressure. (digitimes)
Taesung showcases FC-BGA and glass-substrate processing equipment at TPCA Show 2026 in Taipei, targeting next-generation AI substrate manufacturing. (The Elec)
Compute & Edge
Renesas launches RX74M and RX74N 32-bit MCUs, the first RX-family devices built on a 22nm process and featuring embedded MRAM. (Renesas)
Cadence opens a sovereign AI data center in Munich dedicated to chip design workloads, targeting EU customers requiring local data residency. (StreetInsider)
Texas Instruments issues an above-consensus revenue forecast after demand picks up across its analog and embedded-processing product lines. (Bloomberg.com)
Data Centers
Crusoe plans the Spark Liquid, a 3.5 MW modular liquid-cooled AI factory unit supporting Nvidia DSX POD, and signs a 40 MW Norway deal with Polar. (Crusoe)
Corning approves a $585 million expansion of its Wilmington facility, adding 414 jobs to scale optical-fiber output for data-center demand. (Corning)
Nebius becomes the first customer for HUMAIN and DataVolt’s planned 50 MW AI giga campus at NEOM’s Oxagon, formalizing a framework agreement. (NEOM)
Policy & Trade
Fujifilm is scaling EUV photoresist mass production at its Pyeongtaek facility, deepening its supply partnership with Samsung Electronics. (The Elec)
NY Creates commits to high-NA EUV access for its semiconductor research ecosystem, backing domestic advanced-lithography capability. (Electronics360)
Canon plans to boost lithography equipment production capacity in 2027 to address sustained demand from advanced-node fabs globally. (nippon.com)
Tell us how to make Semi Doped better! Give us your feedback! ❤️