AI hardware demand collides with financing pressure and lengthening supply queues

AI infrastructure spending kept expanding, but financing costs, component shortages and construction restrictions exposed constraints beyond accelerator supply. Google TPUs entered production at Meta, while Google Cloud set a November launch for NVIDIA Vera Rubin instances.

Key developments

Meta’s commitments reach $825 billion

Meta’s future lease and non-cancelable contract commitments climbed to about $825 billion as free cash flow turned negative in Q3 2026. The company raised the lower end of its 2026 capital spending forecast by $5 billion to $135 billion.

Infrastructure procurement is absorbing cash even as Meta increases its spending floor.

Oct 25 · DIGITIMES

Anthropic seeks costly TPU financing

Anthropic is reportedly offering Blackstone a total return above 10% under a proposed arrangement to finance $60 billion of Google TPU purchases.

Financing charges would add to the cost of Anthropic’s compute expansion beyond the hardware itself.

Oct 19 · DIGITIMES

Google TPUs enter Meta production

Google TPUs are serving production AI workloads at Meta. New open-source engines allow PyTorch models to run natively on the chips.

Native PyTorch support lowers a software barrier to using an alternative to NVIDIA accelerators.

Oct 22 · Tom's Hardware

Vera Rubin gets a cloud launch window

Google Cloud scheduled general availability of A5X instances for November, combining NVIDIA Vera Rubin NVL72, ConnectX-9 SuperNICs, Google’s Virgo Network and liquid cooling. European expansion is also planned.

Cloud customers now have a deployment window for the new training, inference and reinforcement-learning infrastructure.

Oct 21 · NVIDIA Newsroom

Chinese buyers reportedly order $35 billion in GPUs

Tencent, Alibaba and ByteDance reportedly placed $35 billion in orders for NVIDIA RTX Pro 5500 workstation GPUs. Demand reportedly far exceeds NVIDIA’s planned production.

The gap between orders and planned output puts availability of this workstation GPU under pressure.

Oct 20 · Tom's Hardware

Component queues stretch beyond a year

Samsung Electro-Mechanics’ MLCC lead times reached 61 weeks as AI server demand strengthened its pricing power, with shortages extending into consumer, automotive and industrial products. Separately, HDD shortage risk is expected to extend into 2028, while fourth-quarter enterprise SSD price negotiations are beginning.

AI infrastructure demand is exposing buyers across several industries to longer waits and pricing pressure for supporting components.

Oct 19–23 · DIGITIMES · TrendForce

Texas suspends data-center permits

Texas suspended 17 data-center construction permits and shut down the automated portal used to approve projects.

Affected projects face scheduling uncertainty before construction can proceed.

Oct 24 · Tom's Hardware

Samsung reallocates capacity toward HBM

Samsung is moving its next-generation M2800 mobile modem to a 2nm process. The freed 4nm capacity will primarily make HBM base dies, which handle control functions at the bottom of memory stacks.

The shift gives AI memory components more manufacturing capacity previously allocated to mobile modems.

Oct 21 · DIGITIMES

Mature-node pricing strengthens

PSMC reported third-quarter revenue of NT$21.3 billion, up 23% quarter-over-quarter, and net profit of NT$6.5 billion. Higher prices and utilization lifted gross margin from 28% to 48%, and the company targets a further 20% price increase in 2027.

Customers buying mature-node memory and power-management chips face further procurement cost pressure.

Oct 22 · DIGITIMES

Taylor’s 2nm capacity target takes shape

Samsung reportedly ordered equipment to fill the remaining space at its first Taylor, Texas, foundry fab. The additions would bring 2nm capacity to 20,000 wafers per month in the second half of 2027.

The orders establish a reported scale and delivery window for additional US advanced-node manufacturing capacity.

Oct 23 · DIGITIMES

Also this week

Compute platforms

• Arm claimed its AGI CPU delivers nearly twice AMD EPYC 9755’s modeled rack performance across seven agentic AI workloads using a methodology developed with Fujitsu, Meta and OpenAI. Tom's Hardware

• Modal is adopting NVIDIA Vera CPUs after its internal AI-agent workload tests showed twice the throughput of Intel and AMD chips. Tom's Hardware

• Samsung is developing Exynos 2700 for broader Galaxy S27 deployment in 2027, while Huawei’s next-generation Kirin chips have entered testing. DIGITIMES

• Qualcomm will begin performance-based layoffs worldwide in November as it shifts resources toward AI infrastructure, data centers, automotive, PCs and edge computing. Tom's Hardware

• AMD and Intel are collaborating to consolidate PyTorch CPU inference optimizations under the oneDNN library with a single code path. Tom's Hardware

• Arm characterized a $2 nominal damages award in its Qualcomm trial as validation, while chip-design experts cited open architectures such as RISC-V as a longer-term option for AI workload flexibility. EE Times

• AMD director Travis Kirsch confirmed that Lisa Su initiated the Ryzen 7 5800X3D relaunch plan before DDR5 shortages. Tom's Hardware

Memory roadmaps

• Micron is ending future 4F² DRAM development after two decades of research, favoring advanced process technologies for further bit-density gains. Tom's Hardware

• Micron’s next-generation G11 NAND promises 42 Gb/mm² density and a 5,400 MT/s interface using more than 400 layers. Tom's Hardware

• Samsung and SK hynix are reportedly upgrading their Chinese memory factories to produce NAND flash with more than 200 layers. DIGITIMES · Tom's Hardware

• Solidigm ended its review of a potential IPO while maintaining plans to build a US 3D NAND fab. SK hynix Newsroom · Tom's Hardware

• Jensen Huang said NVIDIA does not expect memory supply to exceed demand in the coming years and remains open to working with CXMT. DIGITIMES

Factory investment

• Bosch began 200mm silicon carbide sample production at its Roseville facility as part of an investment of up to $2 billion to convert the former TSI Semiconductors fab. Semiconductor Digest

• Aixtron said preliminary third-quarter order intake, operating profit and operating margin significantly exceeded market expectations. lSemiconductor Today News

• NVIDIA and supply-chain partners announced plans to expand global manufacturing capacity for accelerated computing and full-stack AI factory platforms. NVIDIA Newsroom

• Intel is installing production tools at Arizona’s Fab 62, while Ohio One reached its first major structural truss installation. ServeTheHome · Tom's Hardware

• Texas Instruments reported strong third-quarter revenue growth across industrial, automotive and data-center businesses and described chip supply as tight while expanding production and testing capacity. DIGITIMES

• SEMI forecast 11.8% growth in global silicon wafer shipments to 14,420 million square inches in 2026, with record shipment levels expected through 2029. Semiconductor Digest

• The US CHIPS Research and Development Office announced letters of intent with four companies for $389 million in awards, including up to $73 million for Vertical Semiconductor. lSemiconductor Today News · Semiconductor Digest

• Applied Materials presented financial projections supporting its $5 billion EPIC Center investment. EE Times

• HANMI Semiconductor opened HANMI USA to expand its presence in the US semiconductor market. Semiconductor Digest

• Chipmakers, equipment suppliers and EDA companies are adopting NVIDIA CUDA-X libraries and AI tools to connect simulation, AI physics, agents and digital twins across manufacturing. Semiconductor Digest

Packaging and materials

• TSMC’s 104.65-hectare Longtan Science Park Phase 3 expansion restarted after development-plan approval, with advanced packaging such as CoPoS expected to lead its initial focus. DIGITIMES

• NVIDIA unveiled an Optics-on-Interposer prototype that moves optical engines onto the silicon interposer carrying compute chips to increase bandwidth density and reduce power consumption. DIGITIMES

• ACM Research’s panel-level packaging equipment is undergoing customer evaluations in Taiwan as the company invests in South Korea and Oregon. DIGITIMES

• Tongtai Machine & Tool reported PCB drilling equipment order visibility through the end of 2027, linking demand to AI infrastructure’s high-end board requirements. DIGITIMES

• Unimicron Chairman Shan-Chieh Chien projected that substrate sizes could reach 14 times their current scale by 2030 as AI accelerators require more advanced packaging and heterogeneous integration. DIGITIMES

• Photon Bridge raised €10 million to address laser production for optical connections in expanding AI processor clusters. lSemiconductor Today News

• Atomera and IQE signed a joint development agreement for MST-enabled RF GaN-on-silicon wafers, which IQE will manufacture for foundry and integrated device manufacturer customers. lSemiconductor Today News · Semiconductor Digest

• Gigaphoton is extending its laser technology into packaging micro-via drilling, actinic mask inspection and 3D metrology while developing on-site neon recycling. Semiconductor Digest

Data-center systems

• Chief Telecom broke ground on a roughly NT$3.3 billion AI data center at Central Taiwan Science Park, with operations expected in the second half of 2028. DIGITIMES

• NVIDIA introduced containerized DSX PODs and reference designs using closed-loop liquid cooling and dry coolers for near-zero facility cooling water consumption. NVIDIA Newsroom · ServeTheHome · The Next Platform

• Infineon expanded its NVIDIA collaboration on 800-volt direct-current data-center power architectures, focusing on GaN chips for high-voltage intermediate bus conversion. DIGITIMES · lSemiconductor Today News

• NVIDIA and PreussenElektra are studying the reuse of decommissioned German nuclear plant sites as AI data centers with existing grid connections and industrial infrastructure. Tom's Hardware

• NVIDIA’s BlueField-4 STX context-memory server, using a single Vera platform, was demonstrated at OCP 2026. ServeTheHome

• NVIDIA launched Minerva, an AIOps platform combining infrastructure and workload telemetry with AI models to help diagnose AI factory problems. NVIDIA Newsroom

• Leaked footage reportedly shows Chinese factories dismantling thousands of consumer GPUs, including RTX 5090 Founders Editions, for AI servers. Tom's Hardware

• Jensen Huang is reportedly considering enterprise AI software fees of around $2,000 per GPU annually, with pricing and launch timing unconfirmed. Tom's Hardware

Design market

• Electronic system design industry revenue rose 12.2% year-over-year to $5,710.8 million in the second quarter of 2026. Semiconductor Digest

What to watch

• Watch for Google Cloud’s A5X general availability in November, the scheduled release window for its Vera Rubin instances.

• Watch fourth-quarter 2026 enterprise SSD price negotiations; increases above 30% are proving difficult to secure.

• Watch whether Samsung’s reported Taylor equipment additions deliver 20,000 wafers per month of 2nm capacity in the second half of 2027.

• Watch for a decision on Anthropic’s proposed $60 billion TPU financing arrangement; timing has not been announced.