AI hardware expansion brings new rack designs and up to $21 billion in fabs

Microsoft and AMD introduced new rack-scale AI platforms, while GlobalFoundries outlined two new fabs backed by up to $21 billion in investment. Manufacturing expansion remained constrained by equipment supply, with ASML reporting record revenue and memory makers weighing competing uses of capacity.

Key developments

Microsoft makes Maia its inference foundation

Microsoft unveiled the TSMC N2P-based Maia 300, with production deployments planned for 2027 and 96 accelerators per rack. It plans multi-gigawatt deployments and will make Maia its primary inference platform for its own AI services, while continuing to use NVIDIA and AMD GPUs.

Microsoft’s own AI services will shift more inference work onto custom silicon, changing the hardware mix in its data centers.

Oct 12 · DIGITIMES · DIGITIMES · DIGITIMES

GlobalFoundries adds two-fab plan

GlobalFoundries will operate two new semiconductor fabs involving up to $21 billion in capital investment. The facilities will be jointly owned by the U.S. and Japan and produce foundational semiconductors.

The planned facilities would add manufacturing capacity for foundational chips and expand GlobalFoundries’ operating footprint.

Oct 14 · Semiconductor Latest News | SIA · Semiconductor Digest

ASML growth meets an optics constraint

ASML reported a record $13 billion quarter and pledged to expand production capacity as customers order advanced tools years ahead. Its EUV expansion depends partly on Zeiss increasing output of qualified precision optics.

Precision-optics supply limits how quickly chipmakers can receive the EUV tools needed for additional advanced-node capacity.

Oct 14–15 · Tom's Hardware · EE Times · Tom's Hardware

AMD takes CPUs to rack scale

AMD introduced Helion, a rack-scale reference design for agentic AI supporting up to 32,768 Zen 6 Venice CPU cores. CPU-only and all-in-one configurations are scheduled to ship in 2027.

Server vendors gain a reference architecture for building CPU-heavy agentic AI systems at rack scale.

Oct 12 · Advanced Micro Devices, Inc. (AMD) Press · Tom's Hardware

NVIDIA opens Vera Rubin architecture

NVIDIA announced partnerships and contributions to open its Vera Rubin architecture at the OCP Global Summit. It also announced plans with Cisco for USFP, a high-density pluggable optical module supporting 400Gb/s per lane.

Hardware partners gain more ways to participate in NVIDIA infrastructure, including higher-bandwidth optical connectivity.

Oct 12 · NVIDIA Newsroom · NVIDIA Newsroom

Memory commitments tighten expansion choices

Samsung reportedly committed 75–80% of its 2027 memory supply, including HBM, to long-term agreements; Samsung has not confirmed the allocations or named buyers. Separately, some SK hynix HBM4 back-end equipment installations may slip into 2027 as rising conventional DRAM prices create competing investment opportunities.

Buyers seeking additional AI memory face both limited unallocated supply and uncertainty over new HBM4 capacity.

Oct 12–14 · DIGITIMES · DIGITIMES

Samsung advances Taylor while Austin contracts

Samsung plans a November 2026 groundbreaking for its second Taylor fab, while the first prepares to begin operations by year-end and ramp 2nm volume production in 2027. Yield engineers are reportedly being sent from South Korea; meanwhile, Austin’s semiconductor workforce fell 32% in 2025 and local operating spending fell nearly 75%.

Samsung’s Texas manufacturing activity is shifting toward advanced-node production at Taylor rather than growing uniformly across its sites.

Oct 13–16 · DIGITIMES · DIGITIMES · DIGITIMES

Japan approves TSMC’s 3nm upgrade

Japan approved TSMC’s revised plan for its second Kumamoto fab. The facility will produce 3nm logic chips instead of the 6nm and 12nm processes approved in February 2024.

Customers will have a more advanced planned manufacturing option within TSMC’s Japanese footprint.

Oct 17 · DIGITIMES

RTX Spark laptops reach buyers

NVIDIA introduced RTX Spark, and laptops became available from ASUS, Dell, HP, Lenovo, Microsoft and MSI starting at $2,399. The Windows systems combine the GB10 SoC, Blackwell RTX graphics and up to 128GB of unified memory; NVIDIA states up to 1 petaflop of AI compute.

Buyers can now purchase a local AI platform with substantial shared memory alongside graphics capabilities.

Oct 13–17 · NVIDIA Newsroom · NVIDIA Newsroom · ServeTheHome

DRAM repricing lifts Nanya margins

Nanya reported record third-quarter 2026 earnings in unaudited results, with gross margin reaching 85.7%. DRAM average selling prices rose more than 60% sequentially as low-priced long-term agreements expired.

Expiring contracts are exposing memory customers to higher prices while increasing producer profitability.

Oct 12 · DIGITIMES

Also this week

Manufacturing investment

• Applied Materials unveiled its $5 billion EPIC Center in Silicon Valley to bring chipmakers, equipment suppliers and researchers together for process development and commercialization. EE Times · Semiconductor Digest

• SEMI forecast global 300mm front-end fab equipment spending of $149 billion in 2026, up 31%, rising to $227 billion in 2029. Semiconductor Digest

• Samsung affiliates planned KRW12.26 trillion, approximately $9.18 billion, in semiconductor investments across South Korea and Vietnam, with funding dependencies and equipment hurdles affecting parts of the expansion. DIGITIMES

• TSMC confirmed it is considering U.S. capacity expansion beyond Arizona, including Texas, and said participation in Elon Musk’s Terafab project would require it to operate the entire fab. Tom's Hardware

• GlobalFoundries and the CHIPS Program Office amended an award to reallocate up to $955 million in incentives toward $5 billion in investment in Malta, New York, and Burlington, Vermont. lSemiconductor Today News

• Infineon and Wolfspeed signed a multi-year agreement moving their silicon carbide substrate supply arrangement from 150mm to 200mm wafers. lSemiconductor Today News

• Taiwan Mask agreed to acquire Japanese equipment-services provider TCK for $4.75 million, extending into equipment assembly and engineering support in Kyushu. DIGITIMES

• Enterprise Singapore positioned local semiconductor SMEs for larger roles in high-value manufacturing, advanced packaging and silicon photonics. DIGITIMES

Process and packaging

• ASE posted third-quarter revenue of NT$244.161 billion, up 27.79% sequentially and 44.84% year-over-year, as AI-related assembly and testing demand helped September set a third consecutive monthly record. DIGITIMES

• NVIDIA announced semiconductor adoption of CUDA-X libraries, including cuPotential, while Siemens reported up to 12× acceleration of computational lithography workflows with NVIDIA. NVIDIA Newsroom · Semiconductor Digest · Semiconductor Digest

• Rapidus launched its CORE ecosystem for 2nm development, starting with Design Solution Associates supporting customers’ early-stage designs. Semiconductor Digest

• ZEISS SMT expanded into advanced-packaging process control with DUNE for individual wafer-bow correction and NLX for inline X-ray inspection. Semiconductor Digest

• Advantest integrated its V93000 platform with FormFactor’s TRITON prober and introduced a turnkey optical test rack for high-volume silicon photonics manufacturing. Semiconductor Digest

• Langdao Silex demonstrated a laser-produced plasma EUV source, although useful power levels likely remain years away and a practical lithography tool requires further advances. Tom's Hardware

Chips and memory

• Intel named its Crescent Island inference accelerators GPU D-Series, specifying 1.3 PFLOPS of dense FP8 compute for the D270 and up to 480GB of LPDDR5X memory with 1.5TB/s bandwidth across the family. ServeTheHome · Tom's Hardware

• CXMT president Cao Kanyu said its first 4F² DRAM product will be a server DDR5 RDIMM expected by the end of 2026. Tom's Hardware

• YMTC began developing 4F² DRAM targeting 9,600 MT/s while pursuing a CNY33 billion Shanghai initial public offering. DIGITIMES

• Samsung explored stacking HBM directly above graphics processors, while SK hynix advanced high-bandwidth flash to increase AI inference memory capacity. DIGITIMES

• SK hynix detailed H-TPU in-house processor cores that will offload routine I/O work from Arm CPUs in its next enterprise SSD controllers. Tom's Hardware

• NVIDIA reportedly said the RTX 5090 remains in production despite constrained supply, severe shortages and removed retailer listings. Tom's Hardware

Optical connectivity

• Tower Semiconductor and Semtech expanded their silicon-germanium manufacturing partnership through a multi-year agreement adding dedicated capacity for optical and copper AI interconnects. Semiconductor Digest

• Eliyan introduced Nexus and Oscar, its first NuLight optical interconnect chiplets, for high-bandwidth, extended-reach and low-power AI connectivity. Semiconductor Digest

• Wiwynn, Global Unichip and Ayar Labs prepared their first live demonstration of an integrated compute system with real-time telemetry at the 2026 OCP Global Summit. DIGITIMES

• Lumentum, Astera Labs and Corning demonstrated 1060nm VCSEL-based PCIe-over-optics connectivity for AI rack-scale architectures at the OCP Global Summit. lSemiconductor Today News

• Gemtek Technology and Starman New Photonics announced a U.S. joint venture to manufacture advanced optical transceivers for hyperscale AI data-center interconnects. DIGITIMES

Infrastructure components

• NVIDIA showed its first PTFE switch board sample for the MGX server ecosystem and named seven supply-chain partners, including Flexium Interconnect and Taiflex Scientific. DIGITIMES

• TrendForce forecast 43% growth in server CPU shipments in 2027, highlighting their role in AI inference. TrendForce

• Power Integrations and Ideal Power began collaborating on solid-state circuit breakers designed for the 800VDC transition in AI infrastructure. lSemiconductor Today News

• Saras Micro Devices began sampling a 2 mm capacitor for embedding in a PCB or package core to support vertical power delivery for AI chips. Semiconductor Digest

Legal and deployment

• A federal jury ruled for Qualcomm on an architecture breach claim against Arm but awarded $2, sided with Arm on some issues and left the legal battle continuing. Tom's Hardware

• An analysis mapping more than 300 local U.S. data-center moratoriums against the project pipeline found that only three projects had actually been delayed. Tom's Hardware

What to watch

• Watch for Samsung’s second Taylor fab groundbreaking in November 2026 and the first facility’s planned start of operations by year-end.

• Watch for CXMT’s first 4F² server DDR5 RDIMM, expected by the end of 2026.

• Watch whether SK hynix’s HBM4 back-end equipment installations move from the second half of 2026 into 2027.

• Watch for AMD Helion shipments and Microsoft Maia 300 production deployments in 2027.