Memory shortages deepen as Rubin ships and AI hardware expansion accelerates

AI hardware demand ran further ahead of supply this week: HBM requests exceeded twice SK hynix’s available volume, while additional memory supply is not expected until 2028. NVIDIA’s first Rubin production hardware began shipping as cloud providers expanded accelerator commitments and shortages spread across packaging, substrates and circuit boards.

Key developments

Memory relief slips to 2028

SK hynix said customers are seeking more than twice the HBM volume it can currently supply, while reporting third-quarter revenue of KRW 104.6155 trillion and operating profit of KRW 83.3338 trillion. Samsung reported revenue of KRW 195.3 trillion and operating profit of KRW 107.4 trillion, with memory gains driven mainly by higher prices. Both companies are making record equipment investments, but additional memory supply is not expected until 2028.

Accelerator buyers face prolonged competition for HBM and continued memory cost pressure despite suppliers’ increased investment.

Oct 28–29 · SK hynix Newsroom · Samsung Semiconductor Global Newsroom · EE Times

Rubin moves into production

NVIDIA began shipping its first Rubin production hardware and updated platform specifications. NVIDIA and Oracle also announced plans for multi-gigawatt OCI Superclusters using Vera Rubin NVL72 systems, while standalone Vera CPUs became generally available on OCI Superclusters.

Oracle customers can now access Vera independently of integrated GPU systems as Rubin moves from planned infrastructure toward deployment.

Oct 27–Nov 1 · NVIDIA Newsroom · ServeTheHome · Tom's Hardware

Intel prioritizes servers over PCs

Intel’s third-quarter results showed manufacturing bottlenecks and rising memory costs weighing on PC sales despite strong data-center growth. The company is shifting production from consumer to server CPUs and expanding chipmaking and advanced packaging capacity in the US, Ireland and Malaysia.

Prioritizing server processors could further constrain PC processor availability while Intel works through manufacturing bottlenecks.

Oct 30–31 · Intel Corporation (INTC) Press Releases · Tom's Hardware · Tom's Hardware

Packaging costs face upward pressure

TSMC is reportedly considering CoWoS price increases of up to 20%, with equipment deliveries taking up to eight months and separate wafer-price increases of 5%–10% expected in 2027. NVIDIA is moving to secure scarce ABF substrate capacity, while Ibiden announced another JPY151 billion investment in substrates, including capacity for Intel expanding from 2029.

Accelerator production faces higher packaging costs and limited near-term relief from new substrate capacity.

Oct 26–31 · TrendForce · Tom's Hardware · DIGITIMES

PCB shortages reach device makers

AI demand is reportedly diverting PCB manufacturing resources from PC and device makers, with order books full until 2030. Lead times are stretching to 30 months and material costs are rising twelvefold.

PC and device deliveries can be delayed by board shortages even when processors are available.

Oct 29 · Tom's Hardware

Anthropic discloses compute commitments

Anthropic’s IPO filing reveals hundreds of billions of dollars in compute-capacity commitments spanning Google, Amazon, Microsoft, Broadcom and GPU infrastructure providers. Its buildout encompasses TPUs, GPUs and dedicated data centers.

The commitments spread AI infrastructure demand across custom accelerators, GPUs and multiple cloud suppliers.

Oct 27 · DIGITIMES

Arm enters finished server CPUs

Arm is expanding beyond Neoverse IP and compute-subsystem licensing into finished server CPUs for AGI applications. Meta is a co-developer and primary customer, and OpenAI also plans to adopt the products.

Server buyers gain an alternative to developing their own Arm-based processors.

Oct 30 · DIGITIMES

Oracle broadens accelerator suppliers

Oracle signed partnerships involving Cerebras, Google and SambaNova to expand the chips hosted in its cloud infrastructure. The deals accompany its purchase of NVIDIA Vera Rubin platforms.

Oracle cloud customers gain more accelerator choices beyond NVIDIA hardware.

Oct 28 · The Next Platform · DIGITIMES

Maia 300 hardware takes shape

Microsoft showed its Maia 300 platform at OCP Summit 2026 with four accelerators per tray. Each accelerator has 288GB of HBM4 and a 950W power rating.

The configuration establishes concrete memory and power requirements for systems built around Microsoft’s next accelerator.

Oct 28 · ServeTheHome

ASE increases capacity investment

ASE Technology Holding raised its 2026 capital expenditure budget by $1 billion to $11.5 billion. Third-quarter net profit reached NT$26.996 billion, up 28% sequentially and 148% year over year.

The higher budget supports expansion of packaging and assembly capacity needed by chip suppliers.

Oct 29 · DIGITIMES

Also this week

Manufacturing expansion

• Hermes Testing Solutions approved a new Hsinchu R&D centre and a manufacturing centre in south-central Taiwan, with total investment expected to exceed NT$7 billion, approximately US$215 million. DIGITIMES

• Edwards Vacuum began low-volume production at its new 250,000-square-foot New York facility, ending its sole reliance on South Korea for worldwide dry vacuum pump supply. Semiconductor Digest

• Infineon plans to invest about €1.1 billion by 2034 in existing Villach cleanrooms to establish what it describes as Europe’s first high-volume 300mm GaN power-semiconductor facility. lSemiconductor Today News

• Solidigm is reportedly discussing a potential $45 billion investment in a new US flash memory factory. Tom's Hardware

• UMC’s utilization exceeded 90% in the second half of 2026 as AI, high-performance computing, communications and automotive demand strengthened its mature and specialty process business. DIGITIMES

• AUO approved the sale of its Taichung Plant Phase II facilities and infrastructure to TSMC for NT$38.8 billion excluding taxes, and separately terminated its planned Huaya facility sale to Quanta over usage specifications and delivery schedules. DIGITIMES

• Global wafer fabrication equipment spending is forecast to exceed $200 billion in 2027, driven by investment in advanced logic, DRAM and HBM. DIGITIMES

• The 2031 metrology and inspection market forecast rose from $24.7 billion to $31.4 billion, largely because of AI-driven memory demand. Semiconductor Digest

Design and processes

• Cadence announced its Maestro AI Model Portfolio for semiconductor and system-design agents, with early access for select customers planned for the fourth quarter of 2026. The Next Platform · Semiconductor Digest · Tom's Hardware

• Micron plans to use wafer bonding and High-NA EUV for next-generation DRAM after rejecting future 4F² DRAM development. Tom's Hardware

• SK hynix plans to extend its 6F² DRAM architecture through EUV lithography and process improvements while retaining 4F² and 3D DRAM on its roadmap. DIGITIMES

• Lam expanded its collaboration with imec to cover advanced 3D integration and Aether dry-resist development for future High-NA EUV patterning. Semiconductor Digest

Chip businesses

• Intel announced plans to acquire Classiq to create a full-stack quantum computing platform. Intel Corporation (INTC) Press Releases

• Altera announced its Agilex 2X series, led by Agilex 27, with LPDDR6 and 1.6TbE support coming in 2027. ServeTheHome

• SK hynix appointed its AI memory development chief to lead a new US-based Americas unit as the industry moves toward customized HBM. DIGITIMES

• MediaTek reported third-quarter revenue of NT$168.06 billion, up 18.3% year over year, and projected fourth-quarter sequential revenue growth of up to 32% as AI accelerator ASIC shipments accelerate. DIGITIMES

• Renesas expects 2026 revenue of JPY1.6668 trillion and a 33.2% operating margin, citing strong AI data-center, industrial and automotive demand. DIGITIMES

• An internal Intel driver publication and a shipping manifest point to Raptor Lake Next desktop processors approaching availability. Tom's Hardware

• Intel CEO Lip-Bu Tan reportedly discussed possible cooperation with AMD CEO Lisa Su as Intel rebuilds its CPU business and expands foundry operations. Tom's Hardware

• RTX Spark laptops became available from ASUS, Dell, HP, Lenovo, Microsoft and MSI, combining a Blackwell RTX GPU and Grace CPU with up to 128GB of unified memory. NVIDIA Newsroom

• Wolfspeed commercially introduced transparent silicon carbide technology targeting augmented reality, quantum computing and advanced optical applications. lSemiconductor Today News

• The open-source UEFI tool nv-uefi-unlocker lifts NVIDIA’s Windows tensor-rate cap on RTX 20, 30 and 40-series desktop and laptop GPUs, with reported FP16 gains of up to 65%. Tom's Hardware

Infrastructure demand

• AMD and Meta expanded their strategic partnership to support Meta’s superintelligence infrastructure. Advanced Micro Devices, Inc. (AMD) Press

• Global server CPU shipments are forecast to rise 45.7% to 52.6 million units in 2027, supported by agentic AI, AI server expansion, cloud providers’ in-house CPUs and server replacement demand. DIGITIMES

• Google and Foxconn are working to shorten the interval from chip supply to data-center deployment from 12–18 months to about six months. DIGITIMES

• OpenAI CEO Sam Altman met around 50 Samsung executives in South Korea on October 31 to discuss AI and business opportunities while seeking more data-center RAM. Tom's Hardware

• GitHub added four million CPU cores and 160 petabytes of storage to handle a sixfold year-over-year increase in new code associated with agentic AI. Tom's Hardware

• Google said AI agents create bursty workloads that frequently leave CPUs idle, changing requirements for data-center CPU infrastructure. EE Times

Emerging hardware

• Palomino Laboratories received its first commercial purchase order from a strategic hyperscaler customer for GaN micro-LED chips used in AI optical interconnects. lSemiconductor Today News

• PsiQuantum demonstrated an 8×8 laboratory optical switch with typical loss of 1.2 dB against a data-center allowance of up to 3 dB, and plans to develop larger devices. Tom's Hardware

• Nagare Labs emerged from stealth with $5 million from Playground Global to develop a new memory tier targeting AI’s memory bottleneck. Semiconductor Digest

What to watch

• Watch for select-customer early access to Cadence’s Maestro AI models in the fourth quarter of 2026.

• Watch for Altera’s LPDDR6 and 1.6TbE features in the Agilex 2X series during 2027.

• Watch whether Solidigm’s potential $45 billion US flash factory investment becomes a confirmed project; decision timing has not been announced.