Compute & Infrastructure

Nvidia revealed at Computex 2026 that Anthropic, OpenAI, and SpaceX are committed early adopters of its Vera CPU, the Arm-based host processor designed to pair with its Blackwell and next-generation GPU architectures in the NVLink-connected Grace Blackwell successor platforms. This is strategically significant beyond the headline: Vera represents Nvidia's push to own the full compute stack inside AI data centres, not just the GPU. By securing the CPU socket alongside the GPU socket, Nvidia reduces the surface area available to Intel (Xeon) and AMD (EPYC) and tightens its integration story for customers who want optimised NVLink bandwidth between host and accelerator. Bloomberg reports the chip is described as 'upcoming,' meaning these are design-win commitments, not volume shipments — a distinction that matters for capacity planning.

The customer list itself is analytically revealing. Anthropic just raised $65 billion with memory chipmakers Micron, Samsung, and SK Hynix as investors, per Data Centre Dynamics, suggesting it is building out serious proprietary infrastructure rather than relying solely on cloud providers. OpenAI's presence confirms that even as it pursues custom silicon partnerships, it is hedging with Nvidia's latest platforms. The convergence of frontier model labs into direct hardware customers — rather than consuming compute through cloud intermediaries — is a structural shift that accelerates demand for discrete infrastructure and reduces the buffer that hyperscale cloud margins previously provided.

Nvidia announced the RTX Spark Superchip platform at Computex — an Arm-based SoC combining a 20-core CPU with a 6,144-CUDA-core Blackwell GPU and up to 128GB of unified memory, targeting Windows laptops and desktops. Tom's Hardware reports the platform is branded around enabling 'agentic AI' locally on Windows. Jensen Huang also revealed a multi-generation roadmap — RTX Spark followed by Rubin (with LPDDR6 memory) and Rosa Feynman — signalling this is a committed product line, not a one-off venture. Leaked specs confirmed by Tom's Hardware show multiple SKU tiers targeting different price points.

From a supply chain perspective, this is a significant new demand vector on TSMC's advanced node capacity and on LPDDR memory from SK Hynix and Samsung. The 128GB unified memory spec is particularly notable — it matches Apple M-series positioning and signals that local inference capability is becoming a meaningful hardware differentiator at the client tier. Whether OEMs adopt the platform aggressively depends on Microsoft's Windows-on-Arm ecosystem maturity and whether Nvidia can offer a compelling total cost to PC makers accustomed to Intel's supply chain infrastructure. The Arm CPU pivot also puts Nvidia in direct competition with Qualcomm's Snapdragon X Elite in a segment Qualcomm has spent years cultivating.

SoftBank has committed up to $87 billion for AI data centres in France, with Tom's Hardware explicitly noting France's nuclear grid as the decisive site selection factor over US locations. This is one of the most direct public acknowledgements from a major capital allocator that power grid reliability and carbon intensity — not real estate, fibre access, or regulatory environment — now drive hyperscale siting decisions. France's EDF nuclear baseload provides both the power density and the green energy credentials that US grid-constrained markets increasingly cannot. The scale of the commitment — $87 billion — dwarfs most sovereign infrastructure programmes and, if executed, would make France a primary European AI compute hub.

The financial context requires scrutiny. SoftBank carries over $130 billion in debt and took a $40 billion bridge loan in March to fund its OpenAI investment, per Tom's Hardware. This is an announced commitment, not a confirmed build — infrastructure professionals should track whether project finance structures, offtake agreements, and actual construction contracts follow. The gap between announced AI infrastructure spend and capital actually deployed has been substantial across the industry in 2025-2026.

Intel used Computex 2026 to announce Crescent Island, an inference-optimised data centre GPU with up to 480GB of LPDDR5X memory — a deliberately differentiated memory architecture designed to address the HBM supply shortage that constrains Nvidia-based deployments. As Tom's Hardware and Data Centre Dynamics both report, the chip is air-cooled and targets inference rather than training — a pragmatic positioning that cedes the high-end training market to Nvidia while targeting the much larger, cost-sensitive inference deployment base. Intel also launched Xeon 6+ (Clearwater Forest) aimed at agentic AI workloads and new Ethernet E835 network adapters for dense virtualised deployments, per Data Centre Dynamics.

The critical caveat is that Crescent Island and Xeon 6+ have been detailed and announced, not shipped at volume. Intel's previous GPU generations — Ponte Vecchio and Gaudi 3 — shipped in limited quantities and failed to achieve meaningful market share against Nvidia. The LPDDR5X memory strategy is analytically interesting because it sidesteps HBM procurement competition and allows Intel to lean on a broader pool of DRAM suppliers, but the bandwidth per GPU will be materially lower than HBM-based alternatives, limiting applicability to long-context inference rather than high-throughput training or demanding inference tasks.

Anthropic's $65 billion raise at a $965 billion valuation included equity stakes from Micron, Samsung, and SK Hynix, per Data Centre Dynamics. The participation of all three major DRAM and HBM suppliers in a single frontier AI lab's funding round is structurally significant: it creates financial alignment between memory supply and model development that could translate into preferential allocation, co-development arrangements, or early access to next-generation HBM4 capacity. For competitors relying on the same suppliers, this raises questions about neutrality in allocation decisions during supply-constrained periods.

Separately, a top-performing technology fund is initiating a position in SK Hynix specifically on a memory crunch thesis, per Bloomberg, after SK Hynix shares have already rallied approximately 1,000% over the prior year. The dual signal — strategic investors (memory chipmakers themselves) and financial investors both piling into the AI memory trade — reflects consensus that HBM supply will remain structurally tight. This is relevant for infrastructure planning because HBM availability, not GPU compute availability per se, is increasingly the binding constraint on AI accelerator shipments.