Nvidia's Vera CPUs ship to OpenAI, Anthropic, SpaceX - and Wistron's $700M Texas plant is where they get racked

In plain terms - Nvidia has started shipping its Vera CPUs (the successor to Grace) to OpenAI, Anthropic, and SpaceX. At the same time, Wistron officially opened a $700 M Blackwell-server plant in Texas. Two announcements, one message: the Grace/Vera + Blackwell stack is going into physical racks at the top-three customers, in the US.

Contexte

Vera is Nvidia's next-gen ARM-based server CPU, designed to sit next to a Blackwell (or later) GPU in the same NVL rack, with NVLink C2C connecting them. It replaces Grace on the roadmap. Wistron is one of two Taiwanese ODMs (with Foxconn) that manufacture the reference NVL racks for hyperscalers and frontier labs.

Under the hood

• Vera : ARM-based, high memory bandwidth to feed adjacent GPUs. Ships as part of the Vera-Rubin / Vera-Blackwell reference designs.
• Delivery scope : first units confirmed shipped to OpenAI, Anthropic, and SpaceX. Korean and Japanese buyers reportedly next in queue (per ETNews and Nikkei coverage).
• Wistron Taylor, TX : $700 M investment, targets Blackwell-generation servers for US customers. Aligns with DoC "made in USA" incentives and reduces exposure to Taiwan-side supply concentration.
• Racks : Nvidia's NVL72 and successors are the reference form factor. A single fully-populated NVL rack is now a >$3 M line item.

Analyse

Vera-to-OpenAI/Anthropic/SpaceX is not a routine bill of materials - it is Nvidia curating who gets the frontier stack first. That priority list confirms the shape of the frontier customer base: two labs and one anchor tenant (SpaceX for onboard/edge compute + Starlink AI workloads). Korea and Japan getting next-in-queue status matches the sovereign-compute pattern (see fil japan-ai-industrial-stack).

Wistron opening in Texas closes a supply-chain loop that the Trump administration and the Biden CHIPS act both wanted: high-value server assembly in the US, feeding US frontier labs. The unit economics of ODM assembly are thin, but the political option value is huge - Wistron gets US-buyer preference in future federal contracts, and Nvidia gets a second physical rack line outside Taiwan.

For infra buyers: Vera-Blackwell rack availability is now the bottleneck, not GPU count on paper. The queue is real, and it's US-manufactured.

Scénarios

• Base case (60%) : Wistron Taylor ramps to full run rate by Q2 2027; Vera-Blackwell racks account for the bulk of new US hyperscaler capacity.
• Faster (25%) : Foxconn adds a second US site by late 2027; ODM competition compresses assembly margins but expands throughput.
• Slower (15%) : power/permit delays at Taylor slow ramp; Taiwan-side capacity stays dominant into 2028.

Implications

For a hyperscaler procurement lead: the Vera queue is priced-in; get on it now. For a public-market investor: Wistron's US business becomes materially strategic; margins matter less than volume commitments.