← Back to feed

Vera Rubin is in full production ramp. NVIDIA confirmed on July 21 that Vera Rubin NVL72 racks are now running at all four major cloud providers: CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure. The platform’s supply chain now spans 350+ factory sites across 30 countries — one of the largest coordinated chip production buildouts ever assembled.

Anchoring the US portion of that chain: Wistron’s D1 facility in Fort Worth, Texas. The 324,000-square-foot greenfield plant opened on July 21 with NVIDIA founder and CEO Jensen Huang and Wistron Chairman Simon Lin on stage at the opening ceremony. Taiwan government officials and local Fort Worth leaders attended.

The plant currently runs two manufacturing cells: one for the NVIDIA GB300 Grace Blackwell Ultra Superchip and one that will produce the Vera Rubin Superchip. Wistron is scaling D1 to produce tens of thousands of boards per month before the end of 2026.

“Manufacturing is an essential pillar for every economy and every country,” Huang said. “Building chip plants, packaging plants, computer system plants like this, and AI factories all over the United States, has allowed the United States to really reindustrialize for the first time in a long time.”

The Vera Rubin NVL72 Architecture

Vera Rubin NVL72 is a 7-chip, 5-rack-tray system: Vera Rubin NVL72, Vera CPU rack, Groq 3 LPX, Spectrum-6 SPX, and Vera BlueField-4 STX. NVIDIA engineered these as a single integrated unit rather than discrete off-the-shelf components assembled at the rack level.

At the center is NVIDIA’s custom Olympus CPU core: 2x single-threaded performance, 3x core-to-core bandwidth, and 40% lower memory latency versus competing chiplet designs. NVIDIA positions the Olympus core specifically for the coordination overhead of agentic AI workloads — the class of tasks where CPU-GPU communication patterns differ sharply from standard batch inference.

Efficiency Numbers That Move Procurement

CoreWeave’s DeepSeek-R1 benchmark on Vera Rubin NVL72 delivered 10x more throughput per megawatt versus Grace Blackwell NVL72. For power-constrained AI data centers — which is effectively all of them at hyperscale — that metric translates directly to operating cost per token.

The benchmark result arrived at a moment when power procurement has become the primary constraint on AI capacity expansion. Operators who committed to Vera Rubin deployments earlier in 2026 are now receiving production hardware across all four major cloud platforms simultaneously, which marks a step-change from the staged rollouts that characterized Grace Blackwell.

The US Manufacturing Signal

Wistron’s Fort Worth plant is part of a broader reshoring pattern. NVIDIA’s AI manufacturing is now distributed across domestic and allied-nation factories at a scale that would have been unusual five years ago. The Fort Worth D1 facility is a greenfield build, not a conversion — designed from the ground up for the board-level assembly complexity that Vera Rubin’s codesigned rack architecture requires.

Scaling to tens of thousands of boards per month from a single 324,000-square-foot plant signals that Vera Rubin’s production ramp is being treated as infrastructure-level, not product-level — closer in kind to how automotive manufacturers approach platform rollouts than how GPU generations have historically shipped.