NVIDIA's Vera Rubin Countdown to Mass Production! OpenAI, Microsoft Lead Testing, Rack Priced Up to $8 Million

NVIDIA (NVDA-US) is rapidly advancing its next-generation AI server platform, Vera Rubin, toward mass production. Company executives recently made a rare public disclosure of progress, revealing that major customers including OpenAI, Microsoft (MSFT-US), Anthropic, and CoreWeave (CRWV-US) have already received test racks and begun validation.

According to The Information, NVIDIA recently arranged for several senior executives to speak with the media, offering the first detailed look at Vera Rubin’s manufacturing process, testing status, and production planning.

Andrew Bell, NVIDIA’s senior vice president, stated that Vera Rubin’s development and production progress is significantly smoother compared to the previous-generation Blackwell platform. The company holds a “cautious optimism” regarding full-scale production.

The report indicates that dozens of customers have already received small numbers of Vera Rubin test racks, each equipped with 72 GPUs. In addition to OpenAI, Microsoft, and Anthropic, cloud service providers such as CoreWeave and xAI, the AI company under SpaceX (SPCX-US), have also begun testing.

On pricing, a procurement source revealed that each Vera Rubin rack is priced between $7 million and $8 million—significantly higher than the current Grace Blackwell 300 rack, which costs around $5 million.

During the media visit, journalists toured a testing facility near NVIDIA’s headquarters in Santa Clara, California. Approximately 30 server racks were operating on-site, many of them Vera Rubin systems. NVIDIA executives confirmed that OpenAI has already started using some of the equipment for testing.

A single Vera Rubin rack is roughly the height of a large filing cabinet but weighs approximately 4,000 pounds (about 1,814 kilograms)—nearly the weight of a pickup truck.

Bell acknowledged that Blackwell’s initial mass production faced multiple challenges, including issues with hardware, software, diagnostic systems, and manufacturing processes, requiring some design rework. He described it as a “difficult experience” for the team.

Learning from the previous generation, Bell said Vera Rubin has undergone numerous design improvements. The most significant changes include drastically reducing internal cabling within the rack and introducing more robotic automation in the assembly process, lowering manufacturing complexity and improving assembly efficiency and yield.

Bell believes these changes will help avoid the bottlenecks encountered during Blackwell’s early production phase.

On production capacity, he revealed that NVIDIA is currently collaborating with over a dozen manufacturing partners, and the overall supply chain could eventually achieve a maximum production capacity of 1,000 Vera Rubin racks per day.

At $7 million per rack, this theoretical maximum output would translate to over $630 billion in potential quarterly revenue—far exceeding NVIDIA’s approximately $82 billion in revenue for the quarter ending April this year.

However, this represents only the supply chain’s maximum theoretical production capacity. Actual shipments will depend on customer procurement demand, data center construction progress, and installation speed.

Ian Buck, vice president at NVIDIA, said the company prioritizes allocating GPU resources to customers who have already completed their server room and server installation preparations, ensuring equipment can be deployed quickly. He half-jokingly added that the final decision on resource allocation rests with CEO Jensen Huang.