NVIDIA’s next-generation Vera Rubin server system is expected to achieve large-scale production next year, with several executives disclosing detailed progress. Dozens of customers, including CoreWeave, Microsoft, OpenAI, Anthropic, and SpaceX, have already received test racks, each equipped with 72 GPUs and priced between $7 million and $8 million. The Vera Rubin system features significant design improvements, with nearly cable-free racks and robotic systems handling most assembly tasks, resulting in a markedly superior manufacturing process compared to the previous Blackwell generation. NVIDIA’s manufacturing partners will ultimately be capable of producing up to 1,000 racks per day; at this production rate, quarterly potential revenue could reach at least $630 billion.

Article author and source: Wall Street Journal

NVIDIA's next-generation Vera Rubin server system is advancing rapidly, with mass production expected next year, generating high anticipation among AI developers, cloud service providers, and Wall Street investors.

On July 22, according to The Information, several NVIDIA executives recently welcomed media visits, providing detailed insights into the production progress and technical specifications of Vera Rubin.

Andrew Bell, Senior Vice President at NVIDIA, stated that the production and manufacturing process for Vera Rubin is significantly improved over the previous-generation Blackwell products, and he remains optimistic about the smooth progress of this advancement.

The production ramp-up of the Vera Rubin racks will have a decisive impact on NVIDIA's future revenue. Estimated at maximum capacity, the potential revenue scale far exceeds NVIDIA's current quarterly performance, and market interest in its commercial prospects continues to grow.

Dozens of customers have received their test racks.

According to The Information, dozens of customers, including CoreWeave, Microsoft, OpenAI, Anthropic, and SpaceX AI, have received early shipments of small Vera Rubin racks, each equipped with 72 GPUs.

In terms of pricing, one buyer revealed that each Vera Rubin rack is priced between $7 million and $8 million, compared to the current flagship Grace Blackwell 300 rack, which sells for approximately $5 million, representing a significant increase.

During a visit to NVIDIA’s headquarters in Santa Clara, journalists also toured a semi-confidential testing site, where approximately 30 server racks were operational, many of them Vera Rubin units—NVIDIA executives indicated that OpenAI is using some of these racks.

In appearance, a Vera Rubin rack is about the size of a tall filing cabinet but weighs a massive 4,000 pounds—equivalent to a pickup truck.

The production process is smoother than Blackwell's, and the cable-free design improves assembly efficiency.

Compared to the numerous challenges faced at Blackwell's launch, Vera Rubin's production and manufacturing progress has been significantly smoother.

NVIDIA Senior Vice President Andrew Bell admitted in an interview that Blackwell initially “almost broke everything”—hardware issues, software issues, diagnostic system problems, and manufacturing challenges, ultimately forcing a complete restart.

Bell stated that Vera Rubin introduced significant design improvements: the rack contains almost no cables, and a robotic system handles most assembly tasks, greatly reducing assembly complexity and thereby improving production efficiency and yield.

He is cautiously optimistic that Vera Rubin can avoid the technical bottlenecks that Blackwell encountered.

Maximum daily production capacity of 1,000 racks, with potential quarterly revenue exceeding $630 billion

Regarding capacity planning, Bell revealed that NVIDIA’s current network of more than ten manufacturing partners will ultimately be capable of producing up to 1,000 Vera Rubin racks per day.

Based on this production capacity, the potential revenue per quarter would be at least $630 billion, compared to NVIDIA's actual revenue of $82 billion for the April quarter this year.

However, having the maximum production capacity does not mean that deliveries will occur at that rate; actual shipment volumes will still depend on customer demand and installation timelines.

NVIDIA Vice President Ian Buck said the company’s decisions on GPU allocation are directly tied to whether customers are physically ready to install servers, and jokingly implied that the final allocation decisions rest with CEO Jensen Huang.