Nvidia Pitches Vera Rubin as Full AI Stack, Not Just GPUs
.jpg)
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Nvidia pitched Vera Rubin as a complete AI system at a Santa Clara briefing ahead of AMD's Thursday product event, with the NVL72 rack pairing 36 Vera CPUs with 72 Rubin GPUs in a monolithic — not chiplet — design.
- Vera Rubin NVL72 processes 10x as many tokens per watt as Grace Blackwell, per Nvidia claims, and the company is marketing the system as 'cable-free,' 'hot-swappable,' and 100% liquid-cooled, cutting rack install time from hours to minutes.
- Nvidia will sell the Vera CPU as a standalone product and reportedly told Chinese customers units could ship as soon as August, marking its formal push into CPU territory long dominated by AMD and Intel.
- OpenAI already has one Vera Rubin rack in its hands, per Nvidia, with CEO Jensen Huang naming Microsoft and Oracle as additional early customers ahead of a second-half 2026 ramp.
- AMD unveiled its competing Helios AI chip rack on Sunday, and both companies are chasing multi-year supply contracts from Meta, Amazon, OpenAI, Anthropic, and SpaceXAI.
- Nvidia benchmarks for Vera CPU against AMD and Intel appear to use slightly older generations of competitor chips, a detail buried in the fine print of last week's session led by VP Ian Buck.
Why it matters: Nvidia's Blackwell chips reportedly overheated in custom server racks and forced shipment delays, so on-time Vera Rubin delivery to OpenAI, Microsoft, and Oracle is the credibility test. By selling Vera CPUs standalone — including to Chinese customers as soon as August — Nvidia is trying to claim the same full-stack share of the AI data center that AMD's Helios and Intel's x86 chips currently split.
