Nvidia's Vera Rubin system targets AI dominance with CPU+GPU combo
10x tokens per watt, cable-free racks, and OpenAI already onboard...
Deep Dive
Nvidia unveiled Vera Rubin, its next-gen hybrid superchip system featuring 36 Vera CPUs paired with 72 Rubin GPUs per rack. The company claims 10x tokens per watt over Grace Blackwell, nearly 3x memory bandwidth, and hot-swappable, liquid-cooled racks that install in minutes. OpenAI already has one rack in use. Vera Rubin will ship in the second half of this year.
Key Points
- Vera Rubin NVL72 packs 36 Vera CPUs and 72 Rubin GPUs per rack, offering 10x tokens per watt over Grace Blackwell
- System is 100% liquid-cooled, cable-free, and hot-swappable—reducing rack installation from hours to minutes
- OpenAI already has one Vera Rubin rack in use; standalone Vera CPUs could ship to China by August
Why It Matters
Nvidia aims to own every chip in AI data centers, threatening AMD and Intel's CPU stronghold.