Nvidia has unveiled new performance benchmarks for its upcoming Vera Rubin chip system, signalling a strategic shift from being purely a maker of graphics processing units (GPUs) towards positioning itself as a supplier of complete AI computing systems, including central processing units (CPUs). The move comes as the AI industry increasingly relies on more complex, agentic systems, which demand CPUs capable of orchestrating data flows and networking alongside the GPUs that handle model training and inference. The announcement, made at a technical workshop in Santa Clara, comes just ahead of rival AMD's own product event in San Francisco.

Vera Rubin is the successor to Nvidia's Grace Blackwell system and pairs one Vera CPU with every two Rubin GPUs, with a full NVL72 rack containing 36 CPUs and 72 GPUs. Nvidia claims the new system processes ten times more tokens per watt than Grace Blackwell, offers nearly triple the memory bandwidth, and requires far fewer cables, potentially cutting rack installation time from hours to minutes. OpenAI is reportedly already using one such rack, while Nvidia says standalone Vera CPUs could reach Chinese customers as soon as August; the briefings were led by vice president Ian Buck rather than chief executive Jensen Huang, who was in Japan announcing separate robotics partnerships.