Nvidia announced Monday at Computex in Taipei that it has begun mass production of the Vera Rubin platform, aimed at powering a new generation of AI factories. First unveiled in March, Vera Rubin represents a complete architectural overhaul from its predecessor, Grace Blackwell.
The platform includes new Rubin GPUs, Vera CPUs, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and the Spectrum-6 Ethernet switch, delivering 10x the throughput of previous systems. CEO Jensen Huang said, "One prompt can launch a thousand-step journey of reasoning, retrieval, tool use, and response generation."
Vera Rubin is liquid-cooled, featuring 72 Rubin GPUs and 36 Vera CPUs per rack, designed to train large models using one-fourth the GPUs of Blackwell while delivering 10x throughput at one-tenth the cost per token.
Nvidia is also introducing the world's first co-packaged optics-based network switches via Spectrum-X Ethernet Photonics, offering 5x power efficiency and 1.3x faster deployment. Security is handled by BlueField-4 STX, enabling full-stack confidential computing with data encrypted across interconnects at 800 GB per second.
Key partners Dell, HPE, Supermicro, and Lenovo are manufacturing Vera Rubin servers, with first shipments expected this fall.