Nvidia unveiled the Vera Rubin platform at GTC 2026 - a complete architectural overhaul designed to power the enterprise shift to agentic AI. Named after astronomer Vera Rubin, the system comprises seven new chips: the Rubin GPU, Vera CPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch, and Groq LPX LPU.
The centerpiece is the liquid-cooled Vera Rubin NVL72 - 72 Rubin GPUs and 36 Vera CPUs interconnected via NVLink 6. Nvidia claims it trains MoE models using one-fourth the GPUs required by Blackwell, and delivers 10x inference throughput at one-tenth the cost per token.
The Vera CPU Rack (256 CPUs) accelerates reinforcement learning and agent validation - 50% faster and twice as efficient as x86 servers. The BlueField-4 STX rack serves as context memory for multi-turn agent interactions, boosting inference throughput five-fold. The Groq LPX Rack increases inference throughput per megawatt by 35x when paired with Rubin GPUs.
To address AI’s energy bottleneck, Nvidia introduced the DSX AI Factory Reference Design with DSX Max-Q software - delivering 30% more infrastructure within fixed power envelopes - and DSX Flex for grid-responsive computing. Digital twin capabilities via Omniverse DSX are already integrated by Schneider Electric and Siemens.
Vera Rubin systems will ship in H2 2026 through AWS, Google Cloud, Microsoft, Dell, and Supermicro.