Details
- NVIDIA launched the Vera Rubin platform at GTC, announcing seven new chips now in full production to power agentic AI across pretraining, post-training, test-time scaling, and inference.
- Involves NVIDIA (Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet), plus Groq (Groq 3 LPU); partners include AWS, Google Cloud, Microsoft Azure, Oracle, CoreWeave, Anthropic, OpenAI, Mistral AI.
- Features five racks: NVL72 (72 Rubin GPUs, 36 Vera CPUs for 1/4 GPUs vs. Blackwell, 10x inference throughput/watt), Vera CPU (256 CPUs), Groq 3 LPX (256 LPUs, 35x throughput/megawatt), BlueField-4 STX storage, Spectrum-6 SPX Ethernet; extreme codesign for POD-scale AI factories.
- Succeeds Blackwell with 4x fewer GPUs for MoE training, 10x lower cost per token; shifts from chips/servers to integrated rack/POD systems for efficiency and resiliency.
- Availability from partners in H2 2026; DSX platform enables 30% more infrastructure in fixed-power centers; verified ecosystem expansion with 80+ MGX partners, cloud/system OEMs like Dell, HPE, Supermicro.
Impact
NVIDIA's Vera Rubin solidifies its rack-scale dominance, slashing AI training/inference costs by up to 10x versus Blackwell and enabling trillion-parameter agentic models at unprecedented efficiency. This accelerates adoption by hyperscalers and AI labs like OpenAI and Anthropic, outpacing competitors amid surging demand for POD-scale factories. Expect intensified ecosystem lock-in, with H2 2026 launches reshaping data center economics and energy use.
