· Updated
The NVL72 rack — 72 Rubin GPUs and 36 Vera CPUs — trains large mixture-of-experts models with one-fourth the GPUs of Blackwell and achieves 10x higher inference throughput per watt at one-tenth the cost per token. The Groq 3 LPX rack delivers 35x inference throughput per megawatt for trillion-parameter models.
NVL72's one-fourth GPU requirement compared to Blackwell is the cost story — teams running mixture-of-experts models at scale get the same performance at a fraction of the hardware. Anthropic, OpenAI, Meta, and Mistral AI building on this platform positions Vera Rubin as the standard infrastructure for frontier model training and serving.
NVIDIA Vera Rubin is opening the next frontier of AI. #NVIDIAGTC news: The Vera Rubin platform’s seven chips are now in full production to scale the world’s largest AI factories. Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6 and Groq 3 work together as one AI supercomputer powering every phase of AI. https://t.co/GqYcF1sfRg
