The Vera Rubin NVL72, which combines 72 Rubin GPUs and 36 Vera CPUs in a single NVLink domain, can train a 10-trillion-parameter mixture-of-experts model on 100 trillion tokens in one month with about 75% fewer GPUs than the Grace Blackwell NVL72. Each GPU features 288 GB of HBM4 memory, while new interconnect designs and updated precision formats are intended to improve energy efficiency. For inference, Nvidia expects the price per million tokens to be one-tenth that of Blackwell racks. MLPerf Inference v6.1 and Artificial Analysis benchmarks measured up to 3.7 times the throughput of the GB300 NVL72 on the Qwen3-VL task and approximately 5.7 times the throughput per megawatt on agentic AI workloads. Production shipments began in fall 2026 with support from more than 80 MGX partners, including Google Cloud and Microsoft Azure, and Google Cloud plans to scale to 960,000 Rubin GPUs by November.
Source: Read the original article
