Nvidia’s Vera Rubin NVL72 trains models with 75% fewer GPUs than Grace Blackwell

Nvidia has a habit of making its last chip look obsolete before most customers have finished unpacking it. The Vera Rubin NVL72 is the latest case in point. The new system, which packs 72 Rubin GPUs and 36 Vera CPUs into a single NVLink domain, can train a 10-trillion-parameter mixture-of-experts model on 100 trillion tokens in one month using about 75% fewer GPUs than the Grace Blackwell NVL72 configuration. What the Vera Rubin NVL72 actually is The system is the direct successor to Nvidia’s Blackwell NVL72 racks, which shipped in GB200 and GB300 variants. Each GPU in the Vera Rubin NVL72 carries 288 GB of HBM4 memory. The architecture also introduces new interconnectivity designs and updated precision formats, both aimed at squeezing more useful work out of every watt consumed. The cost reduction on the inference side is equally striking. Nvidia projects that running inference on the Vera Rubin NVL72 will cost one-tenth the price per million tokens compared to Blackwell racks. Third-party benchmarks back up at least part of that story. MLPerf Inference v6.1, released in September 2026, recorded throughput improvements of up to 3.7 times higher than the GB300 NVL72 on the Qwen3-VL...