Nvidia’s Vera Rubin NVL72 trains models with 75% fewer GPUs than Grace Blackwell
Nvidia has launched the Vera Rubin NVL72 system, which trains massive AI models using 75% fewer GPUs than its predecessor, the Grace Blackwell NVL72.
The Vera Rubin NVL72 system can train a 10-trillion-parameter mixture-of-experts model on 100 trillion tokens in one month while using dramatically fewer components. Each GPU in the system carries 288 GB of HBM4 memory, with new interconnectivity designs and updated precision formats aimed at maximizing efficiency.
Performance Improvements
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 support these claims, with MLPerf Inference v6.1 recording throughput improvements of up to 3.7 times higher than the GB300 NVL72 on the Qwen3-VL task.
Artificial Analysis benchmarks showed approximately 5.7 times higher throughput per megawatt than the GB300 NVL72 on agentic AI workloads, demonstrating the system's superior energy efficiency.
Market Adoption
Production shipments began ramping in fall 2026, with support from over 80 MGX partners. The two most consequential names on that list are Google Cloud and Microsoft Azure, with Google Cloud anticipating scaling to as many as 960,000 Rubin GPUs by November.
Related
