NVIDIA Vera Rubin NVL72 on CoreWeave: 10x More Tokens Per Megawatt Than Blackwell
CoreWeave, Tuesday, July 21st, 2026
CoreWeave publishes first Vera Rubin NVL72 benchmarks: 10x more tokens per megawatt than GB200.
CoreWeave published the first measured performance metrics for NVIDIA's Vera Rubin NVL72 GPU system. Running DeepSeek R1 inference at comparable user responsiveness, it achieved 10x tokens-per-second per megawatt versus the NVIDIA GB200 NVL72.
This efficiency lets organizations scale reasoning-model deployments within existing power budgets or cut infrastructure needs.
The gains are attributed to a unified 260 TB/s NVLink 6 fabric, native NVFP4 support, and software optimization via TensorRT-LLM and Dynamo.