CoreWeave: Vera Rubin NVL72 delivers 10x tokens per MW versus Blackwell
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
The claimed 10x was obtained with inference optimizations enabled on a single DeepSeek R1 workload — it is the effect of the whole stack, not just the new accelerators.
CoreWeave has published the first silicon-measured results for NVIDIA Vera Rubin NVL72. At comparable interactivity (tokens per second per user), the system delivered 10 times more tokens per second per megawatt than NVIDIA GB200 NVL72. The comparison ran on the same DeepSeek R1 workload. The measurements were taken with inference optimizations enabled, including expert parallelism, NVFP4 precision, multi-token prediction and separate prefill and decode phases on NVIDIA TensorRT-LLM and Dynamo. The rack consists of 72 Rubin GPUs and 36 Vera CPUs linked by NVLink 6 fabric with 260 TB/s all-to-all bandwidth. Bring-up and validation were completed in early June 2026. The measurement was performed by the cloud operator itself; CoreWeave calls the publication a starting point and promises to update the numbers as optimizations progress.
Sources 1