Story 01 November 20, 2026 issue
Daily AI-generated issue
CoreWeave measures 10× throughput per megawatt with Vera Rubin versus GB200
Top News · Official · coreweave.com
In CoreWeave's DeepSeek R1 evaluation, NVIDIA Vera Rubin NVL72 delivered 10× the tokens per second per megawatt of NVIDIA GB200 NVL72. Both systems were measured at a matched per-user throughput target.
Here's what went into the test:
- The rack connects 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs through a 260 TB/s all-to-all NVLink 6 fabric.
- Native NVFP4 support comes through the Transformer Engine, with NVFP4 precision used in the comparison.
- The inference setup combined large-scale expert parallelism, multi-token prediction, and disaggregated prefill and decode, enabled by NVIDIA TensorRT-LLM and NVIDIA Dynamo.
- CoreWeave completed bring-up and validation in early June, verifying power, cooling, networking, and compute at production scale.
One catch: CoreWeave calls this a starting measurement and expects further gains as it tunes scheduling, networking, and software.
Sources: wf.coreweave.com
This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent.