CoreWeave announced on September 30, 2026, that Nvidia’s Vera Rubin NVL72 is available on its cloud, with Cognition, the lab behind the Devin coding agent, as the first customer anywhere running production workloads on the system. CoreWeave made the announcement during its Fully Connected conference.
The step follows CoreWeave’s bring-up and validation of Vera Rubin NVL72 in early June. Vera Rubin NVL72 is a rack-scale system with 72 Rubin GPUs and 36 Vera CPUs connected over sixth-generation NVLink.
What Cognition measured
Cognition runs training, reinforcement learning and production inference for Devin on CoreWeave. According to CoreWeave’s write-up of the deployment, Cognition’s engineers ran their own benchmark against a GB200 NVL72 baseline:
| Workload | Result reported by Cognition, vs GB200 NVL72 |
|---|---|
| SWE-2 inference | Up to 4.8x total token throughput |
| Reinforcement learning | 3.8x output token throughput |
CoreWeave says production workloads were running within days of the cluster being stood up. These are the customer’s own figures on its own models, as published by CoreWeave.
CoreWeave’s own measurement
Separately, CoreWeave published what it calls the first measured silicon performance for the platform. In its benchmark post, it ran DeepSeek R1 on Vera Rubin NVL72 and on Blackwell NVL72 and reports 10 times the tokens per second per megawatt at a matched interactivity target. The test used large-scale expert parallelism, NVFP4 precision, multi-token prediction and disaggregated prefill and decode, with Nvidia TensorRT-LLM and Dynamo.
Per-megawatt throughput is the figure that matters most to a cloud whose sites are limited by power rather than floor space: at a fixed power budget it sets how many tokens a data hall can serve. It is a vendor-run result on one model with every optimization enabled, so it describes that workload rather than a general speed-up.
Rubin-generation specs are compared with Hopper and Blackwell in the AI chips table, and on-demand GPU-hour rates from CoreWeave and other clouds are listed in GPU prices. Other accelerator news is in AI chip news.
