CoreWeave says Vera Rubin NVL72 is in production, Cognition first

CoreWeave said on 30 September 2026 that NVIDIA's Vera Rubin NVL72 systems are available on CoreWeave Cloud, and that Cognition, the lab behind the coding agent Devin, is the first customer running production workloads on them anywhere. Cognition and CoreWeave stood up a Vera Rubin cluster in early September, per the release.

The performance numbers in the release are Cognition's and CoreWeave's own benchmarks. Cognition's Silas Alberti is quoted saying engineers saw up to 4.8 times total token throughput on SWE-2 inference versus a GB200 NVL72 baseline, and 3.8 times output token throughput on reinforcement-learning workloads. CoreWeave separately cites a measured 10 times token throughput per megawatt on DeepSeek R1 versus GB200 at matched interactivity. Those figures are workload-specific claims, not a general service-level promise.

Why it matters on this map: CoreWeave is a specialized AI cloud that rents NVIDIA-based capacity. Disclosing a new GPU generation in production with a named customer is evidence that next-wave rack-scale systems are leaving the lab and serving live traffic at least at one provider. The release does not give GPU counts, megawatts or contract value. Dell is named as a deployment partner; NVIDIA is the silicon vendor. Neither is recolored by this signal.

CoreWeave moves to green. The compute-and-power condition adds a supporting data point without a status change. Production at one neocloud is not a gigawatt total. Cognition's Devin throughput claims are not treated as a consumer-agent adoption data point under the map's evidence rule for that condition. Nothing turns red.

What would change the picture: CoreWeave disclosing megawatts or GPU counts of Vera Rubin capacity; a second named production customer; or Cognition publishing usage or revenue figures that meet the agent-row evidence standard.