CoreWeave Brings NVIDIA Vera Rubin NVL72 to the Cloud With Up to 4.8x Token Throughput

CoreWeave announced early access to NVIDIA Vera Rubin NVL72 systems on CoreWeave Cloud and said it has already brought a production cluster online for Cognition. The move gives customers access to the new hardware through CoreWeave's cloud stack, while the headline benchmark came from a narrow early test rather than a broad workload comparison.

For the comparison, Cognition ran AI agents on software-engineering tasks sampled from the FrontierCode dataset. The company tested Vera Rubin NVL72 against GB200 NVL72, and the reported 4.8x figure applies to those early SWE-2 inference workloads.

CoreWeave paired the systems with Spectrum-X 102.4T Ethernet networking. Customers can manage cloud resources through CoreWeave Kubernetes Service, SUNK, Mission Control, Sandboxes and Inference.

NVIDIA said the benchmark results could mean faster real-time code generation and more responsive multistep reasoning for Devin. CoreWeave stood up a production Vera Rubin cluster for Cognition within days of receiving its first production racks.

Cognition's Silas Alberti said the company values having the components of its agentic coding workload on one platform, rather than focusing on a single benchmark result.

A single CoreWeave Vera CPU rack contains 128 processors and 11,264 cores, enough to run more than 11,000 concurrent environments at one core each. CoreWeave also reported a 1.7-times performance gain on Vera CPUs across all passing Terminal-Bench tasks.

How much weight should the early 4.8x SWE-2 result carry when assessing Vera Rubin NVL72's real-world cloud performance?