NVIDIA and CoreWeave bring Vera Rubin NVL72 to production for AI agents

CoreWeave announced production availability of NVIDIA Vera Rubin NVL72 systems on its cloud and their use for production AI workloads. The announcement was made at CoreWeave Fully Connected in San Francisco, taking place this week.

Cognition, the company behind the Devin AI software engineer, is the first customer running production workloads. In early software-engineering tests, Vera Rubin NVL72 delivered up to 4.8x higher total token throughput than GB200 NVL72.

CoreWeave also announced NVIDIA Vera, a CPU designed for agentic workloads, and launched CoreWeave Forge, an environment for training, evaluating and continuously improving models and agents.

Cognition scales Devin

Cognition uses CoreWeave for Devin training, reinforcement learning and production inference. The company scaled to thousands of GPUs on the platform within nine months.

For its comparison with GB200 NVL72, Cognition sampled tasks from FrontierCode and assigned AI agents to solve them. Cognition founding team member Silas Alberti said agentic coding requires long contexts, high concurrency and attention to token cost; the test results are intended to support faster code generation and more responsive reasoning in Devin.

Vera supports isolated environments

A CoreWeave Vera rack contains 128 CPUs and 11,264 cores, enough for more than 11,000 concurrent environments at one core each. CoreWeave Sandboxes hardware-isolates those environments for training, agent tool use and model evaluation.

CoreWeave said its tests showed agent sandbox startup times more than three times faster on Vera. On Terminal-Bench, it reported a 1.7x performance gain across all passing tasks.

Forge connects production and training

Forge combines Weights & Biases, OpenPipe’s post-training expertise and the open-source marimo notebook project. It supports models, frameworks and clouds from different providers while feeding production data into subsequent training cycles.

The platform includes the generally available CoreWeave ARIA and Sandboxes services, as well as the new Agent Lens service. CoreWeave said Agent Lens improves failure detection by 20% and cuts issue-fixing costs by half, based on analysis of tens of millions of production agent traces.

CoreWeave reported that serverless reinforcement learning in Forge trains 1.4x faster at 40% lower cost than a self-managed setup. Canva, Capital One and MasterClass are among the first companies building on Forge.