From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI

At CoreWeave Fully Connected in San Francisco, CoreWeave announced NVIDIA Vera Rubin NVL72 availability with Spectrum-X 102.4T Ethernet. Cognition, builder of the Devin coding agent, is cited as the first customer on production Vera Rubin workloads, reporting up to 4.8x total token throughput on SWE-2 inference versus GB200 NVL72 in early tests.

CoreWeave will also offer NVIDIA Vera, described as the first CPU built for AI agents, and launched CoreWeave Forge for training, evaluating, and improving models and agents. NVIDIA’s Ian Buck highlights multi-generation residual value, noting V100-era iron still earning on CoreWeave.

We’re watching agentic coding become the killer app that fills new racks.

4.8x token throughput on a real SWE workload is the kind of number that changes how many agent steps you can afford. Forge — train, eval, improve in one environment — is how you close the loop after swarming failures. Useful infrastructure meets agent ops where cost-per-token decides ship/no-ship.

I’m treating early Cognition benchmarks as directional until more customers publish. Vera CPUs for agents sound purpose-built; we’ll wait for scheduling and memory stories.

We’ll cover CoreWeave as the specialist cloud that ships silicon first. Agentic AI only scales if the factory loop includes evaluation, not just bigger GPUs.

NVIDIA Blog on CoreWeave Fully Connected announcements, late September 2026.