Back

AI cloud company CoreWeave starts offering NVIDIA's next generation of AI computers and will add a processor built for AI agents

Two presenters unveil an NVIDIA server tray on stage beside the CoreWeave logo at CoreWeave's Fully Connected event
Image: NVIDIA

NVIDIA and CoreWeave said on Wednesday, Sept. 30, that NVIDIA's Vera Rubin NVL72 systems are now available on CoreWeave's cloud, with Cognition, the lab behind the Devin coding agent, the first customer running production work on them. CoreWeave will also offer NVIDIA Vera, which the companies call the first CPU built for AI agents. The speed gains they cite come from early tests.

NVIDIA's next generation of AI computers, Vera Rubin, is now running on CoreWeave's cloud for early-access customers. On Wednesday, Sept. 30, at CoreWeave's Fully Connected event in San Francisco, NVIDIA and CoreWeave said Vera Rubin NVL72 systems, with NVIDIA's Spectrum-X 102.4T Ethernet networking, are available on CoreWeave Cloud. Both companies are pitching the capacity for agentic AI, AI that is able to carry out tasks on its own.

Cognition, the lab behind the Devin coding agent, is the first customer running production work on Vera Rubin, NVIDIA's blog says. In early tests on coding tasks sampled from FrontierCode, Cognition saw the new systems process up to 4.8 times as much text in the same time as NVIDIA's GB200 NVL72 systems when running its SWE-2 model.

CoreWeave also said it will offer NVIDIA Vera, which both companies call the first CPU built for AI agents. CoreWeave says a single rack of Vera holds 128 CPUs and 11,264 cores, with room for more than 11,000 one-core work spaces running at once. In its own testing, CoreWeave says, sandboxes, the walled-off spaces where agents run code, started more than 3 times faster on Vera than on an x86 CPU.

Alongside the hardware, CoreWeave launched Forge, one connected workspace that brings together Weights and Biases, OpenPipe's expertise in extra training for models that are already trained, and the open-source marimo notebook project. The stated goal is to feed what models and agents do in real use back into how they are trained and tested. NVIDIA's blog names Canva, Capital One and MasterClass among the first companies building on Forge.

The 4.8 times and 3 times figures come from early tests, and it is not yet clear how they will hold up in wider use. Nor is it known how quickly Vera Rubin capacity will reach customers beyond Cognition, or how Vera's sandboxes perform when packed as densely as CoreWeave says they can be.

More on NVIDIA