Agentic cloud capacity just moved from roadmap talk to a named availability day. On Wednesday, Sept. 30, at CoreWeave Fully Connected in San Francisco, NVIDIA and CoreWeave said Vera Rubin NVL72 systems with Spectrum-X 102.4T Ethernet are available on CoreWeave Cloud. The news is production infra for agent workloads, not a safety-platform software release.
Cognition, the lab behind the Devin coding agent, is the first customer running production workloads on Vera Rubin, NVIDIA's blog says. In early tests, Cognition reported up to a 4.8x increase in total token throughput for SWE-2 inference versus a GB200 NVL72 baseline. That 4.8x figure is a customer early-test claim on a sampled FrontierCode workload, not a third-party audit.
CoreWeave also said it will offer NVIDIA Vera, which both companies frame as the first CPU built for AI agents. CoreWeave's company news says rack-scale Vera puts 128 CPUs and 11,264 cores in a single rack, with room for more than 11,000 concurrent one-core environments. In testing, CoreWeave says it saw more than 3x faster agent sandbox startup times on Vera versus an x86 CPU. Those density and startup figures are company claims.
Alongside the hardware, CoreWeave launched CoreWeave Forge, a connected environment that unifies Weights and Biases, OpenPipe post-training, and the open-source marimo notebook project. The stated goal is to close the loop from production behavior back into training and evaluation for models and agents. NVIDIA's blog notes Canva, Capital One, and MasterClass among early Forge builders.
What remains open is how the 4.8x and 3x claims hold outside early tests, how fast Vera Rubin capacity lands for customers beyond Cognition, and how Vera CPU sandboxes behave at the claimed concurrent density. This story is distinct from NVIDIA's Open Agent Safety platform work on OpenShell and Sentry. It is also separate from fellowship announcements and from capital-raise filings.