Summary
CoreWeave has announced NVIDIA Vera Rubin NVL72 availability on its cloud, with Cognition as the first customer running production workloads. In an early software-engineering test, Cognition reported up to 4.8× the total token throughput of a GB200 NVL72 system.
CoreWeave has announced that NVIDIA Vera Rubin NVL72 systems are available on its cloud, with Cognition the first customer running production workloads on the platform. Cognition says an early test of software-engineering inference workloads delivered up to 4.8× the total token throughput of a GB200 NVL72 system. The result comes from a specific coding workload, rather than a general comparison across AI applications.
The announcement also includes plans to offer NVIDIA Vera CPUs and the launch of CoreWeave Forge, a connected environment for training, evaluating and improving AI models and agents.
What the early Vera Rubin tests measured
Cognition runs training, reinforcement learning and production inference for its Devin AI software engineer on CoreWeave. For its comparison, it sampled tasks from FrontierCode and deployed agents to solve them. The company benchmarked Vera Rubin NVL72 against a GB200 NVL72 baseline and reported up to 4.8× higher total token throughput for SWE-2 inference workloads. NVIDIA’s announcement says the improvement can mean faster code generation and more responsive multistep reasoning for Devin.
The test is relevant to agentic systems, which can make repeated model calls and use tools or execute code while working through a task. Their infrastructure needs to handle both model inference and the environments in which agents act. CoreWeave says its Vera CPU configuration places 128 CPUs and 11,264 cores in a rack, enough for more than 11,000 concurrent environments at one core each. Its Sandboxes service provides isolated CPU or GPU environments for agent tool use, reinforcement learning and evaluation.
CoreWeave reports that tests on Vera CPUs produced agent sandbox startup times more than 3× faster than its comparison setup. It also reports a 1.7× performance gain on Terminal-Bench across all passing tasks. CoreWeave says it will offer the Vera CPU on its cloud; the announcement describes the Vera Rubin NVL72 as available, while early-access customers can use the capacity.
Forge connects deployment with model improvement
CoreWeave launched Forge as a connected environment for working across the model-development cycle. It brings together Weights & Biases, post-training expertise from OpenPipe and the open-source marimo notebook project. CoreWeave says the environment is open across models, frameworks and clouds.
Among the announced components, CoreWeave ARIA and CoreWeave Sandboxes are generally available. ARIA can analyse experiment runs, suggest experiments and code changes, and save those changes in GitHub. Sandboxes provide isolated execution environments for agents, tool calls, reinforcement learning and evaluations. CoreWeave Agent Lens, described as a new service, uses production-agent observability to generate insights for improving systems. The company says it improves failure detection by 20% and fixes issues at half the cost.
The rationale for linking these tools is that activity in production can inform later training and evaluation. Forge is intended to bring that feedback into the same environment used to develop and improve agents, while Vera Rubin provides the newly announced cloud hardware for production workloads.