At its CoreWeave Fully Connected conference in San Francisco on September 30, 2026, CoreWeave announced the production availability of NVIDIA Vera Rubin NVL72 systems paired with Spectrum-X 102.4T Ethernet networking on CoreWeave Cloud. Cognition, the creator of the Devin AI software engineer, has become the first customer running production workloads on the new infrastructure platform.
Vera Rubin Delivers Early Token Throughput Gains
According to CoreWeave, Cognition benchmarked the Vera Rubin NVL72 against a GB200 NVL72 baseline using tasks sampled from FrontierCode. In early tests on SWE-2 inference workloads, Cognition observed up to a 4.8x increase in total token throughput, which speeds up real-time code generation and multi-step reasoning for agentic coding.
CoreWeave deployable capacity is managed via CoreWeave Kubernetes Service, SUNK, CoreWeave Mission Control, CoreWeave Sandboxes, and CoreWeave Inference, allowing customers to deploy clusters in days.
Dedicated AI Agent CPU and Execution Environments
CoreWeave is also bringing the NVIDIA Vera CPU to its cloud platform to target agentic environments that require thousands of simultaneous, isolated sandboxes. A single rack configuration houses 128 Vera CPUs and 11,264 cores, offering hardware-isolated CoreWeave Sandboxes for reinforcement learning, tool execution, and evaluation.
In internal testing, CoreWeave reported more than 3x faster sandbox startup times on NVIDIA Vera CPUs compared to previous environments, alongside a 1.7x performance improvement across passing tasks on Terminal-Bench.
CoreWeave Forge Unifies Model Training and Observability
Alongside the hardware rollouts, CoreWeave launched CoreWeave Forge, an integrated environment connecting Weights & Biases, OpenPipe post-training tools, and the open-source marimo notebook project. The platform introduces CoreWeave ARIA for automated experiment analysis and code recommendations, as well as CoreWeave Agent Lens, which CoreWeave claims improves agent failure detection by 20%.
Forge also integrates NVIDIA Dynamo for inference and private previews of RL Rollouts to update model checkpoints dynamically during runtime. Initial users adopting Forge include Canva, Capital One, and MasterClass.
AI-generated editorial illustration; not a documentary photograph.
Source: From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI
SIGNAL TITANS
Leave a Reply