CoreWeave Deploys NVIDIA Vera Rubin NVL72 for Agentic AI Workloads

CoreWeave and NVIDIA have extended a nearly decade‑long co‑engineering relationship by putting the new Vera Rubin NVL72 platform into production on CoreWeave Cloud. The rack‑scale system pairs 72 GPUs with Spectrum‑X 102.4 Tbps Ethernet, delivering a full‑stack AI factory that can be provisioned in days.

Cognition, the lab behind the Devin AI software engineer, was the first customer to run real‑world agentic workloads on the new hardware. In early benchmarks using a FrontierCode‑derived software‑engineering suite, Vera Rubin NVL72 produced up to 4.8 times the total token throughput of a GB200 NVL72 baseline, translating into faster code generation and more responsive multi‑step reasoning for Devin.

Beyond raw compute, CoreWeave introduced CoreWeave Forge — a connected environment for training, evaluating, and iterating models and agents on NVIDIA‑accelerated infrastructure. Forge integrates dataset management, reinforcement‑learning loops, and production‑grade inference endpoints, allowing teams to move from experiment to deployment without changing platforms.

Why it matters for GPU and AI infrastructure

The Vera Rubin launch demonstrates how a cloud provider can preserve investment across GPU generations while instantly offering the latest silicon. For buyers, this means lower total cost of ownership, reduced migration risk, and the ability to match the right accelerator — whether a GPU, CPU, or purpose‑built AI processor — to each stage of an agentic pipeline.