CoreWeave Launches NVIDIA Vera Rubin AI Infrastructure and Forge Platform for Production-Ready Agentic AI

Image: NVIDIA · Source

CoreWeave introduces NVIDIA Vera Rubin NVL72 systems and Vera CPU to its cloud, enhancing AI agent workloads with up to 4.8x faster inference and over 3x faster sandbox startups. It also debuts CoreWeave Forge, a unified environment to streamline AI model training and deployment.

CoreWeave, a cloud provider specializing in AI workloads, has announced the deployment of NVIDIA's next-generation AI infrastructure within its cloud platform, marking a significant step in supporting production-ready agentic AI applications. Leveraging nearly a decade of close collaboration with NVIDIA, CoreWeave now offers the NVIDIA Vera Rubin NVL72 system equipped with Spectrum-X 102.4T Ethernet networking to its customers, delivering substantial improvements in inference performance.

Cognition, an applied AI lab known for their Devin AI software engineer, is the first to operate production workloads on Vera Rubin NVL72. Early benchmarking showed that the new infrastructure provides up to a 4.8 times increase in token throughput for software engineering inference tasks compared to NVIDIA's previous GB200 NVL72 system. This gain translates into faster real-time code generation and more responsive multi-step reasoning for the Devin agent, essential for handling complex workloads characterized by long contexts and high concurrency.

Alongside this, CoreWeave is among the first cloud providers to introduce the NVIDIA Vera CPU, a processor designed from the ground up for agentic AI workloads. Its architecture supports running thousands of isolated environments concurrently, a critical requirement for the scaling needs of agent training and evaluation. Testing demonstrated that CoreWeave’s deployment of Vera CPUs achieved more than three times faster agent sandbox startup times, accelerating isolated reinforcement learning and evaluation environments. The Vera CPU racks provide up to 11,264 cores per rack, enabling over 11,000 simultaneous isolated agent workloads.

Complementing this hardware rollout is the launch of CoreWeave Forge, an integrated platform designed to close the AI development loop by unifying training, evaluation, and continuous improvement workflows. Forge combines tools like Weights & Biases, expertise from OpenPipe, and the open-source marimo notebook project to offer an open, connected environment suitable across various AI models, frameworks, and clouds. New features include CoreWeave ARIA for experiment analysis and code iteration, Agent Lens for improved production agent monitoring and failure detection, and CoreWeave Sandboxes for running isolated AI agent tasks safely and efficiently.

CoreWeave Forge supports serverless supervised fine-tuning and reinforcement learning, reportedly training models 1.4 times faster and at 40% lower cost compared to self-managed setups. Additionally, NVIDIA's open-source inference framework Dynamo powers CoreWeave’s managed inference services and supports RL Rollouts, enabling continuous reinforcement learning without redeployment.

The combined NVIDIA and CoreWeave platform has attracted a diverse range of users from startups to enterprises. For instance, Ennoble Care, a large healthcare provider, uses CoreWeave’s infrastructure for clinical AI inference to support documentation and decision-making across multiple states. CoreWeave also holds distinguished industry rankings, including Platinum status in several MLPerf benchmarks, and serves nine of the top ten AI labs.

These advancements position CoreWeave and NVIDIA as key enablers for transitioning AI agents from experimental stages to robust production systems capable of software development, clinical support, and other complex real-world tasks. The new offerings were announced and showcased at CoreWeave Fully Connected, a technology event held in San Francisco.