Nvidia and CoreWeave are collaborating to integrate the Vera CPU into AI infrastructure to address performance bottlenecks. While GPUs manage model reasoning, CPUs handle the execution of tasks, including API calls, tool usage, and SQL queries, which are critical for agentic AI operations.
CoreWeave plans to offer standalone Vera CPU capacity to its clients. The company utilizes its Sandboxes environment for reinforcement learning and inference tasks. Testing with Vera CPUs demonstrated a 3x improvement in Sandbox startup times compared to previous configurations.
The Vera Rubin hardware tray integrates the Vera CPU with BlueField-4 data processing units. This architecture supports the Open Agent Safety Platform, which combines OpenShell runtime controls running on the CPU with Nvidia Sentry monitoring on the DPU to ensure infrastructure security.
Engineering priorities for CoreWeave include managing power, cooling, and automation to support the scaling of AI systems. The companies aim to ensure that networking, storage, and processing components operate cohesively to prevent performance limitations in agentic AI workflows.
