CoreWeave Will Offer Nvidia CPU Racks It Says Can Run Over 11,000 AI-Agent Environments

CoreWeave will add Nvidia’s Vera CPUs to its cloud for AI workloads, including the environments where agents carry out their work. The September 30, 2026 announcement promises dense computing capacity alongside the company’s training fleet, but gives no availability date.

CoreWeave presented the addition at Fully Connected, its AI cloud conference. Its explanation starts with a division of labor: GPUs power model training and reasoning, while CPUs carry out the surrounding software work.

• Tool calls and code execution: carrying out the software work surrounding model reasoning.
• Isolated sandboxes: separate environments in which agents can execute their work.
• Data pipelines and reinforcement-learning environments: preparing data and supporting training episodes.

That demand can arrive in bursts. CoreWeave says a single step can require thousands of environments for an hour, followed by almost no demand until the next run. Its infrastructure pitch therefore combines density with the ability to scale as demand changes.

128CPUs per rack

CoreWeave says its rack-scale Vera configuration contains 128 CPUs.

11,264Cores per rack

The same configuration has 11,264 CPU cores, according to CoreWeave.

More than 11,000Concurrent agent environments

CoreWeave says the rack has capacity for more than 11,000 simultaneous agent environments.

CoreWeave says Vera will run on bare metal, directly on hardware, under the same platform and consumption models as the rest of its fleet. The offering will support CoreWeave Sandboxes.

Through Sandboxes, the company says agents can operate in fully hardware-isolated environments alongside the training jobs they support. The rack also includes Nvidia BlueField-4 data processing units and Spectrum-X Ethernet switching. CoreWeave describes these networking components as providing secure, high-performance connections between Vera nodes.

The company identifies two measures that matter: how many isolated environments can run at once, and how consistent each environment’s performance remains as that count rises.

CoreWeave says its testing achieved more than three times faster agent sandbox startup on Vera than on an x86 CPU. That measures bringing an environment into operation, not how quickly an agent completes an entire assignment.

The announcement does not identify the x86 processor or disclose the test configuration behind that comparison. That limits readers’ ability to reproduce or assess the claimed startup advantage.