TL;DR
CoreWeave is adding NVIDIA's Vera CPU, the first chip purpose-built for AI agents, to its cloud fleet, delivering 11,264 cores per rack and 3x faster sandbox startup times at a moment when agentic workloads are straining conventional CPU infrastructure.
What happened
- CoreWeave announced NVIDIA Vera CPU availability at its Fully Connected conference, attended by more than 4,500 customers, partners, and developers.
- 128 CPUs and 11,264 cores per rack fit into a single Vera rack, enabling more than 11,000 concurrent isolated agent environments.
- BlueField-4 DPUs and Spectrum-X Ethernet switching provide secure, high-performance connectivity between Vera nodes.
- 3x faster agent sandbox startup times recorded on Vera versus a standard x86 CPU in CoreWeave's own testing.
- Vera runs on CoreWeave as bare metal under the same platform, consumption models, and economics as the rest of the fleet.
Why it matters
- Agentic AI loops are CPU-bound, not just GPU-bound: sandboxes, reinforcement learning environments, tool calls, code execution, and data pipelines all run on CPU, and that work now sets the pace of the entire AI loop.
- Demand is bursty and hard to predict: a single loop step can spike to thousands of environments for an hour, then drop to near zero, making density and elastic scaling critical.
- 11,000-plus concurrent environments per rack is a step-change in sandbox density, directly cutting the infrastructure cost and latency of post-training and agent evaluation pipelines.
- CoreWeave already holds the only Platinum ranking in SemiAnalysis ClusterMAX three consecutive times, so Vera lands on a platform with a proven performance track record.
- First-mover positioning matters: whoever hosts the most agent sandboxes at lowest latency becomes the default substrate for the next wave of AI development.
What to watch next
- Customer adoption benchmarks: whether hyperscalers or frontier labs shift post-training workloads to Vera-based infrastructure, confirming the CPU-for-agents thesis.
- Competitive response from AMD and Intel: both supply x86 CPUs to cloud AI fleets and will need to answer Vera's sandbox density and startup-time claims.
- CoreWeave pricing and availability details: consumption-model economics are promised but not yet quantified, and unit economics will determine whether Vera displaces general-purpose CPU instances at scale.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.