CoreWeave to offer NVIDIA Vera, billed as the first CPU built for AI agents, on bare metal

Why a hardware story is about CPUs rather than GPUs. The technical explanation in the release is that agentic AI operates through a continuous AI loop of run, observe, curate, improve and evaluate. GPUs power model training and reasoning, while CPUs execute the work surrounding every step: isolated sandboxes, reinforcement learning environments, tool calls, code execution and data pipelines. Demand for data preprocessing pipelines and reinforcement learning training episodes is bursty and hard to predict: a single step in the loop can demand thousands of environments for an hour, then virtually nothing until the next run. The release draws the conclusion that as post-training and agentic workloads grow, this CPU-intensive work increasingly determines how quickly the loop can turn.

The concrete specs. The release states that rack-scale Vera at CoreWeave puts 128 CPUs and 11,264 cores in a single rack, with BlueField-4 DPUs and Spectrum-X Ethernet switching delivering secure, high-performance connectivity between Vera nodes, which is room for more than 11,000 concurrent environments. On how it runs, the release stresses that Vera on CoreWeave runs the way the rest of the fleet does: bare metal, under the same platform, consumption models and economics. Operation-wise that sentence carries more information than the core count: the new hardware does not come with a new procurement and scheduling model.

The performance number and its basis. The release states that in testing, CoreWeave achieved more than three times faster agent sandbox startup times on NVIDIA Vera CPU compared with an x86 CPU. Chen Goldberg, executive vice president of product and engineering at CoreWeave, is quoted saying general-purpose infrastructure bottlenecks agentic AI and that Vera is the first CPU explicitly designed to accelerate it, noting that products like CoreWeave Sandboxes enable Vera natively out of the box so teams can instantly spin up thousands of isolated environments, removing operational friction and accelerating the entire AI loop on day one. The figures in this section come from the vendor's own testing; the release does not attach an independent benchmark report.

The release also quotes a third party, and the point is a trend judgment rather than a score. Ryan Shrout, president and general manager at Signal65, says AI infrastructure conversations have centered on accelerators and that agentic systems are changing that: a growing share of every agent cycle is CPU work, and that share will only increase as post-training and agentic deployments scale. He puts the emphasis on the surrounding platform: what decides whether teams can use that capacity is the platform around it, and running it on bare metal, beside the training fleet, under orchestration that already exists, is where the operational difference shows up.

Placed back into a longer arc, this announcement reads two ways. The first is hardware-side: for two years the keywords of AI infrastructure were GPUs and rack-scale interconnect, and if the claim that CPU work is a growing share of every agent cycle holds, CPUs move from a supporting component to a bottleneck component. The CoreWeave release is written on exactly that argument. The second is procurement-side: the release also highlights MLPerf inference and training results and a SemiAnalysis ClusterMAX rating, which is where cloud providers compete in the new agentic-workload narrative. Worth noting: the specs and performance comparison come from CoreWeave's own testing and third-party commentary it cites, not from an independent benchmark.

One note on how to read this. What is public has clear edges: confirmed are that CoreWeave will offer NVIDIA Vera CPU, the rack spec numbers, the BlueField-4 and Spectrum-X pairing, the target of more than 11,000 concurrent environments, and the tested figure of more than 3x faster sandbox startup. Not confirmed: when it will be generally available, how it is billed, and how it performs on real customer workloads. To judge a story like this, look at whether it changes how you calculate cost. If, inside a single agent task, the time and money to stand up environments starts to rival running the model, that is the real turning point.