CoreWeave says it will offer NVIDIA’s Vera CPU in its cloud, positioning the chip for the CPU-heavy work surrounding AI agents. The rack-scale system is designed to run more than 11,000 isolated agent environments at once, and CoreWeave says its tests showed sandbox start-up times more than three times faster than on an x86 CPU.
NVIDIA Watch analysis
What happened
The announcement expands CoreWeave’s compute portfolio with NVIDIA Vera, which the company describes as the first CPU designed specifically for AI agents. CoreWeave says Vera will run on its existing platform, alongside its CoreWeave Sandboxes service.
A Vera rack at CoreWeave combines 128 CPUs and 11,264 cores, with BlueField-4 DPUs and Spectrum-X Ethernet switching. The company says the system can support more than 11,000 concurrent environments, where agents can run in hardware-isolated sandboxes beside training jobs.
CoreWeave’s announcement describes a test in which Vera delivered more than three times faster agent-sandbox start-up than an x86 CPU. That is a company-reported result, not a like-for-like independent benchmark. The announcement was distributed by Business Wire and carried by the Financial Times markets site.
Why it matters
AI infrastructure talk often puts GPUs centre stage. But agents also need CPUs to run code, call tools, manage data pipelines and create isolated environments. If those supporting jobs become a bottleneck, faster model reasoning alone will not keep the whole loop moving.
CoreWeave’s pitch is that Vera can provide dense CPU capacity for bursts of agent work without forcing customers to rebuild their wider cloud setup. Whether that translates into better performance or lower costs for customers will depend on real workloads, not rack specifications alone.
Our read
This is a meaningful sign that AI infrastructure is broadening beyond accelerators: the machines that organise and execute an agent’s work may matter nearly as much as the machines that run the model. The eye-catching speed figure deserves a proper comparison, but the capacity and intended workload make the launch worth watching.
What to watch
- When Vera instances become available to CoreWeave customers and on what terms.
- How the sandbox start-up comparison was measured and whether independent tests reproduce it.
- Whether large agent workloads show gains in throughput, cost or reliability.
Discussion spark: For AI agents, should cloud providers prioritise more CPU capacity and isolated environments, or invest first in faster accelerators?
Sources and evidence
- CoreWeave to Offer NVIDIA Vera, the First CPU Built for AI Agents – Company Announcement – Financial Times (30 September 2026, 13:00 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.