CoreWeave says a test of NVIDIA Vera CPUs cut startup times for its agent sandboxes by about threefold, as it plans to offer standalone Vera CPU capacity alongside its existing infrastructure. The result points to CPUs becoming a more visible part of the kit for running AI agents, not just a supporting act to GPUs.
CoreWeave Watch analysis
What happened
In an interview at the Fully Connected event, CoreWeave senior director of product Harsh Banwait said the company had recently tested its Sandboxes offering with Vera CPUs and saw “about a 3x improvement” in startup times. Sandboxes provide isolated environments where agents can execute tasks during reinforcement learning and inference, according to SiliconANGLE’s report of the discussion.
Banwait also said CoreWeave plans to offer standalone Vera CPU capacity. NVIDIA director of product marketing Hannah Coutand described the company’s view that tool calls, API calls and SQL queries add CPU work as agents move from answering questions to executing tasks. She also outlined NVIDIA’s Open Agent Safety Platform, combining OpenShell runtime controls with NVIDIA Sentry on BlueField-4 DPUs.
Read SiliconANGLE’s report. The report notes that theCUBE’s coverage of the event was sponsored by CoreWeave, with sponsors having no editorial control, according to SiliconANGLE’s disclosure.
Why it matters
If agents spend more time calling tools and carrying out tasks, infrastructure needs to start those workloads promptly as well as serve model reasoning. Faster sandbox startup could matter to teams running repeated agent experiments or inference workflows, while standalone CPU capacity would give CoreWeave another way to serve that work. The reported result is a company-described test, not a general performance guarantee.
Our read
This is a useful signal about the less glamorous side of agent infrastructure: the pause before an agent can act is part of the experience too. A threefold improvement is worth attention, but the practical questions are the baseline, the workloads tested and whether customers see the same gain. A faster starting line is promising; it is not yet the whole race.
What to watch
- When CoreWeave makes standalone Vera CPU capacity available, and on what terms.
- Whether CoreWeave shares the test conditions and baseline behind the sandbox startup result.
- How CPU capacity and NVIDIA’s proposed safety platform fit into real customer deployments.
Discussion spark: If agent workloads increasingly depend on tool calls and isolated execution, should cloud buyers treat CPU performance as a first-order AI infrastructure choice alongside GPUs?
Sources and evidence
- Nvidia and CoreWeave rethink agentic AI infrastructure – SiliconANGLE (5 October 2026, 16:46 UTC)
not affiliated with or endorsed by CoreWeave