CoreWeave says its Vera Rubin NVL72 systems are now available, with Cognition already running production workloads on the new hardware. The company also announced a CPU built for AI agents, betting that the infrastructure around the GPUs deserves a place in the spotlight too.
CoreWeave Watch analysis
What happened
At its Fully Connected conference in San Francisco, CoreWeave said the first customer running production workloads on Vera Rubin NVL72 is Cognition. The company says those workloads delivered more token throughput than on the previous generation, though it did not provide a figure in its conference recap. CoreWeave also announced NVIDIA Vera CPU, which it describes as built for the work around AI agents, from sandboxes to tool calls.
In CoreWeave’s conference recap, the company says one rack supports more than 11,000 concurrent environments and that its tests found sandboxes started more than three times faster than on an x86 CPU. Those performance figures are CoreWeave’s claims. The announcements also included the CoreWeave Partner Network, a group of integrations the company says have been tested on its infrastructure under production load, and a partnership with care provider Ennoble Care to run clinical AI.
Our top picks
- Vera Rubin NVL72 reaches production
Cognition is the first customer CoreWeave identifies as running production workloads on NVIDIA’s new system. - A CPU aimed at agent workloads
Vera CPU targets sandboxes and tool calls, the supporting work that can pile up around AI agents. - More than 11,000 environments per rack
CoreWeave says one rack can support this many concurrent environments, a concrete scale claim for agent infrastructure. - Faster sandbox starts in company testing
CoreWeave says its sandboxes start more than three times faster than on an x86 CPU. - A partner network tested under production load
CoreWeave says integrations including CrowdStrike, VAST Data, ClickHouse and Reflection have been tested on its infrastructure. - Clinical AI gets a named deployment partner
Ennoble Care, which CoreWeave says serves about 50,000 patients a year, selected the company to run its clinical AI.
Why it matters
AI infrastructure is not only a contest to put more accelerators in a rack. Agent workloads also need computing capacity for the environments, tools and background tasks that make an agent useful. CoreWeave’s CPU announcement is a bet that this supporting work will become an increasingly important part of the market.
The useful detail is that the company has named a production customer for Vera Rubin, alongside measurable claims about its agent-oriented CPU. The next question is whether those figures hold across workloads beyond CoreWeave’s own tests. A rack can be impressive; what customers can reliably get done with it is the more interesting unit.
Our read
This is a substantial infrastructure story because it joins a new production deployment to a specific hardware pitch for agent workloads. CoreWeave is making a case for a broader AI cloud, not just a place to rent GPUs. Operators should watch for independent performance comparisons and real customer workloads before treating the company’s speed claims as a new baseline.
What to watch
- Whether Cognition or other customers share workload details and comparable performance figures.
- How Vera CPU performs across different agent workloads and against alternative systems.
- Which integrations enter the Partner Network and what customers can actually deploy.
- What Ennoble Care says about the clinical AI deployment and its use in practice.
Discussion spark: As AI agents take on more background tasks, should cloud providers build specialised CPUs and tightly integrated platforms, or is a mix of general-purpose hardware and specialist tools the better bet?
Sources and evidence
- Fully Connected 2026: Superintelligence Happens When Pioneers Connect (2 October 2026, 00:00 UTC)
not affiliated with or endorsed by CoreWeave