CoreWeave has made Nvidia’s Vera Rubin NVL72 rack-scale platform available in its cloud, with AI software company Cognition as the first customer running production workloads on the system. The company also announced support for Nvidia’s Vera CPU, aimed at the growing CPU-side demands of agentic AI.investors.coreweave+1
First Vera Rubin customer
Cognition, the developer of the Devin AI software-engineering agent, deployed a Vera Rubin NVL72 cluster on CoreWeave in early September. The company then ran what CoreWeave describes as the first customer-executed inference benchmark on the platform.investors.coreweave+1
In early testing, Cognition reported up to a 4.8x increase in total token throughput for SWE-2 inference workloads compared with an Nvidia GB200 NVL72 baseline. It also measured a 3.8x increase in output-token throughput for reinforcement-learning workloads.investors.coreweave+1
These results come from Cognition’s benchmarks on CoreWeave Cloud, not from an independent third-party performance audit. Still, they provide an early indication of how Vera Rubin systems may perform for agent-heavy workloads that repeatedly read code, reason over long contexts, and execute tasks.
Why agentic AI needs more than GPUs
CoreWeave says agentic workloads place unusual demands across the entire infrastructure stack. AI agents may maintain long conversation contexts, issue repeated model calls, run tools, execute code, and manage thousands of simultaneous tasks.
Chen Goldberg, CoreWeave’s executive vice president of product and engineering, said the company’s goal is to make compute, networking, and software operate as one system. The company argues that higher throughput can allow more concurrent agent sessions per GPU, faster research cycles, and lower cost per session without sacrificing generation speed.investors.coreweave
CoreWeave already operates Nvidia GB200 and GB300 NVL72 systems, including Dell PowerRack configurations. The company says existing customers can use Vera Rubin capacity with the same tooling and operational processes they use for those earlier systems.investors.coreweave
Vera CPU for agent environments
Alongside Vera Rubin, CoreWeave is adding support for Nvidia’s Vera CPU. While GPUs handle most model training and inference, CPUs support surrounding tasks such as isolated execution environments, reinforcement-learning jobs, tool calls, code execution, and data pipelines.coreweave
CoreWeave’s initial Vera deployment uses rack-scale configurations with:
128 Vera CPUs.
11,264 CPU cores per rack.
Nvidia BlueField-4 DPUs.
Spectrum-X Ethernet switching.
CoreWeave says this configuration can support more than 11,000 concurrent agent environments, assuming one core per environment.coreweave+1
A platform play for AI infrastructure
The announcements show CoreWeave positioning itself as more than a GPU rental provider. By combining rack-scale Nvidia systems, high-speed networking, and software designed for agent workloads, the company is trying to reduce the operational work required to move each new Nvidia generation into production.
For customers building autonomous software agents, coding assistants, or reinforcement-learning systems, the key question is whether the infrastructure can sustain many simultaneous, long-running tasks efficiently. CoreWeave’s Vera Rubin and Vera CPU offerings are designed to address both the GPU-intensive and CPU-intensive sides of that workload.coreweave+1
The early Cognition results are promising, but broader adoption will depend on pricing, availability, workload compatibility, and how the performance gains hold up across different types of AI applications.
