Cloud provider CoreWeave has integrated the Nvidia Vera Rubin NVL72 system into its infrastructure. The AI company Cognition became the first client to deploy production workloads on this new platform.
In SWE-2 inference tasks, the platform delivered up to 4.8 times higher total token throughput compared to the GB200 NVL72. For reinforcement learning, throughput per GPU increased by 3.8 times, enabling the Devin agent to handle more concurrent sessions and accelerate multi-step inference.
The Vera Rubin NVL72 system features 36 Vera CPUs and 72 Rubin GPUs, utilizing a full liquid cooling setup. The modular rack design reduces the installation time for individual trays from approximately two hours to five minutes.
Energy efficiency metrics in DeepSeek-R1 testing showed a 10-fold increase in token throughput per megawatt compared to the GB200 NVL72. CoreWeave also offers Vera CPUs separately, with up to 128 processors per rack providing 11,264 cores.
This CPU configuration supports over 11,000 isolated environments simultaneously. According to CoreWeave, launching agent sandboxes on these processors is more than 3 times faster than on standard x86 architectures.
