Nvidia has begun shipping the Vera Rubin NVL72, a system with 72 GPUs and 36 CPUs. Wait, correcting my output: the line above is the untranslated source. Here is the Korean translation: 엔비디아가 72개의 GPU와 36개의 CPU를 탑재한 시스템 Vera Rubin NVL72의 출하를 시작했다.
Nvidia is shipping the Vera Rubin NVL72, a liquid-cooled AI computing rack with 72 Rubin GPUs and 36 Vera CPUs. The system is built for large-scale AI training.
The NVL72 is a rack-scale system rather than a standalone processor. Its GPUs and CPUs connect within a single NVLink domain, Nvidia's high-speed interconnect. Nvidia called the rollout "the fastest product ramp in the company's history."
As of Nov. 16, the Vera Rubin platform delivered up to twice the per-GPU performance of its predecessors in MLPerf Training v6.1, an AI training benchmark suite. It led in six of seven categories. The performance comparison is per GPU, while the category tally counts results across the suite.
Deployment has also advanced through Nvidia's partners. Supermicro began shipping integrated liquid-cooled racks on Sept. 23, CoreWeave announced early availability on Sept. 30, and Oracle Cloud Infrastructure did so on Oct. 27. CoreWeave has put NVL72 systems into production and opened customer access through its cloud platform. Its production rollout made Cognition the first customer to run production workloads on the platform.
Cognition also produced a separate performance comparison. Testing software-engineering AI inference tasks, it measured total token throughput and recorded up to 4.8 times the throughput of Nvidia's GB200 NVL72. This was a different metric and comparison from MLPerf's per-GPU training results.
Concerns about AI safety had been raised earlier this fall. In an essay published Sept. 12, Anthropic CEO Dario Amodei called for slower advances in the most powerful AI models, warning that such systems could enable cyberattacks and become harder for their creators to control. OpenAI CEO Sam Altman voiced support for Amodei's position that day. OpenAI paused its largest planned frontier training run on Sept. 20 and resumed it on Oct. 1 after adding security measures.