NVIDIA says Vera Rubin NVL72 is ramping with partners and 10x tokens per megawatt vs Blackwell

NVIDIA said its Vera Rubin NVL72 rack-scale AI platform is ramping into production, with systems running at CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure. The company described a supply chain spanning more than 350 factory sites in 30 countries and said the platform is co-designed across seven chips and five rack trays. CoreWeave reported a DeepSeek-R1 benchmark on live Vera Rubin NVL72 hardware showing 10x more throughput per megawatt than Grace Blackwell NVL72. NVIDIA and partners framed tokens per megawatt as the metric that matters for power-constrained AI factories. Google Cloud said A5X bare-metal instances built on Vera Rubin NVL72 are running for London startup Ineffable Intelligence, with claimed up to 10x lower inference cost per token and 10x higher token throughput per megawatt than the prior generation. NVIDIA also said Vera Rubin underpins an expanded Microsoft and Mistral partnership for European AI infrastructure under a multibillion-dollar agreement. Mistral is adding GPU capacity drawing on thousands of Vera Rubin GPUs. Mistral Medium 3.5 and OCR 4 are available in Microsoft Foundry, with models integrated into Copilot Studio. DeepInfra benchmarks cited by NVIDIA showed the Vera CPU supporting up to 1.6x more concurrent AI agents and up to 2.2x faster orchestration than alternative CPUs. The company said NVL72 trays have no cables, fans, or hoses, cutting compute tray assembly time from hours to one minute, and that a 45-degree Celsius liquid cooling inlet design enables chiller-free dry-cooler operation intended to cut water use at new AI factories.