CoreWeave claims first working Nvidia Vera Rubin NVL72 rack in race with Microsoft

CoreWeave says it has the first fully operational Nvidia Vera Rubin NVL72 server rack running, and Dell Technologies is backing that claim directly. Specifically, Michael Dell confirmed the deployment on LinkedIn, describing the hardware as the Dell Technologies PowerEdge XE9812, a fully liquid-cooled system built in partnership with CoreWeave for next-generation AI workloads.

The announcement, however, comes with an asterisk worth reading carefully. Back in March, Microsoft stated it was the first cloud provider to bring up an Nvidia Vera Rubin NVL72 system, though that claim was framed around validation rather than full operational deployment. CoreWeave, by contrast, is now claiming the first fully working rack. The distinction between validation and operational use is a fine one, and both companies clearly want the milestone on their record.

So what exactly is the Vera Rubin NVL72? At its core, it is Nvidia’s rack-scale system packing 36 CPUs and 72 GPUs into a single unit. The Vera CPU succeeds Nvidia’s Grace, while the Rubin GPU takes over from Blackwell. According to Nvidia, Rubin delivers roughly five times the inference performance and three and a half times the training performance of the Blackwell generation it replaces. Those are significant jumps on paper, and as a result, AI infrastructure operators have been watching the production timeline closely since Jensen Huang confirmed full chip production began in January 2026.

Beyond raw compute, the NVL72 also introduces a design change that anyone who has spent time in a data center will appreciate immediately. The system runs entirely on liquid cooling and uses a cable-free modular tray design. Furthermore, Nvidia claims that shift cuts installation time from around two hours down to five minutes per unit. For operators deploying racks at scale, that difference compounds quickly across a large build-out.

The broader Vera Rubin platform additionally brings in several other components alongside the CPU and GPU, including the NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet Switch. Together, these form an interconnected system designed to handle memory-intensive, high-throughput workloads that modern AI training and inference demand.

For CoreWeave, consequently, being first to operational status on a new Nvidia generation carries real commercial weight. Cloud customers shopping for AI compute consistently notice which providers have the latest hardware running and which are still waiting.