CoreWeave said on Wednesday that it has joined several Nvidia Vera Rubin NVL72 racks into one scale-out cluster on its cloud, putting hundreds of Rubin GPUs behind a single system.
The contents of each rack are substantial. There are 72 Rubin GPUs and 36 Vera CPUs inside. Networking comes from NVLink 6 within the rack, plus ConnectX-9 SuperNICs and BlueField-4 DPUs. CoreWeave stitches racks together over Spectrum-X Ethernet.
Agentic work is the target. A single task fans out into many model calls and tool invocations, and storage latency compounds at each hop. Each GPU therefore carries two SuperNICs for 1.6 Tb/s of scale-out bandwidth, and CoreWeave says the fabric supports roughly 128,000 GPUs per rail without a redesign as racks get added.
Operations run through Mission Control, whose Rack LifeCycle Controller covers detection, firmware, validation, power and cooling. Racks go through full-rack workload tests against Nvidia field diagnostics and only enter production once the whole system passes.
Storage changed as well. Cross-region write acceleration lands checkpoints locally while replicating them elsewhere in the background, with applications seeing one bucket. A new Archive tier charges nothing for retrieval, early deletion or reads, aimed at checkpoints and datasets teams would otherwise throw out.
CoreWeave pointed to record MLPerf runs and top ClusterMAX rankings from SemiAnalysis as proof it can run the hardware at full tilt.
