CoreWeave deploys multi-rack NVIDIA Rubin NVL72

CoreWeave announced on 16 September 2026 the launch of multi-rack NVIDIA Vera Rubin NVL72 clusters on its cloud platform. The infrastructure links hundreds of NVIDIA Rubin GPUs into a single scale-out environment designed for agentic artificial intelligence training and inference.

A single NVL72 rack combines 72 Rubin GPUs with 36 Vera CPUs, NVLink 6, ConnectX-9 SuperNICs, and BlueField-4 DPUs. CoreWeave links these racks with NVIDIA Spectrum-X Ethernet networking. Each Rubin GPU features two ConnectX-9 SuperNICs that provide 1.6 Tb/s of scale-out bandwidth, supporting roughly 128,000 GPUs per rail in a non-blocking fabric.

The company also launched two updates for CoreWeave AI Object Storage. A cross-region write acceleration feature writes data locally while replicating it to a remote region in the background, avoiding pauses during checkpoint creation. In addition, a new Archive tier cuts long-term retention costs without charging fees for retrieval, early deletion, or reads.

CoreWeave noted that its Local Object Transport Accelerator caching system cuts latency by 8x compared to standard storage clusters while delivering up to 7 GB/s of throughput per GPU.