CoreWeave Validates Multi‑Rack Nvidia Vera Rubin NVL72 Deployment

Overview

• CoreWeave disclosed Wednesday that it has put seven Nvidia Vera Rubin NVL72 racks into production across two regions, giving customers access to 504 Rubin GPUs in a single scale‑out configuration.
• The company says the hardest work was networking and operations because stitching NVL72 racks shifts traffic, congestion control and fault isolation from the vendor to the operator.
• CoreWeave launched AI Object Storage features including the Local Object Transport Accelerator (LOTA) for a local NVMe cache, cross‑region write acceleration for background replication, and an Archive tier; the firm claims LOTA can cut read latency up to eightfold and deliver as much as 7 GB/s per GPU.
• To manage power, cooling and failures at rack scale CoreWeave extended its stack with Racky, Valvey and a Rack LifeCycle Controller and it deliberately forces traffic over the backend network during tests to expose switch, cabling and software faults.
• The announcement gave CoreWeave and Nvidia modest stock gains and reinforces CoreWeave’s first‑mover claim, but prior reporting warns that heavy cash burn, debt and lender risks remain material to the company’s ability to scale beyond these initial multi‑rack deployments.