Overview
- CoreWeave said Wednesday that it has put seven Vera Rubin NVL72 racks into production across two regions, a 504‑GPU multi‑rack cluster that the company says is the first validated deployment of Nvidia’s NVL72 at scale.
- Each NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs and uses Nvidia NVLink 6, ConnectX‑9 SuperNICs and BlueField‑4 DPUs, with ConnectX‑9 quoted as offering up to 1.6 Tb/s of scale‑out connectivity per GPU and an architectural target of roughly 128,000 GPUs per rail.
- During validation CoreWeave forced traffic across the backend network by disabling direct GPU‑to‑GPU links to expose multi‑rack networking faults and ran diagnostics and distributed workloads before promoting systems to production.
- The company launched CoreWeave AI Object Storage updates including the Local Object Transport Accelerator (LOTA), cross‑region write acceleration and an Archive tier, with LOTA billed as reducing read latency up to 8x and delivering up to 7 GB/s per GPU under favorable conditions.
- Markets reacted modestly with CoreWeave shares rising about 3% in premarket trading and Nvidia edging up roughly 1%, while the technical milestone strengthens CoreWeave’s competitive positioning even as prior coverage notes rapid revenue growth, widening losses and material refinancing and debt pressures.