• CoreWeave brought up a multi-rack NVIDIA Vera Rubin NVL72 cluster, linking hundreds of Rubin GPUs over Spectrum-X Ethernet
  • The company also added cross-region write acceleration and an Archive tier to AI Object Storage

The fact

CoreWeave said on 16 September that it had brought up a multi-rack NVIDIA Vera Rubin NVL72 cluster on CoreWeave Cloud, connecting hundreds of Rubin GPUs through NVIDIA Spectrum-X Ethernet. Each NVL72 rack combines 72 Rubin GPUs with 36 Vera CPUs, alongside NVLink 6, ConnectX-9 SuperNICs and BlueField-4 DPUs.

CoreWeave also announced two changes to AI Object Storage. Cross-region write acceleration allows applications to complete writes locally while replication to another region continues in the background. A new Archive tier is intended for data that needs to be retained but accessed less frequently.

The company described automated rack validation covering hardware detection, firmware, power, cooling and workload testing before systems enter production. It did not disclose the number of Rubin clusters deployed, total available GPU capacity or how much capacity customers can currently reserve.

The assessment

Bringing hundreds of GPUs into one cluster increases the amount of compute a customer can put behind a workload, but that capacity is useful only when data can reach it quickly enough. Training, inference and agentic applications can repeatedly read, write and move large datasets, so storage and network delays can leave expensive accelerators waiting.

Cross-region write acceleration changes one part of that path. A workload can finish its local write without waiting for the remote copy to complete, while CoreWeave handles replication afterwards. That can reduce the time an application spends waiting on cross-region storage, although customers still need to understand when the second copy becomes available and where it is stored.

For BTW readers, the Rubin announcement is therefore more than a GPU milestone. CoreWeave is also working on the storage path around those accelerators, because faster compute delivers less value when jobs repeatedly stop for data movement.

What to watch

Watch for customer deployments that disclose reserved Rubin capacity and performance on named workloads. Storage results should include cross-region write latency, replication time and pricing for the new Archive tier. Those figures would show whether the storage changes reduce waiting time and cost alongside the increase in available compute.