Meta's Prometheus Using Object Storage for Massive AI Training Jobs
Blocks & Files, Tuesday, October 20th, 2026
Meta moved its gigawatt Prometheus training cluster from file storage to multi-tiered cached object storage to cut GPU idle time.
Meta has described its Prometheus gigawatt-scale AI training cluster in New Albany, Ohio, which runs hundreds of thousands of GPUs and is being scaled toward about a million.
The cluster combines Catalina (GB200 NVL72) and Clemente (GB300 NVL72) platforms, with a Vera Rubin-based platform called San Miguel to follow.
Its storage has migrated from file-based systems to multi-tiered cached object storage, chosen to keep GPUs fed and reduce idle time during massive training jobs.
Blocks & Files details the hardware platforms and how the storage architecture supports checkpointing and data loading at this scale.