Meta details AI storage blueprint at scale

Meta has published a technical blog post describing its AI storage architecture evolution and optimization work for GPU utilization and research velocity.

• Meta says it rebuilt the metadata subsystem into a unified, flat schema backed by ZippyDB, removed the dataplane proxy, and now deploys a regional BLOB-storage stack colocated with GPUs in every AI region.
• The post also says Meta uses a distributed data cache on GPU hosts, a read-plan metadata cache, hedged reads, and dynamic concurrency control; it reports an average cache hit rate of 80% and 1-2 ms metadata access from the read-plan cache.