Meta redesigned its storage architecture to stop AI GPUs from stalling, citing storage bottlenecks as a major cause of delays that increase operating costs and slow research. The company operates hundreds of exabyte-scale storage clusters supporting services including Facebook, Instagram, Reality Labs, Meta AI, advertising systems, databases, and internal data warehouses, built on a foundational layer called Tectonic. Meta shifted from traditional file storage toward BLOB storage because massive AI datasets require unified access methods with higher performance. Repeated metadata lookups across layers created latency that interrupted AI training pipelines. Meta rebuilt the metadata subsystem into a unified schema backed by ZippyDB, allowing clients to retrieve data directly from storage servers via an embedded client streaming from the Tectonic layer. It placed regional BLOB-storage systems beside GPU clusters to reduce transfer delays and introduced distributed caching using unused GPU host memory, achieving an average cache hit rate of 80%, with metadata accessible within 1 to 2 ms. Engineers added hedged reads and dynamic concurrency controls, stating that "the new BLOB-storage stack is now capable of serving AI workloads without causing GPU stalls." To address lengthy delays before training began, Meta created multiple cache layers: GPU memory as L1, SSDs inside GPU hosts as L2, and regional flash-backed BLOB storage as L3, while traditional HDD storage remained the authoritative data source.