Google Cloud Storage Rapid Bridges Object Durability and Low-Latency AI Workloads

Google Cloud expanded its object storage ecosystem with the general availability of Cloud Storage Rapid, introducing two purpose-built capabilities for high-performance AI/ML and analytics workloads: Rapid Bucket and Rapid Cache. Rapid Bucket delivers zonal object storage backed by Google’s Colossus distributed storage architecture, offering sub-millisecond latency, up to 15 TB/s of aggregate throughput, and up to 20 million queries per second within a single zone. Rapid Cache provides an SSD-backed zonal read caching layer yielding up to 2.5 TB/s read throughput over standard multi-region or regional buckets, enhanced with an ingest-on-write mechanism that populates the cache concurrently during object writes. For infrastructure architects and MLOps practitioners, this release fundamentally alters the economics and complexity of AI training infrastructure. Modern multi-node GPU clusters often experience significant blocked time when reading multi-modal training datasets or executing frequent model checkpointing. Traditionally, teams mitigated these I/O stalls by deploying dedicated, high-maintenance parallel file systems alongside standard object stores. Rapid Bucket reduces GPU blocked time by up to 50% and accelerates checkpoint writes by 3.2x, while Rapid Cache accelerates checkpoint restores by up to 2.2x. This keeps high-cost accelerators saturated without requiring engineers to manage separate storage tiers. This development reflects a decisive industry-wide transformation: cloud object storage is evolving from a passive, high-latency archive into an active, high-throughput compute engine tier. Similar to architectures like AWS S3 Express One Zone, the shift toward zonal object namespaces and co-located storage nodes acknowledges that high-concurrency AI workloads cannot tolerate cross-regional network hops or traditional HTTP request latency. Furthermore, by integrating native gRPC protocols and streaming append operations, object stores are absorbing features previously restricted to block and high-performance file systems. In practice, platform teams should evaluate workload characteristics before transitioning storage tiers. Rapid Bucket requires evaluating availability tradeoffs; because it is pinned to a specific zone, disaster recovery strategies must explicitly manage replication to regional or multi-region tiers. For existing architectures where refactoring bucket namespaces is impractical, enabling Rapid Cache provides an immediate, transparent throughput boost with zero code modifications. Teams should monitor cache hit ratios and configure appropriate Time-to-Live (TTL) parameters to balance performance gains against cache storage costs.