Google Cloud Storage Rapid Delivers Zonal Object Storage and Ingest Caching for AI Workloads

Google Cloud has released Cloud Storage Rapid, a suite of performance-oriented object storage features tailored for artificial intelligence, machine learning, and data-intensive analytics. The offering includes Rapid Bucket, a zonal object storage class built on Google's Colossus distributed storage architecture that delivers sub-millisecond access latency, up to 15 TB/s aggregate read throughput, and up to 20 million queries per second (QPS). It also incorporates Rapid Cache, an SSD-backed caching tier designed to accelerate read bandwidth on existing regional and multi-region buckets without requiring code or API modifications, alongside stateful streaming object appends and ingest-on-write capabilities. For AI practitioners and infrastructure architects, storage performance is directly tied to GPU economics. In distributed training clusters, accelerators frequently stall while waiting for batch reads or writing intermediate state checkpoints to remote storage. Google reports that Rapid Bucket reduces blocked GPU time by 50%, accelerates multimodal data loading by up to 2.5x, and delivers up to 5x faster checkpoint restores alongside 3.2x faster checkpoint writes compared to standard object storage tiers. Crucially, native support for object appends removes the traditional requirement to rewrite immutable objects completely when updating logs or stateful streaming data. This release reflects a decisive industry shift away from treating object storage merely as a cold, decoupled target for data lakes and archives. As models scale and parameter checkpoints swell into tens of terabytes, hyper-scalers are redesigning object storage engines around zonal locality and hierarchical namespaces. By blending block-like sub-millisecond latency and parallel file-system throughput with standard S3/GCS API compatibility, cloud providers are attempting to eliminate the operational overhead of managing dedicated scratch filesystems like Lustre for intermediate training stages. In practice, engineering teams should evaluate Rapid Buckets for latency-critical training runs, checkpoint staging, and high-concurrency model serving. However, because Rapid Buckets are bound to a single availability zone, architects must factor in zonal durability trade-offs and implement automated lifecycle policies to drain checkpoints or finished datasets into regional or multi-region buckets. Workloads with existing large datasets should test Rapid Cache first, enabling immediate read throughput improvements and caching without refactoring data pipelines or bucket configurations.