Google Cloud Storage Rapid Cache Brings 2.5 TB/s SSD Caching to Object Stores

Google Cloud has updated its high-performance object storage architecture with Rapid Cache (formerly Anywhere Cache), a fully managed, SSD-backed zonal read cache designed to accelerate data retrieval from Cloud Storage buckets. The capability co-locates cached object data within the exact availability zone of the client compute instances, achieving up to 2.5 TB/s in read throughput. Bandwidth autoscales dynamically, beginning at a baseline of 100 Gbps and growing at 20 Gbps per 1 TiB of stored cache data. The service provides bucket-level and prefix-level ingest-on-write mechanics alongside customizable Time-to-Live (TTL) policies, ensuring objects are pre-warmed in local SSDs upon write. Modern AI/ML model training runs, high-throughput inference engines, and parallel analytical query engines overwhelmingly rely on cloud object storage as their primary data lake due to virtually unlimited scale and durability. However, pulling massive model checkpoints or training datasets repeatedly across distributed nodes often hits bandwidth ceilings and incurs substantial data transfer expenses when accessing multi-region buckets. Rapid Cache changes this dynamic by converting standard object storage into a low-latency, zonal read tier. By serving read-heavy assets locally from SSDs, infrastructure teams can drastically reduce time-to-first-byte latency while cutting cross-region network transfer bills. This release aligns with an industry-wide transformation of cloud object storage from passive cold archives into active, high-throughput engines for compute-heavy workloads. As AI model sizes expand and GPU clusters require constant data feeding to prevent idle cycles, the boundary between object storage and high-speed scratch storage is dissolving. Rather than forcing engineers to deploy, manage, and patch complex external distributed caching tiers or expensive shared file systems, cloud providers are integrating caching and low-latency access patterns directly into the native object store plane. In practice, platform engineers and architects should selectively apply Rapid Cache to workloads characterized by frequent reads of relatively static assets, such as base model weights, stable training partitions, or reference datasets. Teams should configure prefix-level ingest-on-write rules to warm only active directories, avoiding unnecessary caching costs on transient outputs. Furthermore, teams must account for architectural nuances: because object metadata operations bypass Rapid Cache and resolve against the underlying Cloud Storage bucket, optimization strategies should focus on streaming large binary blobs rather than heavy metadata-listing loops.