Google Cloud launches 15 TB/s tier

Google Cloud said on May 11 that it launched Cloud Storage Rapid, a new family of object-storage products aimed at AI and analytics workloads that need higher throughput and lower latency than standard cloud object storage. The launch followed announcements at Google Cloud Next ’26 in April, where the company said storage had become a bottleneck for training and inference jobs running on large GPU and TPU clusters. Google said the new family starts with two products: Rapid Bucket, a zonal object-storage offering, and Rapid Cache, a read cache for existing buckets. The company positioned the products as a way to keep accelerators busy by moving data closer to compute. ### What exactly did Google launch? Cloud Storage Rapid is not a single SKU but a product family inside Google Cloud Storage, according to Google’s May 11 blog post and documentation. Google said the initial lineup includes Rapid Bucket, formerly called Rapid Storage, and Rapid Cache, formerly called Anywhere Cache. (cloud.google.com) Rapid Bucket is a zonal bucket type that stores data in the Rapid storage class, Google’s documentation said. Rapid Cache is an SSD-backed zonal read cache that works with existing regional, dual-region or multi-region buckets without requiring API changes, according to the same documentation. ### Where does the 15 TB/s figure come from? (cloud.google.com) Google’s documentation said Rapid Bucket delivers sub-millisecond latency, up to 15 TB/s of aggregate throughput and support for up to 20 million queries per second. The company said those capabilities are designed for AI model training, checkpointing, serving and other I/O-intensive workloads. (docs.cloud.google.com) The AI Hypercomputer documentation repeats the same throughput figure and says Rapid Bucket is intended to bring data into the same zones as compute workloads. Google said Rapid Cache, the companion product for read-heavy workloads, provides 2.5 TB/s of aggregate throughput. ### Why is Google emphasizing zonal storage instead of standard buckets? (docs.cloud.google.com) Zonal placement is central to the product. Google said Rapid Bucket lets customers define a zone as a bucket’s location so storage sits alongside AI accelerators, lowering latency and raising throughput relative to other Cloud Storage options. (docs.cloud.google.com) Google also said zonal buckets support appendable objects and streaming reads and writes, features it tied to checkpointing, evaluation, logging, messaging queues and real-time serving. In the company’s description, that is meant to reduce full-object rewrites and make data visible to consuming applications while it is still being written. (docs.cloud.google.com) ### Who is using it? Thinking Machines Lab was named by Google as an early participant. Google said James Sun, a member of technical staff at Thinking Machines Lab, spoke at a Next ’26 session about the high-performance storage needs of the company’s AI and machine-learning workloads. (docs.cloud.google.com) Anthropic was not named in the Google materials reviewed here as an early Cloud Storage Rapid user. Anthropic has an existing relationship with Google Cloud and said in October 2025 that it planned to expand its use of Google Cloud technologies, including up to one million TPUs, but that announcement did not mention Cloud Storage Rapid. (cloud.google.com) ### How does this fit into Google’s broader AI infrastructure push? At Google Cloud Next ’26 on April 22, Google said Cloud Storage Rapid was part of a broader set of storage announcements that also included a Dynamic tier for Google Cloud Managed Lustre and new management and metadata features. Google said those changes were aimed at training, inference and agent workloads that need faster access to large data sets. (anthropic.com) Google tied the storage launch to its wider AI Hypercomputer stack, where documentation now compares Cloud Storage Rapid with other storage services for model training and inference. That material says customers should choose among object, cache and parallel file-system options based on throughput, latency, concurrency and scale requirements. (cloud.google.com) ### What can customers check next? Google’s May 2026 documentation says Rapid Bucket is generally available and points customers to setup pages for creating zonal buckets, using objects in zonal buckets and reviewing supported locations. Google also said customers can explore the features through Cloud Storage Rapid documentation and Google Cloud Next ’26 storage sessions. (cloud.google.com) (docs.cloud.google.com)