Google TPUs move from 'experimental' to first-class in Ray 2.55

Ray 2.55, released on the Google Developers Blog on July 20, 2026, promotes Google's TPUs (Tensor Processing Units) from an experimental, bring-your-own-container path to a first-class accelerator. The same Ray APIs that teams already use for tasks, actors, and Ray Serve deployments now officially target TPU slices on Google Kubernetes Engine.

For platform engineers, the new thing to plan around is the slice. A TPU slice is a fixed group of host VMs whose chips share a dedicated high-speed interconnect called the ICI (Inter-Chip Interconnect). A multi-host training job has to land on one whole slice, the way a multi-GPU job has to fit on one NVLink-connected box. If workers end up split across slices, their collective operations (all-reduce) never finish, and the job hangs. Ray plus GKE handle gang-scheduling onto an intact slice.

Users request topology, the shape of a slice such as 4x4 for 16 chips, rather than a raw chip count. Google is calling this post Part 1 of a series, with Part 2 set to walk through each Ray library on TPU. The post does not address independent benchmarks or production cost comparisons, so those questions are left for teams to measure on their own workloads.