Google Cloud announced general availability of Cloud Pathways on Google Kubernetes Engine (GKE) on November 9, 2026. The release covers regions that support the named TPU models: Cloud TPU v5e, TPU v5p, Trillium (TPU v6e), and Ironwood (TPU 7x). Google says its versioned GKE container images support JAX 0.10.1 and later. Google’s announcement sets out the release scope.
How Cloud Pathways coordinates TPU slices
Cloud Pathways uses a single-controller design: one JAX client coordinates multiple TPU slices and presents them as a unified system. In the default multi-controller approach, separate Python processes instead have local views of their devices. Google’s technical overview of Pathways on Cloud describes the runtime’s components, including its resource manager, IFRT proxy, TPU workers, and sidecars.
Google also describes elastic training that uses host-memory snapshots and can reshard work across N−1 slices after a hardware interruption, without crashing the client. That offers a recovery mechanism for interrupted workloads; it is not a promise of zero downtime.
Treat general availability as the release status, and check that your team has the GKE tooling and TPU experience Google’s deployment guide assumes.
What operators need to know
Google’s overview lists Kubernetes tools, the Google Cloud CLI, and enabling the Kubernetes Engine API as prerequisites. It assumes operators already have experience running TPUs on GKE, including multislice deployments. The announcement therefore marks the runtime’s availability, while the guide describes a deployment path for teams with those foundations.
The separate Cloud Pathways integration through Cluster Director for Slurm was in Preview in Google’s November 9 announcement; that status is distinct from the GKE release.