Google announced Cloud Pathways as generally available on Google Kubernetes Engine (GKE) on November 9, 2026. The announcement covers GKE; Google described integration for Slurm TPU clusters through Cluster Director separately as Preview. For teams considering the GKE setup, Google’s deployment guide lists command-line tools, Kubernetes Engine API access and sufficient TPU quota or a reservation—and warns that the example creates billable resources. (Google’s announcement; GKE deployment guide)

How Cloud Pathways coordinates TPU slices

Cloud Pathways, also called Pathways on Cloud, uses a single-controller model: one JAX client coordinates work across multiple TPU slices and presents them through a unified device view. In default multi-controller JAX, separate Python processes participate, and each sees its locally connected devices. (Google Cloud technical documentation)

Operating dimensionDefault multi-controller JAXCloud Pathways single-controller
ControlMultiple processes participate, each with its own Python interpreter.One JAX client coordinates work across the system.
Device viewEach process sees its locally connected devices.The client gets a unified view across TPU slices.
Programming patternsPrimarily SPMD: devices run the same program.Also supports patterns such as asymmetric pipeline parallelism and computational sparsity.

For teams whose workload needs one JAX client to coordinate across slices—or programming patterns beyond SPMD—the key difference is how control and device visibility are organized, not a stated performance guarantee.

Google says Cloud Pathways also supports elastic training, interactive workloads, Python sidecars colocated with TPU hosts, persistent compilation caching in Cloud Storage and asynchronous XLA HLO dispatch. Its description of elastic training says host-memory snapshots can let a workload be reshaped to N−1 TPU slices after a hardware failure without crashing the client. (Google’s announcement)

The announced GKE scope covers Google Cloud regions that support Cloud TPU v5e, TPU v5p, Trillium (TPU v6e) and Ironwood (TPU 7x). Google also said its versioned Cloud Pathways container images support JAX 0.10.1 and later. (Google’s announcement)

What Google’s GKE setup requires

For the documented deployment, Google lists the gcloud CLI, kubectl, gke-gcloud-auth-plugin, the Kubernetes Engine API, and enough TPU quota or a reservation for the slices being created. The guide’s manual example also uses CPU and TPU node pools, installs the JobSet API and controller, and uses Cloud Storage for temporary files and compiled-code caching. For its node-metadata configuration, Google says to grant the node service account roles/storage.objectAdmin on the bucket. (GKE deployment guide)

Secure TPU quota or a reservation before creating the example cluster, and delete the cluster when the work is done: its GKE, CPU and TPU resources incur charges.

The guide recommends deleting the cluster after use to avoid continued charges for its resources. It does not give a Pathways-specific price or a total deployment estimate. (GKE deployment guide)