Google Cloud 将 Pathways Runtime 在 GKE 上正式可用

Google Cloud on November 9, 2026, announced that Cloud Pathways is generally available on Google Kubernetes Engine (GKE), bringing a single-controller runtime for large-scale training and inference to customers. Google said users processed more than 130 million Cloud TPU chip-hours with Pathways on GKE during the preceding 90 days.

The runtime, also called Pathways on Cloud, is a core runtime pillar of Google Cloud AI Hypercomputer. Originally engineered by Google DeepMind to train and serve Google’s frontier models, including Gemini, it is designed for massive, heterogeneous accelerator fleets. The release extends technology used for Google’s largest AI workloads to Cloud customers.

One Controller Across TPU Slices

The software targets the coordination overhead of default multi-controller JAX. In that setup, every TPU host runs a separate Python interpreter with only a local view of its connected devices, while all hosts execute the same single-program, multiple-data program in lockstep. Google says this makes heterogeneous multi-program workloads and resource management across slices complex and brittle as jobs scale.

Cloud Pathways replaces that arrangement with a unified controller built on the open-source Interim Framework Runtime. Setting JAX_PLATFORMS=proxy with pathwaysutils lets one JAX client view and orchestrate multiple large TPU slices spanning thousands of chips as though their devices were connected locally. The controller manages graph evaluation, persistent compilation caching in Cloud Storage and asynchronous dispatch of XLA work across slices.

Colocated Python sidecar containers, with shared-memory support, run compute-intensive tasks such as distributed data loading and checkpointing directly on TPU hosts. Elastic training uses host-memory snapshots to reshard a job onto one fewer slice after a hardware interruption without crashing the client. Interactive mode supports notebooks and terminals, while Shared Pathways Service lets multiple users and jobs draw on a long-running pool of TPU slices.

Production Use and Elastic Recovery

Google Cloud reports that a large consumer technology company is running production batch and reinforcement-learning inference on tens of thousands of TPU v5p, Trillium v6e and Ironwood TPU7x chips. That deployment consumes more than 100 million chip-hours per quarter, and the customer co-designed elasticity primitives with Google as it expands into pre-training.

ZenteiQ is applying the same stack to scientific foundation models. It uses Cloud Pathways on GKE with Cluster Toolkit, colocated sidecars and host snapshotting for elastic MaxText pre-training and single-controller reinforcement learning with MaxText and Tunix across multi-slice Trillium clusters.

Founder Sashikumaar Ganesan said the controller monitors TPU topology and dynamically re-meshes and resizes workloads when a chip or slice goes offline. “Combining Pathways with host snapshotting in production pushes recovery time down to a minimum when disruptions occur, keeping our available TPU capacity fully utilized with very little stall time,” he said.

GKE Availability and Slurm Preview

Cloud Pathways is generally available on GKE in all Google Cloud regions that support Cloud TPU v5e, TPU v5p, Trillium TPU v6e and Ironwood TPU 7x. Google provides versioned container images supporting JAX 0.10.1 and later.

A separate integration with Cluster Director extends Cloud Pathways to Slurm TPU clusters, supporting unified multi-slice JAX execution, elastic training and interactive notebooks. That integration is in Preview, a distinct status from the GKE runtime’s general availability.

For operators running multi-day jobs on high-cost accelerators, the release ties orchestration directly to utilization: one controller coordinates the TPU slices, and snapshot-based recovery is designed to keep the client job alive when an individual slice fails.