Google has officially launched its Pathways system on Google Kubernetes Engine (GKE), making it available to external cloud customers. The platform serves as a core runtime for the Google Cloud AI Supercomputer, enabling users to train and run large-scale machine learning models.
The release, which became generally available on November 9, 2026, is designed for teams managing heavy AI workloads on Google's custom TPU chips. Pathways allows a single JAX client to orchestrate tasks across hundreds or thousands of chips, simplifying the management of large-scale operations.
The system supports batch training, interactive workloads, and multi-host inference for massive models like Llama 3.1 405B. Google provides the PathwaysJob API to help teams manage their workloads, with full availability for GKE clusters via XPK or the gcloud tool by mid-November 2026.
Pathways integrates with existing GKE features, including TPU node pools and JobSet APIs, allowing teams to utilize their current infrastructure. New reliability features include suspend-resume resiliency for long training runs and elastic training for projects with fluctuating compute requirements.
