Cloud Pathways, a distributed ML runtime for massive accelerator fleets, now GA

Cloud Pathways, originally developed by Google DeepMind for training and serving frontier models like Gemini, is now generally available on Google Kubernetes Engine (GKE) as part of the Google Cloud AI Hypercomputer. The runtime supports heterogeneous accelerator fleets, including Cloud TPU v5e, TPU v5p, Trillium (TPU v6e), and Ironwood (TPU 7x), allowing customers to process workloads across multiple TPU generations with a single-controller architecture. Over the last 90 days, leading AI companies and global enterprises have used Pathways on GKE to run more than 130 million Cloud TPU chip-hours, demonstrating its scalability for large-scale AI training and inference.

Cloud Pathways replaces traditional multi-controller setups with a unified single-controller design built on the open-source Interim Framework Runtime (IFRT). This architecture enables a single JAX client process to orchestrate thousands of TPU chips across multiple slices as if they were locally connected, simplifying heterogeneous workflows and cross-slice resource management. The runtime also introduces colocated Python sidecars for compute-intensive tasks like data loading and checkpointing, as well as resilient and elastic training that dynamically reshard workloads during hardware failures without interrupting jobs.

A Preview integration with Cluster Director extends Cloud Pathways to Slurm clusters, supporting unified multi-slice JAX execution, elastic training, and interactive notebooks. This allows customers to run multi-host reinforcement learning, disaggregated inference, and pre-training workloads across Slurm-managed TPU clusters while maintaining the same single-controller benefits. Interactive mode and the Shared Pathways Service further enhance usability by enabling real-time debugging, notebook-based workflows, and shared access to long-running TPU slice pools across multiple users.

Frontier AI labs and global enterprises have adopted Cloud Pathways for mission-critical workloads, including hyperscale inference and scientific foundation model training. Customers like ZenteiQ use Pathways with MaxText and host snapshotting to minimize recovery time during disruptions, reducing manual intervention and maximizing TPU utilization. The runtime is available today in all Google Cloud regions supporting Cloud TPUs on GKE, with versioned container images for JAX 0.10.1+, and in Preview for Slurm TPU clusters via Cluster Director.