Google Cloud brings DeepMind's Pathways runtime to general availability on GKE

Google Cloud announced on November 9, 2026, that Cloud Pathways, the distributed machine learning runtime originally built by Google DeepMind to train and serve models such as Gemini, is now generally available on Google Kubernetes Engine (GKE), according to Google Cloud's announcement. The release gives external customers access to the same single-controller runtime DeepMind uses internally for training and serving frontier models on Cloud TPUs.

Key details of the GA release

• Cloud Pathways reached general availability on GKE on November 9, 2026.
• It uses a single-controller architecture built on the open-source Interim Framework Runtime (IFRT), instead of traditional multi-controller JAX setups.
• It supports Cloud TPU v5e, TPU v5p, Trillium (TPU v6e), and Ironwood (TPU 7x).
• It supports elastic fault tolerance with N-1 slice recovery, so jobs survive hardware failures without crashing.
• It adds colocated Python sidecar containers with shared-memory support for custom data loading.
• It includes an interactive mode and a Shared Pathways Service for debugging large distributed jobs from notebooks or terminals.

What it means for AI builders

Pathways was built inside Google DeepMind to handle the orchestration, fault tolerance, and debugging that Gemini-scale training and inference require, and turning it into a GA product on GKE puts that same tooling in front of any team running large distributed workloads on TPUs. Google Cloud reported that more than 130 million Cloud TPU chip-hours moved through Pathways on GKE in the preceding 90 days, and said one major consumer-tech customer uses more than 100 million chip-hours per quarter on it for inference, a sign of the production scale the runtime already handles.

Availability and next steps

Cloud Pathways is available immediately on GKE for teams running JAX 0.10.1 or later, with container images published at us-docker.pkg.dev/cloud-tpu-v2-images/cloud-pathways. Google Cloud also extended preview availability to Slurm clusters managed through Cluster Director, giving teams that orchestrate TPU jobs outside Kubernetes an early path to the same runtime.

Sources

• Cloud Pathways, a distributed ML runtime, is GA (Google Cloud, November 9, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.