01 What happened
On November 9, 2026, Google Cloud announced the general availability (GA) of Cloud Pathways, a distributed machine-learning runtime built by Google DeepMind and used to train and serve flagship models including Gemini. It runs as the core runtime pillar of the AI Hypercomputer on Google Kubernetes Engine (GKE).
02 Key details
- Cloud Pathways unifies multi-slice TPU clusters under a single JAX controller, replacing traditional multi-controller management.
- Available now in all Google Cloud regions that support TPU v5e, v5p, Trillium (v6e), and Ironwood (7x) on GKE.
- Preview integration with Cluster Director extends Cloud Pathways to Slurm clusters.
- Customers processed more than 130 million Cloud TPU chip-hours on Pathways in the last 90 days.
- Elastic training with host-memory snapshots reshaders workloads to N-1 slices during hardware failures without crashing the client.
03 Why it matters
Cloud Pathways gives every Google Cloud customer the same single-controller runtime, colocated Python sidecars, and elastic fault tolerance that power Google’s largest AI workloads, simplifying management of massive accelerator fleets.
04 Who it matters to
AI developers, machine learning engineers and organizations using large-scale TPU infrastructure.
Original source