Google Cloud Makes Pathways Runtime GA for Large-Scale TPU Workloads
What happened
On November 9, 2026, Google Cloud announced that its Pathways runtime is now generally available on Google Kubernetes Engine (GKE). This software, originally developed by DeepMind, allows users to manage large-scale TPU workloads through a single controller. The technology has already processed over 130 million Cloud TPU chip-hours in the previous 90 days. The release brings the same infrastructure used for Google's largest AI models to all Google Cloud customers.
Key details
- Pathways runtime is available on GKE in all regions supporting Cloud TPU v5e, v5p, Trillium v6e, and Ironwood TPU 7x.
- A single controller can now coordinate thousands of TPU chips across multiple slices.
- Elastic training and snapshot-based recovery are supported, reducing downtime if hardware fails.
- A large tech company is already running production workloads on tens of thousands of TPUs, consuming over 100 million chip-hours per quarter.
- Integration with Slurm clusters is in Preview, while GKE support is fully available.
Why it matters
Organizations using Google Cloud for large AI workloads can now manage and recover from hardware failures more efficiently, reducing job interruptions and improving resource use. This change benefits companies running multi-day or large-scale training jobs, as they can share TPU resources among multiple users and jobs, and recover quickly from disruptions. Cloud customers now have access to the same scalable AI infrastructure that Google uses internally, enabling more advanced and resilient AI model development.