Google hands outside customers the engine that trained Gemini

Google flipped the switch on Pathways on Cloud on November 9, 2026, making the system generally available on Google Kubernetes Engine. The company is pitching it as a core runtime for its Cloud AI Supercomputer. The target: teams pushing big AI workloads through TPUs, Google's own machine-learning chips.

Pathways was built to train Google's frontier models, Gemini included, and stayed locked inside the company for years. Now outside customers rent the same plumbing. The headline trick is scale with simplicity — a single JAX client can orchestrate workloads across hundreds to thousands of TPUs, so engineers don't have to babysit each machine separately.

It covers batch training and inference, real-time interactive workloads, and multi-host inference on models too big for one machine. The research findings point to Llama 3.1 405B as the example, an open-weight model from another company. That choice signals Google isn't selling Pathways as a Gemini-only tool.

Google is shipping tooling alongside the runtime. The PathwaysJob API gives teams a defined way to submit and manage workloads, and customers can spin up Pathways-enabled GKE clusters using XPK or the gcloud command-line tool — the research findings put that availability at mid-November 2026. It also plugs into TPU node pools and JobSet APIs, so teams already inside GKE don't have to rebuild their setup.

Two reliability features stand out. Suspend-resume resiliency lets a job pause and pick up again when a training run stretches on for a long time, and elastic training targets frontier-scale projects whose compute needs shift during a run.

For Google Cloud the play is straightforward: make GKE the production home for large-scale AI work and convert an internal engineering edge into a reason to pick its cloud. The research findings frame the launch as democratizing capabilities that used to be limited to Google's own infrastructure.