Google 2015 logo (Wikimedia Commons, public domain)
The internal system behind Gemini training is now a core runtime for Google's Cloud AI Supercomputer, open to outside customers on GKE
Google has made Pathways on Cloud generally available on Google Kubernetes Engine (GKE). Google is positioning it as a core runtime for its Cloud AI Supercomputer.
Pathways is the system Google built for training its own large models, including Gemini. For years it stayed in-house. Now outside customers can rent the same plumbing.
What Google actually shipped #
According to the research findings, general availability arrived on November 9, 2026. The release targets teams running large AI workloads on TPUs, Google’s custom chips built for machine learning.
The headline capability is scale with simplicity. Pathways lets a single JAX client orchestrate workloads across hundreds to thousands of TPU chips.
JAX is a machine learning framework popular with researchers. The practical upshot: one program can steer a whole fleet of chips, so engineers don’t have to babysit each machine separately.
Pathways covers several types of work, including:
Batch training and inference, the big scheduled jobs that build and run models.
Interactive workloads, where developers poke at models in real time instead of waiting for a queued job.
Multi-host inference on large models. The research findings cite Llama 3.1 405B as an example, a model too big to run on a single machine.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The tooling around it #
Google is shipping Pathways with documentation and tools rather than a bare runtime. That includes the PathwaysJob API, which gives teams a defined way to submit and manage Pathways workloads.
Customers can spin up Pathways-enabled GKE clusters using XPK or the gcloud command-line tool. The research findings place that availability by mid-November 2026.
Pathways also plugs into features GKE customers may already use, such as TPU node pools and JobSet APIs. Teams already living inside GKE shouldn’t need to rebuild their setup from scratch.
Two reliability features stand out. Suspend-resume resiliency lets a job and pick up again, which matters when a training run can stretch on for a long time.
Elastic training is the other. It is aimed at frontier-scale projects whose compute needs shift during a run.
Background: from internal backbone to product #
Pathways has long served as a critical backbone for Google’s most demanding training jobs. It was the kind of specialized infrastructure that only a company running its own frontier models would bother to build.
Kubernetes is the widely used open-source system for running and coordinating software across many machines. GKE is Google’s managed version of it.
Putting Pathways on GKE marries Google’s internal AI tooling with an environment developers already know.
The research findings frame this as democratizing access to capabilities previously limited to Google’s internal infrastructure.
What this means #
The Llama 3.1 405B example is telling. Google is not pitching Pathways as a Gemini-only tool. Running a large open-weight model from another company signals that Google wants the platform to host whatever customers bring.
For Google Cloud, the strategy is clear enough. By making GKE the production home for large-scale AI work, Google is trying to turn its internal engineering edge into a reason to choose its cloud. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.