As of Ray 2.55, Google Cloud TPUs are a first-class accelerator in Ray with official pre-built images and support across core libraries. A TPU slice is a fixed group of chips connected via ICI (Inter-Chip Interconnect), and all workers must land on one intact slice or distributed operations hang. GKE provisions slices and labels hosts via the Ray TPU webhook, while Ray Core's `slice_placement_group()` API atomically reserves a whole slice. The Ray Operator add-on for GKE installs KubeRay and the TPU webhook, letting you declare a topology in a RayCluster manifest without hand-writing placement code. In practice, Ray AI libraries (Train, Serve, Data) call `slice_placement_group()` internally, so users only need to declare a topology. Part 2 will cover using Ray AI libraries on TPU for LLM serving with vLLM, data feeding with Ray Data, and training with JaxTrainer.
