A Google Cloud Next-style session covering the full software stack for building, training, fine-tuning, and serving large language models on Google TPUs. The talk walks through four key tools: vLLM on TPU for inference (with paged attention, prefix caching, continuous batching, and a new diffusion-style speculative decoding achieving ~3x speedup), Tunix for post-training and GRPO-based reinforcement learning, MaxText as an open-source JAX/XLA reference implementation for large-scale pre-training, and Torch TPU for running PyTorch workloads on TPUs with minimal code changes. Live demos show serving Gemma 4 31B on 8 TPU chips via vLLM, fine-tuning a 4B model with GRPO using the Food-101 dataset, and a multimodal food-logging agent using tool calling. The session also introduces Kinetic, a Keras-based project that simplifies TPU cluster setup via decorators, and OpenXLA as the underlying open-source compiler powering both JAX and PyTorch backends.