Google Cloud TPU v6e / v7 Pods (contract)

Pod-scale AI accelerators featuring Google's TPU v6e Trillium with 32 GB of HBM per chip in pods of 256, and TPU v7 Ironwood with 192 GB of HBM3e per chip at 7.4 TB/s in pods of up to 9,216 chips delivering 42.5 exaflops of FP8 compute per pod.

Designed for training and serving frontier models at a scale no single rack reaches, with JAX, PyTorch/XLA, and vLLM support, Google's Pathways runtime for multi-pod jobs, and Vertex AI for managed serving, on the same silicon that trains Gemini.

Lives in Google's data centers with liquid cooling and optical circuit switching between chips; nothing to install, and the trade is contract commitment and data residency inside Google Cloud.

Contract pricing only. Google Cloud committed-use agreements; on-demand v6e from roughly $2.70 per chip-hour for reference.