Google ships four TPUs in one year as agents push AI past human scale

Google's eighth-generation TPU lineup grows to four systems in a single year: TPU 8t for large-scale training, TPU 8i for low-latency reasoning, TPU 8tx for confidential inference on regulated data, and TPU 8i Max, which pairs two chips per package to double compute, SRAM, and HBM bandwidth for million-token contexts. Alongside the hardware, Google open-sourced TorchTPU, TPU Sync, and an XLA serving engine, and made Cloud Pathways and GKE dynamic slicing generally available.