Google unveiled its eighth generation of Tensor Processing Units at Google Cloud Next, introducing two purpose-built chips: TPU 8t for training and TPU 8i for inference. These chips power Google's custom-built supercomputers and have been driving foundation models like Gemini for years. The new architecture addresses the demands of AI agents that must reason through problems and execute multistep workflows in continuous learning loops.

TPU 8t delivers nearly 3x compute performance per pod over the previous generation, scaling to 9,600 chips in a single superpod with 2 petabytes of memory and 121 ExaFlops of compute. TPU 8i targets inference workloads with 288 GB of high-bandwidth memory and 384 MB of on-chip SRAM—3x the previous generation—reducing latency by up to 5x. Both deliver up to 2x performance-per-watt over the prior Ironwood generation.

Early customer results show real gains: Citadel Securities ran demanding workloads 2-4x faster at 30% lower cost with 7th generation TPUs. Google co-designed these chips with DeepMind specifically for modern reasoning models, marking the culmination of over a decade of TPU development.