Google Unveils Two New TPUs for AI Era

Google introduced two new custom AI chips at Cloud Next 2026: the TPU 8t for large-scale training and the TPU 8i for inference. The split architecture marks a shift from the single-flagship Ironwood approach, targeting the distinct demands of building and running AI models in the agentic era. TPU 8t delivers 2.7x performance-per-dollar over Ironwood for training, networking up to 9,600 chips per pod. TPU 8i offers 80% performance-per-dollar improvement for inference, featuring three times more SRAM and a custom Boardfly ICI network topology that cuts communication hops by up to 50%. Both chips deliver twice the performance-per-watt of the previous generation.