Google splits its TPU line in two for the agentic era

Google split its TPU roadmap into two chips: TPU 8t for training and TPU 8i for inference.

TPU 8t keeps the 3D torus design and SparseCores for large-scale training workloads.

TPU 8i adds more memory capacity and bandwidth, plus collectives acceleration and the new Boardfly network to cut latency.

The move reflects how training and inference now need different hardware profiles, especially for agentic AI and Mixture-of-Experts systems.