Google releases two specialized AI chips for the agentic era — one optimized for training, one for inference

Google unveiled a new generation of Tensor Processing Units (TPUs — custom chips designed to run AI models), split into two specialized versions: one for training (teaching AI models with data) and one for inference (running trained models to answer questions or complete tasks).

Unlike previous TPU designs that tried to do both jobs, splitting them allows each chip to be optimized for its specific workload — the training chip handles the heavy math of learning patterns in massive datasets, while the inference chip runs lightweight, fast operations to serve answers to users. This matches Google's bet that AI agents (software that makes decisions and takes actions on its own) will need different hardware profiles than past AI systems.

For cloud customers and AI teams building products on Google Cloud, this means faster model training pipelines and cheaper inference costs — translating to lower bills for running chatbots, search systems, or automated workflows. For Google itself, faster custom chips strengthen its competitive position against Nvidia, which currently dominates the AI chip market.

Google has begun rolling out these TPUs to select cloud customers; broader availability and pricing details are expected to follow as the hardware moves into production.