The tech giant claims its Ironwood tensor processing units (TPUs) are more power- and cost-efficient than previous generations. Google says they’re also particularly well-suited for AI agents—assistants that can carry out tasks for users—as well as large language models (LLMs) with reasoning capabilities. (The Logic)

Talking point: Most AI developers—including Google itself—still train and deploy their models on Nvidia’s super popular semiconductors. But cloud giants are also building their own in-house chips to supplement what they buy. When used alongside Google’s software and other hardware, the firm’s TPUs offer “the best intelligence per unit cost overall for our customers,” said Amin Vahdat, vice-president of cloud AI, although he wouldn’t provide a direct comparison with Nvidia’s new Blackwell chips. On Tuesday, Google also announced new automation features in its office software for businesses, and a new system for AI agents from different developers to work together.