Google Drops Two New TPU Chips for AI Training and Inference

Google Cloud just pulled the curtain back on its next-generation tensor processing units — and this time, it's a two-chip affair.

The new lineup features the TPU 8t, purpose-built for AI training workloads, and the TPU 8i, optimized specifically for inference. It's a deliberate split that signals Google is done trying to make one chip do everything.

Both chips are homegrown silicon designed in-house by Alphabet's Google Cloud division. They represent the latest evolution of Google's TPU architecture, which has been central to the company's AI infrastructure strategy for years.

General availability for the new TPU generation is slated for later in 2026. That timeline puts Google in direct competition with Nvidia and other custom silicon efforts from rival cloud providers during a critical buildout period for AI infrastructure.