Google adds confidential AI and high-memory TPU chips

The TPU 8tx, designed with AMD and based on the TPU 8t, is intended to train and run AI models while keeping sensitive data and model weights isolated in hardware, Google said in a blog post. A custom inference network allows it to run reasoning alongside training, including distilling and training smaller models on the fly. Google said that could let businesses train models on proprietary data without moving it between devices.

The TPU 8i Max combines two of the inference-focused TPU 8i chips Google introduced in April into one package. The company says the design doubles memory capacity, processing power and bandwidth, enabling models to work with a broader range of information when answering queries.

Both chips build on Google’s eighth-generation TPUs, introduced in April with increased inference capacity and support for large training workloads. The launch comes as cloud providers and AI companies develop their own specialized chips and software to reduce reliance on Nvidia. Microsoft, Amazon, OpenAI and Anthropic are among those pursuing such efforts.