Google has expanded its eighth-generation Tensor Processing Unit (TPU) portfolio with two new AI chip systems, stepping up its challenge to NVIDIA’s dominance in AI computing infrastructure.

The company unveiled TPU 8tx, designed for secure and confidential AI inference, and TPU 8i Max, built for memory-intensive workloads involving large context windows and increasingly sophisticated AI agents.

A TPU is Google’s custom-designed processor that accelerates AI workloads, including training and running large language models, more efficiently than general-purpose chips.

The launch is part of Google’s strategy to expand its custom accelerators beyond its own data centres and attract external customers. Its TPU business now has nearly 100 customers, including frontier AI laboratories, trading firms, retailers and universities.

A major challenge in competing with NVIDIA is overcoming dependence on its CUDA software ecosystem, which has long given its GPUs a competitive advantage. Google is addressing this through broader software compatibility. Last month, it open-sourced TorchTPU, enabling PyTorch code and Hugging Face models to run natively on its chips. It has also introduced an open-source TPU agent suite, currently in preview, supporting coding tools such as Claude Code, Cursor and Google Antigravity.

The TPU 8tx incorporates hardware-level security to protect data and AI models even during processing. This is aimed at enterprises and government agencies seeking to train or fine-tune models using sensitive proprietary information without exposing it.

The TPU 8i Max targets applications requiring extensive memory and processing capacity, including AI agents analysing large software codebases and lengthy videos. Google said it offers twice the computing capacity and memory bandwidth of its predecessor, TPU 8i.

The new systems complement TPU 8t, designed for AI model training, and TPU 8i, optimised for inference and real-time reasoning.

Google is also allowing customers to deploy TPUs directly in their own data centres, besides accessing them through Google Cloud.

The move reflects a broader industry shift towards custom AI chips to reduce dependence on external suppliers. Amazon is pursuing a similar strategy with its Trainium processors through TorchNeuron, which enables PyTorch workloads to run with minimal code changes.