On November 9, 2026 Google announced two new variants of its tensor processing units (TPUs). The TPU 8tx is built for confidential inference, keeping model weights and data protected inside the chip. The TPU 8i Max combines two inference‑focused TPU 8i chips into a single package, doubling memory, compute and bandwidth. Both chips belong to the eighth generation that Google first introduced in April 2024.
Why the new chips matter
The 8tx adds hardware‑level encryption so regulated firms can fine‑tune or run models on sensitive data without moving that data to another system. Google says the chip can run training and inference together with about half the latency of the earlier TPU 8t, while keeping performance loss minimal. The 8i Max targets the “memory wall” that slows down large‑scale agents. By packing two 8i chips together, it delivers twice the SRAM, HBM bandwidth and compute density in the same rack footprint, letting models handle million‑token context windows.
Broader context
Google’s move follows similar efforts by Microsoft, Amazon, OpenAI and Anthropic to design their own AI‑specific silicon and reduce reliance on Nvidia’s GPUs. Google notes that its TPU business now serves nearly 100 customers, ranging from research labs to retailers, and that four different TPU systems have been released in a single year – a first for the company. The chips are offered through Google Cloud and can also be installed in private data centers.
Why it matters
Enterprises that must keep proprietary data on‑premise can now fine‑tune models without exposing that data to external servers, thanks to the 8tx’s built‑in security. Companies running large language agents will see fewer slow‑downs because the 8i Max supplies more memory and bandwidth in the same rack space.
