Google launched 2 AI chips, 1 built to protect model weights while in use.

Google published "Introducing TPU 8tx and TPU 8i Max" on blog.google on November 9, 2026, with page metadata of 13:00 Universal Time, and we read it there. The author is Amin Vahdat, Senior Vice President and Chief Technologist for AI and Infrastructure. TPU 8tx is described as the third eighth-generation TPU system, aimed at secure, confidential inference and also suited to small-scale training. It uses the same silicon as TPU 8t, paired with 5th Gen AMD EPYC central processing unit hosts. It "adds hardware-level confidential compute that protects data and model weights while in use." It uses the Boardfly network from TPU 8i, which Google says gives up to 2 times lower latency than TPU 8t with "a near-zero performance penalty." Google says it supports fine-tuning, distillation and reinforcement learning on one system, without moving sensitive data between clusters. TPU 8i Max combines 2 TPU 8i chips onto a single package, which Google says doubles the compute, the static memory and the high-bandwidth memory bandwidth of TPU 8i. It fits twice the silicon in the existing rack footprint, which Google calls the highest compute density per rack in the portfolio. Other figures in the post belong to the earlier systems, not to these 2: the 2.7 times training and 80% inference price-performance gains are TPU 8t and 8i against the previous generation, and the "up to half" cut in inter-chip communication delay is TPU 8i. Google says TPUs have nearly 100 customers, including frontier laboratories, trading firms, retailers and universities, and names none for these systems. Not published anywhere in the post: absolute throughput, high-bandwidth memory capacity, static memory size, bandwidth figures, pod or cluster sizes, chips per rack, power or thermal numbers, any training figure for 8tx against 8t, any inference benchmark specific to either system, the baseline for the 2 times latency claim, how "near-zero" was measured, how the confidential compute is attested, availability dates, regional availability, pricing, and whether 8tx can run inside a customer's own data centre. Alphabet shares rose more than 2% before the New York open on the announcement, at $293.04, on a market value of $3.59 trillion.

1. Apr 23

   Google announces its eighth-generation TPU as 2 chips for the first time, TPU 8t for training and TPU 8i for inference, at Cloud Next 2026, on a 2 nanometre process, with networks scaling to 1 million TPUs per cluster.

2. Oct 27

   Google agrees to put its own AI chips inside a rival's cloud for the first time, with Oracle. Edition 064.

3. Oct 30

   Google runs a model on a data centre chip in orbit and gets an answer back. Edition 067.

4. Nov 5

   Alphabet's quarterly holdings filing, accepted at sec.gov at 23:21 Universal Time, shows a $548 million Qualcomm stake acquired through an AI acquisition, as Qualcomm targets AI data centres.

5. Nov 9

   Google publishes TPU 8tx and TPU 8i Max at 13:00 Universal Time, its third and fourth TPU systems of the year. No price, no availability date, no performance figure for either.