Google TPU 8tx and 8i Max: What's New, What Isn't

Google says this is the first time it has introduced four TPU systems in a single year. The count is accurate. “New” is the part to question.

On Monday, November 9, Google announced the TPU 8tx and the TPU 8i Max. Neither is a fresh piece of silicon. Google’s own post calls them “systems,” and the 8tx is built on the same chip as the TPU 8t, which Google showed at Cloud Next in April. The 8i Max takes two TPU 8i chips from the same April launch and puts them in one package.

What the 8tx actually adds

Data is usually encrypted on disk and on the network, but it has to be unlocked in memory while a processor works on it. That gap is what confidential computing is meant to close. According to Google, the 8tx keeps data and model weights “protected and isolated even while in active use,” and it’s aimed at governments and regulated companies running sensitive workloads.

Network World reported that Google Cloud’s Lohmeyer described the whole system as attested, meaning it can prove it’s what it says it is, and that data is encrypted from the host to the TPUs. Chosun Biz, citing a Google Cloud briefing in Korea, reported the same encryption path across the CPU, the TPU and its memory.

Google also claims up to 2x lower latency than the 8t “with a near-zero performance penalty” for the security features. That’s Google’s number. I couldn’t find an independent test of it, and the 8tx isn’t shipping yet.

What the 8i Max adds

Memory, mostly. Google says the Max has double the compute, double the on-chip SRAM and double the high-bandwidth memory bandwidth of the 8i, in the same rack space. Network World gave the figures: 384 GB of HBM (33% more than the 8i), 17.2 TB per second of bandwidth and 768 MB of SRAM. Those numbers match Google’s April specs for the 8i (288 GB, 384 MB, about 8.6 TB per second), once you double the last two. Google’s launch post doesn’t list absolute figures itself.

System	Introduced	Aimed at
TPU 8t	April 2026	Large-scale training
TPU 8i	April 2026	Low-latency reasoning
TPU 8tx	November 9, 2026	Confidential inference and small-scale training
TPU 8i Max	November 9, 2026	Very long context windows, dense racks

What Google hasn’t said

Bloomberg, as relayed by Quartz, reported that both systems arrive next year, according to Google chip chief Amin Vahdat. The company’s own post, in the text I read, gives no date. It also gives no pricing for either system.

The hosts differ by page, too. Google’s blog.google post says the 8tx pairs with 5th Generation AMD EPYC CPUs, while the Google Cloud version of the same post says “custom AMD CPU hosts.” My read: probably the same hardware described two ways, but the two posts don’t say so.

Why Nvidia is the subtext

Bloomberg framed the launch as another step in Google’s challenge to Nvidia, and noted that Anthropic holds an option on up to $150 billion of TPUs between 2028 and 2031, per a regulatory filing. Oracle has already said it will offer TPUs to its cloud customers, which I covered in the earlier post on Google TPUs coming to Oracle Cloud.

My read: the quick release rhythm is the real story. Repackaging existing silicon for a specific buyer, whether that’s a bank worried about data in memory or a team stuffing a whole codebase into a prompt, is much faster than designing a chip. Vahdat told Bloomberg the thesis is that “the future is going to be about specialization.” Whether regulated customers trust Google’s attestation as much as they’d trust a rival’s is a question only audits will answer, and none are public yet.

Sources

Note on sources: Bloomberg’s original article wasn’t opened; its claims here come from Quartz’s and Investing.com’s summaries. The April 8i specs come from a Google deep-dive cited by third parties, which I did not open directly.