Subscribe

Artificial Intelligence

Google Adds TPU 8tx and TPU 8i Max, New Systems Built on Existing TPU Chips

The two eighth-generation systems target confidential inference and long-context work; Google's post gives no date or price, and 2027 availability comes from an executive interview.

DailyMech Staff

Published

Last updated / Fact-checked on

Google Adds TPU 8tx and TPU 8i Max, New Systems Built on Existing TPU Chips
Tho-Ge / Pixabay — Pixabay Content License

Network equipment rack with patch cables, switches and a DSL router

Illustrative generic networking equipment, not Google hardware or a TPU system. Photo: Tho-Ge / Pixabay, Pixabay Content License.

Google added two systems to its eighth-generation TPU lineup on November 9: TPU 8tx, for confidential inference, and TPU 8i Max, for very large context windows. Neither is a new chip. Google describes both as new configurations built around existing eighth-generation silicon. With TPU 8t and TPU 8i, introduced at Cloud Next in April, Google says this is the first time it has introduced four TPU systems in one year.

Google's announcement gives no availability date, price or regions. Amin Vahdat, Google's SVP and Chief Technologist for AI and Infrastructure, told Bloomberg the variations will be available next year, meaning 2027. Reuters, relaying Bloomberg's report on the new variants, said it could not independently verify it. The 2027 timing rests on that one interview.

At a glance

System Google's stated role Basis
TPU 8t Large-scale training Eighth-gen chip, April
TPU 8i Low-latency reasoning Eighth-gen chip, April
TPU 8tx Confidential inference; small-scale training TPU 8t silicon with AMD EPYC hosts
TPU 8i Max Massive context; highest density per rack Two TPU 8i chips in one package
Availability Not in Google's post 2027, per Vahdat to Bloomberg

What TPU 8tx adds for confidential computing

Google pitches TPU 8tx at regulated enterprises and governments running sensitive data alongside frontier or proprietary models. It uses the same silicon as TPU 8t, paired with fifth-generation AMD EPYC host processors.

Confidential computing means protecting data while it is being processed, not only when it is stored or sent over a network. Google says TPU 8tx builds that protection into hardware, keeping data and model weights isolated while in use. It also says teams can fine-tune, distill models and run reinforcement learning on the same system without moving sensitive data between clusters.

AMD's release says its EPYC chips handle host work for TPU 8tx and use AMD's Secure Encrypted Virtualization to extend protection to the connected accelerators. AMD calls TPU 8tx the first cloud TPU system with hardware-enforced confidential computing, and Google Cloud's Mark Lohmeyer, quoted in that release, makes the same "first" claim. That is a vendor characterization.

Google also says TPU 8tx uses its low-latency Boardfly network and delivers up to 2x lower latency than TPU 8t with a near-zero performance penalty. No independent test of those figures was identified in this research.

What TPU 8i Max doubles

TPU 8i Max combines two TPU 8i chips on a single package. Google says that gives double the compute, double the on-chip SRAM and double the HBM bandwidth of TPU 8i, with twice the silicon in the existing rack footprint.

Those multiples are hardware resources relative to TPU 8i, not measured application speedups. Google frames the target as memory-bound agent work, such as many agents working across a codebase or hours of video. Vahdat told Bloomberg that long context matters for drug discovery, math and coding.

TechTarget reported 768 MB of on-chip SRAM and 17.2 TB/s of HBM bandwidth for TPU 8i Max. Those figures do not appear in Google's post and are TechTarget's reporting.

What Google has not said

Google's post lists no prices, regions, general-availability dates or quotas for either system, and TechTarget also reported that Google provided no availability dates or pricing. The post does not say how either system will be offered, though AMD's release describes TPU 8tx as a Google Cloud system.

The post also highlights software, including TorchTPU and TPU Sync, and cites benchmark results. Those are Google claims and were not tested here.

What to watch

Watch for Google Cloud availability and pricing pages, independent latency and density benchmarks, technical documentation on how TPU 8tx verifies its confidential-computing protections, and named customer deployments of either system.

How we reported this

We read Google's November 9 post by Amin Vahdat and AMD's November 9 release in full. Bloomberg's report was read through licensed syndication on Moneycontrol, stamped 6:25 AM PST; Bloomberg's own page was not read. The Reuters item on The Economic Times relays the same Bloomberg report and is not separate confirmation. We also read TechTarget's article, which drew on a Google press briefing. DailyMech did no hands-on testing and contacted no companies.

Published and fact-checked November 9, 2026.

Sources

DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.

In Artificial Intelligence

From other sections

Sections

The Daily BriefingSample

Intelligence, distilled. Every morning.

Sample briefing. A five-minute read of the most important stories in AI and technology, once this list is real.

Sample form. It does not work. The address is not stored or sent.