Google announced two new AI chip systems on Monday, November 9, and the count is the headline: with the TPU 8tx and the TPU 8i Max, Google has now introduced four TPU systems in a single year, which the company says is a first. The TPU 8t and TPU 8i arrived at Google Cloud Next in April. These two complete the set, and they solve two very different problems: keeping sensitive AI work private, and feeding models with enormous context windows.
One thing to get straight before the details: these are new systems, not new chips. Both are built from silicon Google already announced, rearranged into packages aimed at specific jobs. That matters, because it means Google is competing on packaging and system design now, not just on raw transistor counts, and it can ship these faster than a brand-new chip would allow.
TPU 8tx: AI that never lets your data out of the lockbox
TPU 8tx is built for banks, governments, and other regulated organizations that want to run AI models and agents on their most sensitive data. It uses the same silicon as the TPU 8t training chip, paired with 5th Generation AMD EPYC host processors, and Google says it keeps data and model weights protected and isolated even while they are in active use. The idea is confidential computing applied to AI inference: the data stays encrypted and locked away from anything else running on the machine, including the cloud operator’s own software, while the model is working on it.
Google claims up to two times lower latency than TPU 8t for this setup, which matters because confidential computing usually costs speed. The system is also suited to small-scale training, according to Google, so the same locked-down environment can be used to fine-tune a model on private data, not just run one. AMD confirmed its side of the arrangement the same day, announcing that its 5th Gen EPYC processors power the host infrastructure for TPU 8tx, with the chipmaker’s SEV-based confidential computing features doing the isolation work.
Who is this for? Not you, probably, and not most companies either. It is for organizations that have compliance teams and auditors, and where the answer to “can we put this data in the cloud” has until now been no. Hospitals, central banks, defense contractors, and intelligence agencies are the natural customers, and the test of the system will be whether their security reviewers accept Google’s design once it is published in full. So far, the attestation design has not been published, and no customer has been named.
TPU 8i Max: two chips in one package, for the memory wall
TPU 8i Max goes after a different bottleneck, the one Google calls the memory wall. As AI agents work with longer conversations, bigger documents, and more tools, the limiting factor stops being raw compute and becomes memory: how much context the chip can hold close by, and how fast it can move it. Google’s answer is to put two TPU 8i chips on a single package, doubling the compute, the on-chip SRAM, and the high-bandwidth memory bandwidth of a single 8i, while fitting twice the silicon into the same rack footprint.
The underlying 8i design pairs 288 GB of high-bandwidth memory with 384 MB of on-chip SRAM, so the doubled package is aimed squarely at workloads with giant context windows and multiple cooperating agents, the kind of setup where several models pass work back and forth and each one needs the full conversation in memory. It is the infrastructure version of the agentic AI trend: the models are getting longer memories, and the hardware has to keep up.
Four TPUs, four jobs
| System | Announced | Built for | Status |
|---|---|---|---|
| TPU 8t | April 2026 | Large-scale training | Generally available later in 2026 |
| TPU 8i | April 2026 | Real-time inference | Generally available later in 2026 |
| TPU 8tx | November 9, 2026 | Confidential inference, small-scale training | Arriving in 2027, no price yet |
| TPU 8i Max | November 9, 2026 | Massive context windows, multi-agent inference | Arriving in 2027, no price yet |
What Google did not say
The announcement gives no price, no general-availability date, and no named customer for either system. Amin Vahdat, Google’s chief technologist for AI and infrastructure, told Bloomberg the systems will be available next year, meaning 2027, but Google’s own post gives no date at all. Every performance figure in the announcement is Google’s own, with no independent benchmark published for either system. That is normal for a launch-day announcement, and it is also a reason to treat the numbers as a promise rather than a result.
Google did say its TPUs now serve nearly 100 customers, including frontier labs, trading firms, retailers, and universities, but it named none of them. The customer count is up from a year ago, and the mix matters: if frontier labs are buying TPUs in volume, that is the clearest sign yet that Nvidia’s grip on AI training is loosening, at least at the margins.
Why this matters in the fight with Nvidia
Google is the only cloud provider with a serious in-house alternative to Nvidia’s GPUs, and it is now shipping that alternative in four flavors instead of one. The timing is not accidental. OpenAI has reportedly explored buying Google’s TPUs, which would have been unthinkable a year ago, and Google just spent $17 billion on a data center campus in Wyoming after buying an unfinished Crusoe project there. The chips, the data centers, and the customer list are all moving in the same direction: Google wants to be the second source for AI compute, and it wants the sale to happen on its own terms.
For readers, the practical effect is indirect but real. TPUs already power much of Google’s own AI, from Gemini to the search features that answer questions directly, and cheaper, faster inference hardware is part of why those services keep getting cheaper to run. Google’s AI answered its first question from space last month on its own infrastructure, and the company is clearly planning to put AI compute in places Nvidia’s roadmap does not reach. The 8tx and 8i Max will not change your phone or your laptop. They will change what AI services cost to run, and who is allowed to run them on sensitive data.
What to watch next
Three things will tell you whether this announcement matters. First, the availability date and pricing, expected when Google opens the systems to customers in 2027. Second, the first named customer for TPU 8tx, because a bank or a government agency putting its name on a confidential-computing system is worth more than any benchmark. Third, independent performance numbers, which will show whether the doubled memory bandwidth of the 8i Max translates into real speed for long-context work. Until then, this is a capability statement, not a product you can buy, and the right way to read it is as Google’s opening bid for the next round of the AI hardware race. OpenAI just raised $11 billion to buy more compute. Google just announced the compute it would like them to buy.
