Google TPU 8tx and 8i Max: Private AI and Twice the Memory

Google announced two new TPU systems on November 9, 2026: TPU 8tx, built for confidential AI inference, and TPU 8i Max, built for massive context windows. Together with the TPU 8t and TPU 8i introduced in April, they complete an eighth-generation lineup of four systems launched inside a single calendar year, which Google says is a first for the company.

One system is aimed at the customers Google says it hears from most urgently: regulated enterprises and governments that run frontier models and autonomous agents alongside their most sensitive data. The other is aimed at a different bottleneck entirely, the memory wall that appears when dozens of agents reason across an entire codebase or hours of video at once. Both are real announcements with real engineering behind them. Neither has a price, a full specification sheet, or a date you can book. Bloomberg reported that both systems arrive in 2027, and Google's own announcement gives no availability date at all.

What Google actually announced

The eighth-generation TPU portfolio now has four members, each aimed at a different stage of the AI lifecycle:

Google introduced the first two systems at Google Cloud Next in April 2026, claiming up to 2.7 times better training performance and 80% better inference price performance than the previous TPU generation. Both figures are Google's own claims, not independent measurements, and that caveat applies to every performance number in this article. Google also says its TPU business now has nearly 100 customers, including frontier labs, trading firms, retailers and universities.

One distinction in the announcement matters more than the rest. Google is careful to call the 8tx and 8i Max new systems, not new chips. Designing a new chip takes years; a system is how the compute, networking and storage around a chip are packaged, and Google says that modular approach lets it ship specialized configurations in months. The 8tx reuses the same silicon as the TPU 8t. The 8i Max packages two existing TPU 8i chips together. That is the whole story of this launch: no new silicon, two new ways of putting existing silicon together.

TPU 8tx: confidential inference, with AMD inside

TPU 8tx is Google's third eighth-generation system, and it is aimed at regulated enterprises and governments running frontier or proprietary models and autonomous agents alongside their most sensitive data. The problem it addresses is simple to state and hard to solve: data is normally encrypted while it is stored or sent over a network, but it has to be decrypted in memory while a computer processes it. For a bank fine-tuning a model on customer records, or a government agency running an agent over classified material, that exposed moment is the risk.

The 8tx is built on the same silicon as the TPU 8t, paired with 5th Generation AMD EPYC CPU hosts. According to Google, it builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while they are in active use. AMD said in its own announcement that 5th Gen EPYC processors serve as the host CPUs and that the 8tx is the first cloud TPU system with hardware-enforced confidential computing, using AMD's Secure Encrypted Virtualization (SEV) to extend protection from the CPU to the attached accelerators, as reported by WebProNews.

The system also borrows the low-latency Boardfly network Google created for the TPU 8i, and Google says that gives it up to 2x lower latency than the TPU 8t with what it calls a near-zero performance penalty. That is a Google claim, and no independent test exists yet. Pairing training-grade compute with a fast inference network also means one system can run a model and customize it at the same time: teams can fine-tune on private data, distill large models into smaller ones, and train agents through reinforcement learning, all without moving sensitive data between separate clusters.

Two honest limits belong here. First, Google's announcement names no specific trusted-execution technology and publishes no attestation design, a gap noted in independent analysis by beri.net. Until Google publishes the mechanism, the confidential-computing claim is a design promise, not something a security team can audit. Second, the 8tx is not the only confidential option on Google Cloud today. Confidential GPU computing is already available: NVIDIA's RTX PRO 6000 on G4 instances went generally available on October 21, 2026 in 36 zones, and Google extended Confidential Bare Metal support to A4X Max instances powered by GB300 NVL72 systems in October. If you need confidential inference now, those are the documented routes; the 8tx is the route to watch for 2027.

TPU 8i Max: two chips, one package, double the memory

TPU 8i Max is built for a different problem: the memory wall. When dozens of AI agents work together across an entire software codebase or hours of video at once, memory bandwidth quickly becomes the bottleneck, and processors stall waiting for data. Google's answer is to take two low-latency TPU 8i chips and combine them onto a single package. The result, according to Google, delivers double the compute, double the SRAM and double the HBM bandwidth of the TPU 8i, while fitting twice the silicon within the existing rack footprint, so data centers can deploy more computing power per square foot and per watt.

Google did not publish a specification sheet for the 8i Max itself, but the arithmetic follows from the published TPU 8i specifications: each 8i chip carries 288 GB of HBM, 384 MB of on-chip SRAM, 10.1 petaflops of FP4 compute and 8.6 TB/s of memory bandwidth, as reported when the chip was announced in April. Doubling those figures gives the table below. Treat it as arithmetic from Google's "double" statement and the published 8i specs, not as figures Google printed for the 8i Max.

What that buys in practice is headroom for the key-value cache, the working memory a model keeps while it reasons over long context. The 8i's 384 MB of on-chip SRAM was already three times the previous generation, and Google says the chip can host a larger KV cache entirely on silicon. Doubling it again matters most for the workloads Google names: agentic coding across whole codebases, and video understanding over hours of footage. The 8i was also designed to cut inter-chip communication delays by up to half, another Google claim, and the Max inherits that design.

What it costs: no price yet, but the rate card shows the shape

Google has not published a price for the TPU 8tx or the TPU 8i Max. It has not published one for the TPU 8t or 8i either. The Google Cloud TPU pricing page, checked on November 9, 2026, lists only earlier generations, priced per chip-hour in US dollars:

Charges accrue while a TPU node is in a READY state, and billing is shown in the console in VM-hours, so a multi-chip VM costs a multiple of the per-chip rate. For readers outside the US, Ironwood's headline $12.00 per chip-hour works out to roughly £8.95 or CA$16.93 at November 9, 2026 market rates (1 USD = 0.7455 GBP, 1.4105 CAD), and the London $13.20 rate to roughly £9.84 or CA$18.62. Those are approximations from the day's rates, not Google's prices in those currencies; Google bills TPUs in US dollars.

The honest reading of this table is that $12.00 an hour is the current ceiling, not a prediction. The eighth-generation systems use newer silicon and will almost certainly sit above it when they are priced, but Google has said nothing, and no analyst estimate is a substitute for the rate card. There is also an open question about how the new systems will be sold at all. Until 2026, TPUs were only rentable inside Google Cloud. Alphabet began selling TPU systems to third parties for their own data centers this year, and Oracle has said it will resell TPU capacity from 2027. Whether the 8tx and 8i Max arrive as rentable chip-hours, as systems you buy, or both, is unconfirmed.

The software push matters as much as the silicon

The November 9 post is as much a software announcement as a hardware one, and the software is available now, on TPUs you can already rent. Last month Google open-sourced TorchTPU, so PyTorch code and Hugging Face models now run natively on TPUs, alongside out-of-the-box backends for inference engines like vLLM and SGLang. Our TorchTPU vs TorchNeuron comparison explains why that matters: it is the one-line escape from rewriting CUDA code for Google's chips.

Alongside it, Google announced a TPU agent suite in preview that plugs into coding tools like Claude Code, Cursor and Google Antigravity to automate TPU onboarding, diagnose bottlenecks and autotune kernels in plain language, and an XLA Serving Engine that lets teams train models in Python and serve them from a low-latency C++ runtime. On benchmarks, Google claims its native vLLM-TorchTPU stack achieved state-of-the-art price-performance on the industry-standard InferenceX and AgentX tests, which measure realistic agentic coding and long-context work. Those results come from SemiAnalysis evaluations cited by Google, and no independent reproduction has been published.

Three more pieces fill out the picture. Agent Compute Engine software automatically suspends idle agent sessions and shifts working memory off the chip in milliseconds, cutting infrastructure costs for agents that pause while waiting on tools or people. TPU Sync, now open-sourced, turns chip HBM, host DRAM, shared storage and the data-center network into a single memory hierarchy, and Google claims it cuts KV transfer latency and time to first token by more than 90%. Cloud Pathways, now generally available on Google Kubernetes Engine, brings the DeepMind runtime behind Gemini to every Google Cloud customer, and GKE dynamic slicing, also generally available, lets teams assemble TPU topologies on demand; Google claims it cuts job startup times by up to 5x and accelerates fault recovery by up to 4.5x. The pattern is the point: Google spent 2026 removing the reasons teams gave for not using TPUs, and this launch extends that work to regulated and long-context workloads.

What Google did not say

Four gaps are worth stating plainly, because they decide whether any of this is actionable yet:

• No price. No rate for the 8tx, the 8i Max, or even the April 8t and 8i appears on Google's pricing page.
• No date. Google's announcement gives no availability date or region list for either system. Bloomberg reported both arrive in 2027, as noted by beri.net.
• No full specifications. The 8i Max figures in this article are arithmetic from the 8i's published specs, and Google has published no standalone spec sheet for either new system.
• No independent tests. No third party has run either system. Every performance figure, from the 2x latency claim to the 5x startup improvement, is Google's own.

What to watch next: the pricing page, which is where a real rate will appear first; the first independent benchmarks once hardware reaches customers; and whether Google publishes an attestation design for the 8tx, which is what would turn its confidential-computing claim from a promise into something a security team can verify. For context on how Google Cloud is positioning itself more broadly, see our coverage of OpenAI's GPT-6 models landing in Google Cloud's Model Garden and the Google Data Agent Kit.

Who should care

If you run AI agents on sensitive data in a regulated industry, the 8tx is the system to watch, but do not plan 2027 capacity around it yet: no price, no date and no published security mechanism means there is nothing to put in a budget or an architecture document. If you run long-context agent workloads, whole-codebase coding agents or hours-long video understanding, the 8i Max is the more concrete of the two, because its doubling claim can be checked against the published 8i specs. For everyone else, the usable news from November 9 is the software: TorchTPU, TPU Sync, Cloud Pathways and GKE dynamic slicing all work on TPUs you can rent today. The chips are a 2027 story. The software is a today story.

Frequently asked questions

When will TPU 8tx and TPU 8i Max be available?

Google has not given an availability date or a region list for either system. Bloomberg reported that both arrive in 2027. Until Google confirms a date, treat 2027 as reported, not confirmed.

How much will TPU 8tx and TPU 8i Max cost?

Google has not published a price for either system, or for the TPU 8t and 8i announced in April. The current ceiling on Google's TPU rate card is Ironwood at $12.00 per chip-hour on demand in Iowa, or $13.20 in London, with lower rates on one- and three-year commitments.

Is TPU 8i Max a new chip?

No. It packages two existing TPU 8i chips onto a single package. Google says that doubles the compute, SRAM and HBM bandwidth of a single 8i, and fits twice the silicon in the existing rack footprint.

What makes TPU 8tx confidential?

It pairs TPU 8t silicon with 5th Generation AMD EPYC host CPUs, using AMD's Secure Encrypted Virtualization to keep data and model weights protected and isolated while in active use. Google has not published an attestation design or named a specific trusted-execution technology, so the mechanism cannot yet be independently audited.

Can I use TPUs without rewriting my PyTorch code?

Yes. Google open-sourced TorchTPU in October 2026, so PyTorch code and Hugging Face models run natively on TPUs, with out-of-the-box backends for vLLM and SGLang. A TPU agent suite in preview also plugs into Claude Code, Cursor and Google Antigravity to automate onboarding and tuning.

Google's November 9 announcement is a clear statement of direction: confidential computing and massive context are the two workloads the next wave of AI agents will need, and Google now has a system aimed at each. But direction is not delivery. No price, no date, no full specs and no independent tests mean the 8tx and 8i Max are worth tracking, not worth budgeting for yet. The software that shipped alongside them is the part you can use today.