Google's New TPUs: Private AI and Huge Context for Business Agents

Google announced two new AI chips on 9 November 2026: TPU 8tx and TPU 8i Max. One is built for confidential inference, so businesses can run AI on sensitive data without exposing it. The other is built for huge context windows, so AI agents can hold far more of a job in memory at once. Together they mark a shift: not one chip for everything, but specialised hardware for different workloads. Here is what each chip does, what Google has not said, and what it means for your business.

They complete Google's eighth-generation TPU portfolio.

What Google actually announced

Announced on the company blog on 9 November 2026, TPU 8tx and TPU 8i Max are the newest members of Google's eighth-generation TPU family. Google's own announcement is the source for the details here.

TPU 8tx: confidential inference, explained

TPU 8tx is built for one problem: running AI on sensitive data without exposing it. It uses the same silicon as TPU 8t, paired with 5th-generation AMD EPYC host CPUs and Google's Boardfly network. Google claims up to 2x lower latency than TPU 8t with near-zero performance penalty — a vendor claim, not an independent benchmark.

Hardware-backed isolation

TPU 8tx uses hardware-backed isolation to keep sensitive data and model weights protected while in active use. The chip itself enforces the separation, rather than relying on software rules alone. Data stays protected during processing, the moment it is usually most exposed. For patient records, client finances or legal files, that is the gap that matters.

Private fine-tuning without moving data

The chip also supports private fine-tuning, distillation and reinforcement learning without moving sensitive data. Most leaks happen when data moves. Adapt the model where the data lives and the leak surface shrinks. Our RAG chatbot guide makes the same argument: keep company data close.

TPU 8i Max: built for the memory wall

TPU 8i Max tackles a different bottleneck: memory. Agents must hold enormous context in memory at once. When memory cannot keep up, the model forgets or cannot run the job at all. Google calls this the memory wall. TPU 8i Max packs two TPU 8i chips into one package, and Google says it doubles compute, on-chip SRAM and high-bandwidth-memory bandwidth versus TPU 8i — Google's numbers, not independently verified results.

Why huge context windows matter for agents

Context is what separates a chatbot from an agent. A chatbot answers one question at a time; an agent works through a job that can run for hours, holding the brief, history and documents in memory. When the window is too small, the agent loses the thread. Our GDPx coverage found no software agent clears roughly 12% on long-horizon tasks yet.

Why confidential inference matters now

The timing is not an accident. In October 2026, a Cloudflare AI data leak exposed cached information from earlier requests in other customers' responses. The lesson is plain: when AI systems share infrastructure, one customer's data can end up in another customer's answer — exactly what confidential inference is designed to prevent.

This is why the announcement matters beyond Google's data centres. Businesses holding back from AI share one objection: our customers' information could leak. A chip that protects data during processing does not remove the need for governance, contracts and access controls, but it removes a genuine technical objection. Our clinic chatbot guide and our look at Claude handling bank data land on the same point: solve the trust question first.

What this means for businesses running AI agents

Neither chip changes what your business can do tomorrow morning. But together they point to where AI infrastructure is going, with three consequences for companies buying agents.

Regulated businesses get a real option

Finance, healthcare and legal work have been slowest to adopt AI agents: the data is sensitive and a leak is severe. Confidential inference gives them a credible path: protected hardware, data kept in place, private fine-tuning. If you run a clinic, a law firm or a financial practice, ask your vendor: where does my data go while the model works on it? Our clinic chatbot guide covers what to ask first.

Agents that remember more of the job

Longer context means agents that hold a whole job in memory: the full customer history, the complete contract, every message in a support thread. Today the workaround is careful summarising and chunking, which is why our RAG guide spends so much time on retrieval design. Cheaper, larger context does not replace good retrieval, but it makes the agent's job easier.

Compute splits by workload

The deeper trend is that AI compute is splitting by workload: training, real-time inference, private inference and memory-heavy inference are becoming four different jobs, on four different chips. Good news for buyers — no training-grade prices for a simple inference job. "Which AI should we use?" is becoming "which AI, on which hardware, for which job?" See our chatbot versus agent guide and our real estate agent example.

What Google did not announce

An announcement is not a product you can buy. Three things are missing. First, no price — without it, you cannot compare TPU 8tx with existing options. Second, no general-availability date — Google says four eighth-generation systems shipped in one year, but not when either reaches ordinary customers. Third, no named customers — every chip sounds good in a press release. So treat both chips as a signal, not a purchase decision. Planning an AI project for 2027? Ask your vendor how confidential inference and longer context fit their roadmap. Our UK custom software cost guide applies the same discipline: price the real thing, not the announcement.

What to do now

Do the groundwork now, whatever hardware arrives. Start by mapping where your sensitive data goes today: which AI tools touch customer, patient or client data, and where is it processed? That audit is worth doing now — confidential inference only helps if you know what needs protecting. Next, find where your agents lose context: long support threads, multi-document reviews, jobs that stall halfway.

Then ask your vendor when confidential inference and longer context will reach the tools we use? If you are scoping an AI agent project, build it so the infrastructure can change underneath it; a well-built agent should not care which chip runs it. That is the point of AI agents built on your processes, not one vendor's hardware. If the project involves sensitive data, design the data handling now: our logistics automation and SaaS scaling guides show the same lesson — the architecture you choose early is the one you live with. For help scoping an agent that handles sensitive data safely, talk to Coheser about AI automation, see our custom software development work, or contact us.

Frequently asked questions

What is Google TPU 8tx?

TPU 8tx is a new eighth-generation Google TPU announced on 9 November 2026, built for secure and confidential inference. It uses the same silicon as TPU 8t, with 5th-generation AMD EPYC host CPUs and Google's Boardfly network, and hardware-backed isolation protecting data and model weights while in active use. No price or availability date has been published.

What is TPU 8i Max?

TPU 8i Max is a new eighth-generation Google TPU built for the memory wall: massive context windows and the highest compute density per rack. It packs two TPU 8i chips into one package, and Google says it doubles compute, on-chip SRAM and high-bandwidth-memory bandwidth versus TPU 8i. No price or availability date has been announced.

Can I use TPU 8tx or TPU 8i Max today?

No. Google has announced the systems but published no pricing, general-availability date or named customers. Treat this as a direction of travel, not something you can provision now. Planning a 2027 AI project? Ask your vendor how confidential inference and longer context fit their roadmap.

Why does confidential AI inference matter for my business?

Because AI systems share infrastructure, and the October 2026 Cloudflare leak showed one customer's cached data appearing in another customer's responses. Confidential inference protects data while the model works on it — most important for regulated businesses in finance, healthcare and legal work. It does not replace governance, contracts and access controls, but it removes a real technical barrier.