Google’s two new TPU chips, explained for people who build with AI

On 9 November Google announced two new AI chips: TPU 8tx and TPU 8i Max. Three days later, Nvidia’s CEO resigned. So it was a big week for AI hardware.

I don’t buy chips. I build SaaS products with Laravel and Next.js, and more and more of them call an AI model. But chips decide how fast and how cheap those AI calls are. So I read Google’s announcement carefully. Here’s what it says in simple words, and what I think it means for you.

First, what is a TPU?

A TPU (Tensor Processing Unit) is Google’s own chip for AI work. Think of it like this: a normal processor is a good all-rounder. A TPU is a specialist. It does one kind of maths, the maths inside AI models, very fast and with less electricity.

Google uses TPUs to run Gemini, and it rents them to other companies through Google Cloud. Google says nearly 100 customers use them now, including AI labs, trading firms, retailers and universities.

The four TPU chips of 2026

In April Google showed the first two chips of its eighth generation. The two new ones complete the family. Google says it’s the first time it has introduced four TPU systems in one year.

Chip	Made for	In one line
TPU 8t	Training big models	Up to 2.7 times better training performance than the last generation
TPU 8i	Answering fast	80% better price-performance for running models
TPU 8tx (new)	Private, confidential AI	8t silicon with security built into the hardware, on AMD EPYC servers
TPU 8i Max (new)	Very long context	Two 8i chips in one package: double the compute, SRAM and memory bandwidth

All the numbers in this table are Google’s own.

TPU 8tx: AI for data that must stay private

Some companies want AI but can’t send their data just anywhere. Banks, hospitals and government offices are the clear examples.

TPU 8tx is built for them. Google says the protection sits inside the hardware: customer data and the model’s weights stay isolated and protected even while the chip is using them. This is called confidential computing. Data is usually protected when it’s stored or sent. The hard part is protecting it while it’s being processed, and that’s what this chip is for.

Two more details:

• It uses the same chip as TPU 8t, but the servers around it run on 5th Generation AMD EPYC processors instead of Google’s own Arm-based ones. AMD says it’s the first cloud TPU system with hardware-enforced confidential computing.
• It can train and answer on the same system. So a team can fine-tune a model on private data without moving that data from one cluster to another.

I think this matters in Bangladesh too. Banks, fintech and health projects here ask the same question about AI: “Where does our data go?” Hardware like this gives a much better answer than “please trust us”.

TPU 8i Max: for AI that has to read a lot

AI agents read a lot before they answer. An agent fixing a bug may need to look through a whole codebase. Another may need to watch hours of video. All of that is called context, and context needs fast memory close to the chip.

TPU 8i Max puts two TPU 8i chips in one package. Google says that gives double the compute, double the SRAM (the fast memory on the chip) and double the HBM bandwidth of a TPU 8i. It also fits twice the silicon in the same rack, so a data centre gets more AI per square foot and per watt.

In plain words: longer conversations, bigger documents and bigger codebases, answered faster, and over time more cheaply.

The bigger picture: less dependence on one company

Today Nvidia sells most of the world’s AI chips. Big companies don’t like depending on one supplier, so Google, Microsoft, Amazon, OpenAI and Anthropic are all building their own chips and software (TechCrunch). Nvidia had a hard week too: its CEO, Jensen Huang, resigned on 12 November, as Nvidia’s filing confirms.

Google is also working on the software side, which is often the real problem when you change chips. In the same announcement it said PyTorch code and Hugging Face models now run on TPUs through TorchTPU, and that its new open-source TPU tools for AI agents work with Claude Code and Cursor. It also said one new tool cuts the time to the first word of an answer by more than 90% for some workloads.

More competition between chip makers is good news for everyone who pays for AI.

What this means if you build with AI

You will probably never rent a TPU yourself. You will feel these chips in other ways:

1. Prices and speed. When chips get better, the AI APIs you use get cheaper and faster over time. Features that cost too much last year start to make sense.
2. Private AI becomes realistic. If your client is a bank, a hospital or a government office, ask your cloud provider about confidential computing. The answer is getting better.
3. Don’t tie yourself to one provider. The best or cheapest model changes every few months. On this website, the AI writing assistant reaches its model through a few settings, so changing the provider is a configuration change, not a rewrite. I’d recommend the same for any product.

My take

New chips are not a reason to change your plans tomorrow. Nothing here ships today, and the numbers are Google’s own. But the direction is clear: AI is getting cheaper, faster and more private, and more companies are competing to make it. For people who build products, that’s good news. Build in a way that lets you switch when the better option arrives.

Planning an AI feature for your product and not sure where to start? Tell me about it. I reply personally.

References

1. Introducing TPU 8tx and TPU 8i Max, Google, 9 November 2026
2. Google broadens its AI hardware portfolio with two new specialized chips, TechCrunch, 10 November 2026
3. AMD EPYC expands its leadership at the foundation of AI, AMD, November 2026
4. Google adds two new specialized chips to its 8th-gen TPU lineup, Tom’s Hardware, November 2026
5. Nvidia CEO Jensen Huang resigns: live updates, CNBC, 12 November 2026
6. Nvidia Form 8-K: CEO transition, U.S. SEC, November 2026

Figures are from the companies’ own announcements unless noted. Specs, prices and dates may change.