Google Unveils TPU 8tx and TPU 8i Max — Eighth-Generation Chips Built for Agentic AI

Google has widened its eighth-generation TPU lineup with two new systems, and the split between them says a lot about where AI work is heading. The new arrivals are the Google TPU 8tx, built for AI training, and the TPU 8i Max, built for real-time agentic inference. They follow the earlier TPU 8t and 8i, and they land as Google pushes its hardware in two directions at once: bigger training runs on one side, faster live agents on the other.

The announcement came on November 9, 2026, as reported by SXF Brief on November 10. Google has not framed this as a replacement for the existing eighth-generation chips. It is an expansion — two more systems joining a lineup that already covers training and inference, with each new chip aimed at a specific kind of workload. So what does each one actually do, and why does Google need both?

Two new chips, two different jobs

The headline is simple: TPU 8tx is the training chip, and TPU 8i Max is the inference chip. That sounds like a small distinction, but in practice the two jobs pull hardware in opposite directions.

Training is the heavy lift. It means feeding a model enormous amounts of data, adjusting billions of parameters, and running that process for days or weeks. It rewards raw throughput and the ability to keep thousands of chips working together without wasting time. Inference is the opposite problem. It means running a finished model to answer a question, write a line of code, or take an action — and doing it fast enough that a user, or another system, never has to wait.

A chip that is great at one of those jobs is not automatically great at the other. That is why the lineup keeps splitting. Google is not building one chip to do everything. It is building one chip for each job and letting customers pick the right tool.

What the TPU 8tx is for

The Google TPU 8tx is aimed squarely at training. That is the job that has grown fastest as models have got bigger, and it is the job where the cost of the hardware matters most — because training runs are long, and every hour of compute adds up.

Think about what training actually involves. A model learns by working through data in huge batches, updating its internal settings as it goes, and repeating that cycle millions of times. The chips doing that work need to crunch numbers at enormous scale and pass results between each other constantly. If any part of that chain slows down, the whole run slows with it.

That is the problem the 8tx is built to solve. It sits in the eighth generation of Google’s TPU family, which means it inherits the design lessons of the chips that came before it — including the TPU 8t, which arrived earlier in this generation. Google’s own TPU pages on Google Cloud describe the family as the company’s custom-built accelerators for AI workloads, and the 8tx extends that line in the training direction.

What the TPU 8i Max is for

The TPU 8i Max has a different job: real-time agentic inference. That phrase packs in two ideas, so it is worth pulling them apart.

First, inference. Once a model is trained, inference is what happens when it is actually used — answering a prompt, summarising a document, or calling a tool. Second, agentic. An AI agent is a system that does more than answer one question: it plans a task, reasons through the steps, takes actions, and checks the result. That can mean a string of model calls in a row, each one depending on the last.

Now add “real-time” to that picture, and the hardware problem gets clear. An agent that pauses for seconds between steps feels broken. A voice assistant that waits, a coding agent that stalls mid-task, a support bot that takes too long to act — all of them need the model to respond fast, again and again, for as long as the task runs. The 8i Max is aimed at exactly that pattern: many quick responses, one after another, with no room for delay.

It also fits a wider shift in how AI gets used. The first wave of demand was about training bigger models. The next wave is about running them — constantly, in products people touch every day. Google’s bet, if the naming is any guide, is that inference for agents will need its own hardware, tuned for speed rather than sheer bulk.

Why Google keeps splitting its lineup

This is not the first time Google has divided its TPU family by workload, and it probably will not be the last. The eighth generation already had the TPU 8t and the 8i before these two arrived. The 8tx and 8i Max add more options on both sides of the same divide.

There is a practical reason for that. Cloud customers do not all want the same thing. One team is training a model from scratch and cares about getting the most work done per dollar. Another is running an agent in production and cares about response time above all else. A single chip design forces both teams to compromise. A split lineup lets each team pick the system that fits.

It also matches what Google has been doing across its AI business this month. On the software side, Google’s Gemini Omni 1.2 Flash brought fast, short-form AI video generation to its cloud platform. On the infrastructure side, Google bought Crusoe’s Wyoming data centre site and committed $17 billion to the state. The chips, the models, and the buildings all point in the same direction: more AI capacity, built on Google’s own terms.

And it is worth noting what this announcement is not. It is not a move into someone else’s territory. The chips in question are Google’s own design, built for Google’s own cloud — a different announcement from Microsoft’s Maia 300 chip, which the site covered in October. The two companies are solving the same broad problem, but each is doing it with its own silicon.

What to watch next

The announcement, as reported, keeps the focus on the two new systems and the jobs they are built for. That leaves plenty of open questions, and they are the ones worth tracking.

First, how will Google talk about these chips in the coming weeks? The Google blog and Google Cloud’s documentation are the places where detailed specifications, availability, and pricing usually appear. Until those details surface, the story is the shape of the lineup, not the numbers inside it.

Second, watch who actually uses them. Training chips and inference chips serve different customers, and the first real-world deployments will show whether the split matches what developers need. If agents keep growing, the 8i Max side of the lineup could end up doing the heavier lifting.

Third, watch the pattern. Google added two chips to its eighth generation in a single announcement, one for each side of the training-inference divide. If that pattern holds, the next announcements will likely follow the same shape — more options, each aimed at a narrower job, rather than one chip trying to do it all.

For now, the takeaway is straightforward. Google’s eighth-generation TPU family now has four named systems, and the two newest members make the strategy clear: train big on the 8tx, run fast agents on the 8i Max, and let each workload have the hardware it actually needs.