Google TPU 8tx and TPU 8i Max: Two New AI Chips Aimed at Nvidia

Google has added two new systems to its eighth-generation TPU family: the TPU 8tx, a confidential-computing system that pairs TPU 8t silicon with AMD EPYC processors, and the TPU 8i Max, which packs two TPU 8i chips into one package for double the compute, SRAM and HBM bandwidth. Both are expected in 2027, and neither is a new chip design. They are reworked versions of silicon Google first showed in April.

I’ve covered a lot of AI chip announcements this year, and most of them follow the same script: a big new architecture, a bigger benchmark chart, and a vague “later this year” ship date. This one is different, and in a way I find more interesting. Google didn’t design new silicon here. It took two chips it already had and rebuilt them around two specific problems its customers keep raising: keeping sensitive data private while a model runs, and the “memory wall” that slows down huge-context AI work. The timing is hard to miss, too. The news landed in the same week Nvidia’s board opened an independent investigation into CEO Jensen Huang, and Nvidia’s stock slid for a second straight session.

What Google actually announced

The announcement came in a post by Amin Vahdat, Google’s chief technologist and senior vice president for AI and infrastructure. It brings the eighth-generation lineup to four systems. Google says this is the first time it has introduced four TPU systems in a single year.

Here’s a quick recap of how the family fits together now:

Google didn’t publish pricing or exact ship dates for either new system. The 2027 timing comes from Bloomberg’s reporting, not from Google’s own post, so treat it as a target rather than a promise.

TPU 8tx: confidential computing for regulated industries

The TPU 8tx is aimed squarely at governments, banks, healthcare groups and anyone else whose lawyers get nervous when data leaves a locked room. It uses the same silicon as the TPU 8t training chip, but pairs it with fifth-generation AMD EPYC host processors and Google’s Boardfly network. The headline feature is hardware-level confidential computing. Customer data and proprietary model weights stay isolated while the model is actually processing them, not just when they’re sitting in storage.

That last point matters more than it sounds. Encrypting data “at rest” and “in transit” is old news. Protecting it “in use,” while it’s loaded into memory and being crunched, is the hard part, and it’s exactly where regulated customers have been stuck. Vahdat wrote that “some highly regulated customers are asking for specialized encryption capabilities for unique workloads,” and the 8tx is Google’s answer.

Google also claims the 8tx delivers up to 2x lower latency than the TPU 8t, with what it calls a “near-zero performance penalty” for the added security. If that holds up in real deployments, it removes the usual trade-off where you pay for privacy with speed. Google says the system supports private fine-tuning, distillation and reinforcement learning without shuffling sensitive data between clusters.

The AMD angle is worth a second look. Google picking EPYC hosts for a flagship TPU system is a small but real win for AMD in the data center, and another sign that the AI hardware world isn’t one company’s game anymore.

TPU 8i Max: Google goes after the memory wall

The TPU 8i Max is the more technically interesting of the two, at least to me. It takes two TPU 8i inference chips and puts them on a single package. Compared with a regular TPU 8i, Google says that gives you double the compute, double the on-chip SRAM and double the HBM bandwidth, and it fits twice the silicon into the same rack footprint.

Vahdat put it bluntly: the 8i Max “is built for a different problem: the memory wall.” In plain terms, the memory wall is what happens when a processor can do math faster than memory can feed it data. With today’s AI models holding enormous context windows (whole codebases, long documents, dozens of agents talking to each other), the bottleneck often isn’t raw compute. It’s how fast you can move the model’s working memory around. Doubling memory bandwidth per package attacks that directly.

Google’s example use case is dozens of AI agents working together on one software codebase. That’s a very 2026 example, but it’s a fair one. Agentic workloads keep a lot of state in memory, and they punish slow memory access hard.

This also connects to a story we’ve been tracking all year. Memory is the most squeezed part of the AI supply chain right now, and I wrote about how AI data centers have pushed RAM and SSD prices sharply higher. Any design that gets more useful work out of each gigabyte of HBM is worth paying attention to.

The software side Google wants you to notice

Chips get the headlines, but Google spent a big chunk of its post on software, and I think that’s where the real competitive fight with Nvidia sits. Nvidia’s moat has never been just its GPUs. It’s CUDA, and the fact that almost every AI developer already writes code that runs on it.

Google’s answer has several pieces:

• TorchTPU, open-sourced last month, lets PyTorch and Hugging Face models run natively on TPUs without rewriting code. It includes TPU backends for vLLM and SGLang, two popular serving engines.
• TPU Sync, also open source, moves AI working memory (the KV cache) between chips, servers and storage. Google claims it cuts KV transfer latency and time to first token by more than 90%.
• XLA Serving Engine, open source, lets teams train in Python and serve from a low-latency C++ runtime across TPUs, GPUs and CPUs.
• A TPU agent suite, now in preview, which works with coding tools like Claude Code, Cursor and Google Antigravity.
• Cloud Pathways and GKE dynamic slicing, both now generally available, with Google claiming up to 5x faster startup times and up to 4.5x faster fault recovery.

The message is clear: Google wants switching from Nvidia to feel like a configuration change, not a rewrite. Whether developers believe that is the real test.

Why this matters for Nvidia, and why the timing is awkward

Let’s be careful here. Google’s performance claims compare the new systems with its own earlier hardware, not with Nvidia’s latest chips. Nobody outside Google has benchmarked the 8tx or 8i Max yet. So this isn’t proof that Google has beaten Nvidia on raw performance.

What it does show is that the TPU business has become a serious commercial operation. Google says it now has nearly 100 TPU customers, including frontier AI labs, trading firms, retailers and universities. Bloomberg reports that Apple, Meta and Oracle are among them. Alphabet has said TPU system sales meaningfully contributed to Google Cloud’s latest quarter, when the division’s revenue more than doubled year over year to $30.7 billion. Bloomberg Intelligence estimates TPU sales, including related hardware and software, could pass $135 billion by 2028.

Then there’s the timing. This launch landed while Nvidia was dealing with the biggest leadership crisis in its history. Its board has opened an independent investigation into Jensen Huang, and Nvidia shares fell again on Monday. Google didn’t mention any of that, of course. But for a cloud buyer weighing a multi-year AI infrastructure deal, “the other big option has a stable story right now” is not nothing.

The broader trend has been building for a while. Custom AI silicon keeps getting more crowded. Startups like Positron are raising huge rounds for dedicated inference chips, Apple is building its own server silicon with Broadcom, and every hyperscaler wants less dependence on one supplier. Google is simply the furthest along.

My take

I like this announcement more than a lot of flashier ones, because it’s honest about what it is. Google didn’t pretend to have a new architecture. It listened to what paying customers kept asking for (privacy during compute, more memory bandwidth per rack) and shipped targeted configurations. That’s how mature hardware businesses behave.

The catch is the 2027 timeline. A year is a long time in AI hardware, and Nvidia’s next-generation platforms will be shipping by then. If you’re a business planning AI infrastructure today, the 8tx and 8i Max are worth watching, not waiting on. For everyone else, the takeaway is simpler: the AI chip race is a real race now, and that’s good news for prices down the line. You can read Google’s full write-up on the Google Cloud blog.

FAQ

What is the Google TPU 8tx?

The TPU 8tx is an eighth-generation Google TPU system built for secure, confidential AI inference and small-scale training. It uses TPU 8t silicon with fifth-generation AMD EPYC host CPUs and hardware-level confidential computing, which keeps data and model weights isolated while they’re being processed.

What is the TPU 8i Max?

The TPU 8i Max combines two TPU 8i inference chips on a single package. Google says it offers double the compute, SRAM and HBM bandwidth of a standard TPU 8i in the same rack space, aimed at memory-heavy work like huge context windows and multi-agent systems.

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

Bloomberg reports both systems are expected in 2027. Google hasn’t published pricing or exact availability dates.

Are Google’s new TPUs faster than Nvidia’s chips?

We don’t know yet. Google’s performance claims compare the new systems with its own previous TPUs, not with Nvidia hardware, and no independent benchmarks of the 8tx or 8i Max exist so far.