This week: Google unveiled TPU 8tx and TPU 8i Max, two new AI computing systems built for a world where AI agents handle sensitive data and read enormous piles of context at once.

Elsewhere: How Google builds a new system around an existing chip, the open-source software riding along, the fine print, and three fresh AI drops worth a skim.

The New Prime Signal

AI agents have two very human problems. They are asked to handle private stuff, like bank records or government files, and they are asked to remember a lot at once, like an entire codebase or hours of video. On Monday, Google answered both with two new members of its eighth-generation TPU family: TPU 8tx for confidential computing and TPU 8i Max for what Google calls the memory wall. TPUs are Google's custom AI accelerators. It first built them for its own services, and it now says TPUs are a rapidly growing business with nearly 100 customers, including frontier labs, trading firms, retailers, and universities. Think of tonight's launch as one chip that keeps a secret and another that never asks you to repeat yourself.

What shipped: Two systems, not brand-new chips. TPU 8tx is built on the same silicon as TPU 8t, Google's training chip from April, paired with 5th Generation AMD EPYC CPU hosts. Its trick is hardware-level security: Google says data and model weights stay protected and isolated even while in active use, which is normally the moment data has to be unlocked in memory. It also borrows Google's low-latency Boardfly network, and Google says it delivers up to 2x lower latency compared to TPU 8t with a near-zero performance penalty. TPU 8i Max takes two TPU 8i chips and puts them on a single package, delivering double the compute, double the SRAM, and double the HBM bandwidth of TPU 8i. Because it fits twice the silicon in the existing rack footprint, Google says data centers get more computing power per square foot and per watt. Together with TPU 8t and TPU 8i, announced at Google Cloud Next in April, that makes four TPU systems in one year, which Google says is a first.

Why it matters: Plenty of organizations want AI but get nervous about handing their most sensitive data to a model. Google pitches TPU 8tx at regulated enterprises and governments, and says teams can fine-tune models on private data, distill big models into smaller ones, and train agents with reinforcement learning on one system without moving that data between separate clusters. Meanwhile, agents are getting chattier with each other. Google says that when dozens of agents work across a whole codebase or hours of video, memory bandwidth quickly becomes the bottleneck, and TPU 8i Max is aimed squarely at that. The bigger picture is more choice: Decrypt frames the launch as Google taking aim at Nvidia, and more competition in AI hardware is generally good news for anyone paying the compute bill.

The fine print: Before anyone orders a rack, a few caveats. As Decrypt points out, Google compared the new systems with its own earlier hardware, not with Nvidia's latest products, and the company did not disclose pricing or availability dates for the two new configurations. Decrypt also notes that confidential computing can't stop a poorly designed AI agent from mishandling data it is allowed to access. Encryption protects the data from outsiders, not from a sloppy agent with the keys. And all the performance numbers here are Google's own claims, not independent benchmarks. The New Prime Signal take: This is not the flashiest AI news of the day, but it might be one of the more useful. Most people will never touch a TPU, yet the AI tools they use run on hardware like this. A system that keeps data protected while it's being used could nudge cautious hospitals, banks, and agencies off the fence. A system with double the memory bandwidth could make long, many-agent tasks less likely to stall. The real test is price and availability, and Google hasn't shown those cards yet.

Designing a chip takes years. Rearranging everything around it takes months, and that’s the whole idea.

How it works: Google draws a line between a chip and a system. A new chip means laying out billions of transistors in a multi-year process. A system is how compute, networking, storage, and packaging come together around the chip, and Google says that modular approach lets it deliver specialized supercomputers in a matter of months rather than years. The software is getting the same treatment. Google is open-sourcing TPU Sync, which turns chip memory, host memory, shared storage, and the network into one memory hierarchy; Google says it cuts KV transfer latency and time to first token by more than 90%. Its Agent Compute Engine suspends and resumes idle agent sessions while agents wait on tools or humans, Cloud Pathways is now generally available on Google Kubernetes Engine, and GKE dynamic slicing, Google says, cuts job startup times by up to 5x and speeds fault recovery by up to 4.5x.

Try it: Start with Google’s announcement, which explains all four eighth-generation systems on one page.

If you write PyTorch, note the part about TorchTPU, which Google open-sourced last month so PyTorch code and Hugging Face models run natively on TPUs, plus a new open-source TPU agent suite, now in preview, that plugs into coding tools like Claude Code, Cursor, and Google Antigravity.

For outside takes, see TechTarget.

DatacenterDynamics also covered the launch.

The Verge covered the launch too.