Google has added two new systems to its eighth-generation TPU lineup, TPU 8tx for confidential inference and TPU 8i Max for massive context windows, taking the family to four systems launched in a single year. The company says 8tx is the first cloud TPU with hardware-enforced confidential computing, built for regulated businesses and governments running frontier models alongside sensitive data, while 8i Max doubles the compute and memory bandwidth of the reasoning-focused 8i. The launch lands alongside a wider software push, including the general availability of Cloud Pathways, as Google pitches its custom silicon for the agent era.
## TPU 8tx: Security Built Into the Hardware
TPU 8tx is aimed at a problem that has held regulated industries back from running their most sensitive AI work in the cloud. Data is normally encrypted when stored or sent over a network, but it has to be unlocked in memory while a computer processes it.
Google says 8tx closes that gap by building security into the hardware layer, keeping data and model weights protected and isolated even while they are in active use. The system is built on the same silicon as the training-focused 8t and paired with fifth-generation AMD EPYC processors as hosts, with AMD confirming the chips run the host infrastructure.
The security comes with little performance cost, according to Google. By adopting the low-latency Boardfly network built for real-time reasoning, 8tx delivers up to two times lower latency than 8t for multi-step agents, with what the company calls a near-zero performance penalty.
Teams can fine-tune models on private data, distill large models into smaller ones and train agents through reinforcement learning on the same system, without moving sensitive data between separate clusters — a setup Google says keeps private fine-tuning and secure agent execution on one box.
## TPU 8i Max: Doubling Up to Beat the Memory Wall
TPU 8i Max attacks a different bottleneck: memory. When dozens of AI agents work together across an entire software codebase or hours of video at once, memory bandwidth quickly becomes the limit. Google’s answer is to take two low-latency 8i chips and combine them onto a single package.
The result is double the compute, double the on-chip SRAM and double the high-bandwidth memory throughput of a single 8i, so large workloads do not stall waiting on data. Because 8i Max fits twice the silicon inside the existing rack footprint, data centres can deploy far more computing power per square foot and per watt.
Together, the four systems give customers a purpose-built accelerator for each stage of the agentic lifecycle, in Google’s framing: 8t for large-scale training, 8i for low-latency reasoning, 8tx for confidential computing, and 8i Max for massive context windows and the highest compute density per rack.
## The Software Push That Comes With the Chips
The hardware arrives with a wider effort to make TPUs as easy to use as general-purpose hardware. Last month Google open-sourced TorchTPU, so PyTorch code and Hugging Face models now run natively on TPUs, with ready-made backends for inference engines such as vLLM and SGLang.
A new open-source TPU agent suite, now in preview, plugs into coding tools including Claude Code, Cursor and Google Antigravity to automate onboarding, diagnose bottlenecks and tune kernels in plain language. Google has also open-sourced the [XLA Serving Engine](https://geroundtable.com/google-xla-serving-engine-open-source/), which lets teams train models in Python and serve them from a low-latency C++ runtime across TPUs, GPUs and CPUs.
The onboarding push matters because moving real code to TPUs has been hard even for frontier coding agents. In a recent [Princeton benchmark](https://geroundtable.com/coding-agents-fail-tpu-move/), the best agent completed only 15 of 20 real repository migrations to TPUs, a reminder that easier tooling is part of the product, not an afterthought.
Two more pieces went generally available today. Cloud Pathways brings the Google DeepMind runtime behind Gemini to every Google Cloud customer on Kubernetes Engine, and customers have already run more than 130 million Cloud TPU chip-hours through it in the past 90 days.
GKE dynamic slicing, which assembles TPU cubes into larger topologies or splits them into smaller isolated workloads on the fly, is also generally available. Google says the update cuts job startup times by up to five times and speeds fault recovery by up to 4.5 times.
Google is also open-sourcing TPU Sync, a model-state layer that turns chip memory, host memory, shared storage and the data-centre network into one memory hierarchy. The company says it cuts key-value transfer latency and time to first token by more than 90 percent.
A new Agent Compute Engine tackles the other waste problem: agents that pause for seconds or hours while waiting on a database, a tool or a person. It automatically suspends idle sessions and shifts working memory off the chip in milliseconds, cutting the cost of keeping agents ready.
## Why It Matters
The launch shows where Google thinks the AI race is heading: from human-scale chat to agent-scale computing, where dozens of sub-agents run in parallel and demand can jump ten or a hundred times. Google is answering with systems rather than a single new chip, packaging the same silicon differently for each bottleneck.
Confidential computing is the more consequential of the two bets. Regulated customers have had reason to keep sensitive workloads off shared cloud accelerators, and a hardware guarantee — rather than a policy promise — is what unlocks hospitals, banks and governments as TPU customers.
The AMD partnership matters here too: the host processor is now part of the security story, not just the plumbing. And the memory bet is aimed squarely at Google’s rivals — as context windows stretch toward millions of tokens, the vendor whose accelerators stall least on long-context work wins the agent workloads.
The competitive pressure is real. NVIDIA still dominates AI accelerators, and [Oracle has opened its cloud to rival chips led by Google TPUs](https://geroundtable.com/oracle-cloud-rival-chips/), a sign that customers want alternatives. Google’s answer is breadth: nearly 100 TPU customers already, and now a purpose-built system for every stage of the work.
The demand is visible inside Google’s own products, where [Gemini models now generate short AI video in seconds](https://geroundtable.com/gemini-omni-1-2-flash-ga/) on the same infrastructure. If agents become the main way people use AI, the company that owns the cheapest, most secure way to run them has a structural edge — and this launch is Google’s bid to be that company.
## Background
Google introduced the first two eighth-generation TPUs, 8t and 8i, at Cloud Next in April, pitching 8t for training and 8i for real-time reasoning with up to 2.7 times better training performance and 80 percent better inference price performance than the previous generation.
TPUs have grown from an internal project for voice recognition and translation into a business with nearly 100 customers, including frontier labs, trading firms, retailers and universities, and Google now lets customers deploy them in their own data centers as well as on Google Cloud.
Source: [Google’s announcement](https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/) and [AMD’s release](https://www.manilatimes.net/2026/11/09/tmt-newswire/globenewswire/amd-epyc-expands-its-leadership-at-the-foundation-of-ai/2465242)
0 Comments