Google Launches TPU 8tx and TPU 8i Max, Completing Four-System Eighth-Generation AI Chip Lineup

Google has expanded its eighth-generation TPU portfolio with two new systems — TPU 8tx for secure, confidential inference and TPU 8i Max for massive context windows — marking the first time the company has shipped four TPU systems in a single year.

In a blog post published Monday, Google said the two additions are designed for a workload profile that barely existed when the first TPU launched a decade ago: fleets of autonomous agents reasoning over enormous context windows and handling sensitive enterprise data. "With computing moving from human to agent scale, businesses are asking for accelerators that can reason across AI models' massive context windows," the company wrote. "In addition, some highly regulated customers are asking for specialized encryption capabilities for unique workloads." For more context on this story, see our ongoing breaking AI news.

A System for Every Stage of the Agentic Lifecycle

The eighth generation began at Google Cloud Next in April with TPU 8t, built for large-scale training, and TPU 8i, built for real-time reasoning. Google cited up to 2.7 times better training performance for the 8t and 80% better inference price-performance for the 8i compared with the previous generation.

The two new arrivals split the remaining territory. As Google put it, the four systems together give customers "a purpose-built accelerator for every stage of the agentic lifecycle": TPU 8t for scale-up training, TPU 8i for low-latency reasoning, TPU 8tx for confidential computing, and TPU 8i Max for massive context windows and the highest compute density per rack.

TPU 8tx: Security Built Into the Hardware

TPU 8tx is optimized for secure and confidential inference and, Google notes, is also well suited to small-scale training. The target customer is explicit: regulated enterprises and governments running frontier or proprietary models and autonomous agents alongside their most sensitive data.

The engineering approach addresses a long-standing tension in secure computing. Data is normally encrypted when stored or transmitted, but has to be unlocked in memory while a computer processes it — the moment it is most exposed. Built on the same silicon as TPU 8t and paired with 5th Generation AMD EPYC CPU hosts, TPU 8tx builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use.

The system uses the same low-latency "Boardfly" network Google created for real-time reasoning, which means it can handle training-grade compute and fast inference simultaneously. Google says this lets organizations fine-tune models with private data, distill large models into smaller ones, and train agents through reinforcement learning on a single system — without migrating sensitive data between separate clusters.

TPU 8i Max: Attacking the Memory Wall

TPU 8i Max tackles a different constraint entirely. "When dozens of AI agents work together across an entire software codebase or hours of video at once, memory bandwidth quickly becomes the bottleneck," Google wrote. The company's answer is architectural: two low-latency TPU 8i chips combined onto a single package.

The result is a system that scales compute and memory in lockstep — delivering double the compute, double the SRAM, and double the HBM bandwidth of TPU 8i. And because TPU 8i Max fits twice the silicon within the existing rack footprint, Google says data centers can deploy significantly more computing power per square foot and per watt, a meaningful consideration when power and floorspace are the binding constraints on AI capacity expansion.

Why Systems, Not Chips

An important nuance in Monday's announcement is that Google is framing these as new systems rather than new chips. Designing a chip, the company explains, is a multi-year process of laying out billions of transistors bound by physics and manufacturing timelines. A system — how compute, networking, storage, and packaging are arranged around a chip — can be adjusted far faster.

That modularity is the core of Google's AI Hypercomputer strategy, and it is what allowed four specialized systems to ship within a single generation. Google also highlighted how much of its own AI infrastructure now goes back into building the hardware: the company uses DeepMind's AlphaChip reinforcement learning system, and says Gemini now helps deliver chip layouts "in hours rather than weeks." Google calls this "true co-design" and a unique differentiator.

The Competitive Context

The expansion lands amid intensifying competition in AI accelerators. TPU deployments have drawn attention for their role in serving models outside Google's own walls — the company notes TPUs now power training and serving for its Gemini models, agentic workloads, Google Cloud customer services, and most recently direct deployments of TPUs into customers' own data centers. Google describes TPUs as "a rapidly growing business."

Coverage of the announcement, including pieces from TechPowerUp and Electronics Weekly, framed the move in explicitly competitive terms — Google extending its custom-silicon lineup as it courts AI workloads that might otherwise run on NVIDIA hardware. NVIDIA, for its part, has continued shipping successive generations of its own rack-scale systems, leaving AI infrastructure buyers with an increasingly credible choice between merchant and custom silicon.

What It Means for Developers

For developers, the practical significance is less about any single benchmark and more about fit. Teams building agentic products that grind through long documents, large codebases, or extended video context have been hitting memory-bandwidth ceilings that raw FLOPS don't solve — the problem 8i Max is built for. Teams in finance, healthcare, and government that couldn't move sensitive workloads to shared infrastructure because data had to be decrypted during processing now have a hardware-backed option in 8tx.

Google coupled the hardware news with software accessibility messaging, pointing to its recent PyTorch announcements and its stated aim to make TPUs "as easy-to-use as general-purpose hardware." The company hosted a broadcast on Monday to detail the new systems and said customers can explore the full infrastructure lineup on the Google Cloud website.

Stay Ahead of AI

Custom silicon is becoming the strategic battleground of the AI buildout. Stay ahead of AI with AI Buzz Wire — breaking coverage of AI hardware, models, and the companies building them.

Read more AI news →