Google launches two new systems: TPU 8tx and TPU 8i Max

Google has expanded its eighth-generation Tensor Processing Unit (TPU) portfolio with two new systems, TPU 8tx and TPU 8i Max, to meet the computing requirements of enterprises deploying autonomous AI agents.

The two systems target different workloads. TPU 8tx is built for confidential AI inference, allowing enterprises and governments to process sensitive data while models are running. TPU 8i Max, meanwhile, is designed to handle large context windows and workloads involving multiple AI agents.

“With computing moving from human to agent scale, businesses are asking for accelerators that can reason across AI models’ massive context windows,” Google said in a blog post.

The expanded portfolio also includes TPU 8t, built for large-scale model training, and TPU 8i, designed for low-latency reasoning.

TPU 8tx is optimised for secure and confidential inference, particularly for regulated enterprises and government organisations running proprietary AI models and autonomous agents.

The system uses the same silicon as TPU 8t and is paired with fifth-generation AMD EPYC CPU hosts. Google has built security into the hardware layer to protect data and model weights even while they are being processed.

Typically, data is encrypted when stored or transferred over a network but must be decrypted when processed in memory. TPU 8tx is designed to address this gap through hardware-level confidential computing.

Google said the system can deliver up to 2x lower latency compared with TPU 8t, with a near-zero performance penalty.

It also allows organisations to fine-tune models using private datasets, distil larger models into smaller ones, and train AI agents through reinforcement learning without moving sensitive data between separate computing clusters.

The second system, TPU 8i Max, addresses memory bandwidth constraints that emerge when multiple AI agents work across large software codebases or process hours of video.

Google has combined two TPU 8i chips into a single package. The resulting system offers twice the compute, twice the static random-access memory (SRAM) and twice the high-bandwidth memory (HBM) bandwidth of TPU 8i.

The configuration intends to prevent processors from waiting for data during large workloads.

TPU 8i Max also fits twice the silicon within the existing rack footprint, increasing computing capacity per square foot and per watt, according to Google.

The company said the system is built to support massive context windows, including workloads involving million-token contexts.

Alongside the hardware announcements, Google outlined software updates intended to make its TPUs accessible to more developers.

Last month, the company open-sourced TorchTPU, allowing PyTorch code and Hugging Face models to run natively on its TPU infrastructure without requiring code rewrites.

Google has also introduced an open-source TPU agent suite, currently in preview, that integrates with coding tools such as Claude Code, Cursor, and Google Antigravity. The suite can automate TPU onboarding, identify performance bottlenecks, and optimise custom kernels using natural language.

The company also announced the open-source XLA Serving Engine, which allows developers to train models in Python and serve them through a C++ runtime across TPUs, GPUs, and CPUs.

Another update, Agent Compute Engine, looks to reduce infrastructure costs associated with autonomous agents. Since agents can remain idle while waiting for database queries, external tools, or human feedback, the software automatically suspends and resumes idle sessions.

Google is also open-sourcing TPU Sync, a model-state layer that combines chip memory, host memory, shared storage, and data centre networking into a unified memory hierarchy.

The company said the system can reduce key-value cache transfer latency and time to first token by more than 90%.

Google said its TPU business now has nearly 100 customers, including AI research labs, trading firms, retailers, and universities. It is also working to make TPUs available for deployment directly in customers' data centres.