Google launches TPU 8tx and 8i Max for the AI agent era

The motivation is the change in workload profile. AI agents do not just answer questions: they execute tasks in multiple steps, access sensitive data and process huge contexts, like entire codebases or hours of video. Generic hardware does not meet these two extremes.

In this article you will see:

• What TPU 8tx is and who it serves
• What TPU 8i Max is and how it solves the memory bottleneck
• How the four systems complement each other
• What changes for those building agents

TPU 8tx: confidential inference with hardware level security

TPU 8tx is the third system of the eighth generation. It is optimized for secure and confidential inference, and also serves small scale training. The target audience is regulated companies and governments that run frontier or proprietary models alongside sensitive data and autonomous agents.

The differentiator is security at the hardware level. Normally, data is encrypted at rest and in transit, but it needs to be unlocked in memory during processing. TPU 8tx keeps data and model weights protected and isolated even during active use, building security directly into the hardware layer.

The system uses the same silicon as TPU 8t, with fifth generation AMD EPYC CPU hosts, and adopts the Boardfly low latency network created for real time reasoning. The result is up to 2 times less latency compared to TPU 8t, with a performance penalty close to zero.

As the system combines training compute with a fast inference network, organizations can simultaneously run and customize models on a single system: fine tuning with private data, distillation of large models into smaller ones and agent training via reinforcement learning, without migrating sensitive data between clusters.

TPU 8i Max: double compute and memory in the same space

TPU 8i Max solves a different problem: the memory wall. When dozens of agents work together on an entire codebase or hours of video, memory bandwidth quickly becomes a bottleneck.

The solution was to combine two low latency TPU 8i chips in a single package. 8i Max delivers double the compute, double the SRAM and double the HBM bandwidth compared to 8i, keeping processors fed with essential data and preventing large workloads from stalling.

As 8i Max fits double the silicon in the same rack space, data centers can deploy significantly more compute power per square meter and per watt.

How the four systems complement each other

Together, the four eighth generation systems cover every stage of the agentic lifecycle. TPU 8t serves for training at scale, with superpods of up to 9,600 chips. TPU 8i is optimized for real time reasoning, with 80% better performance per dollar in low latency inference for MoE models. TPU 8tx covers confidential compute for regulated companies. TPU 8i Max handles massive context windows and maximum compute density per rack.

This is the first time Google has introduced four TPU systems in a single year. The company also reinforces its commitment to open software: support for PyTorch, open source tools and automation for developers, with the goal of making TPUs accessible to every developer and cloud customer.

How to test the TPU 8tx and 8i Max today

The announcement does not include direct access instructions. Google directs interested parties to the TPUs page on Google Cloud and to the live presentation streamed on the day of the announcement.

For teams already using TPUs in the cloud, it is worth tracking regional availability updates and the new consumption plans the company mentioned, including Agent Compute Engine as an API for cloud customers. Teams running agents with long contexts should evaluate whether 8i Max reduces the need to partition context across multiple chips.