Google launches TPU 8tx and TPU 8i Max for secure AI agents

Hardware for secure and large AI

Google is expanding its custom silicon lineup with the launch of the TPU 8tx and TPU 8i Max. These systems are designed to solve two specific hurdles for companies: keeping data private and handling massive amounts of information at once. The company is positioning these chips to support the next wave of autonomous agents.

The TPU 8tx focuses on confidential computing, which protects information while it is being actively processed. Most traditional systems only encrypt data while it is sitting still or moving between locations. This new hardware allows for the protection of model weights and sensitive data even while the AI is running.

In contrast, the TPU 8i Max is built for high-density workloads that require large context windows. This means the AI can look at long video recordings or huge software repositories without slowing down. It achieves this by combining two TPU 8i chips into a single package to double the available memory and compute power.

Privacy for regulated industries

The TPU 8tx is specifically aimed at government organizations and regulated industries that handle highly sensitive information. It uses the same silicon as the TPU 8t but adds hardware-level security features. It also pairs with fifth-generation AMD EPYC processors to maintain high performance.

One of the main goals of this architecture is to reduce the need for companies to move data between different computing clusters. Organizations can fine-tune models, distill large models into smaller versions, and train agents in one secure environment. This keeps proprietary datasets inside a single protected perimeter.

Google claims the TPU 8tx can offer up to twice the latency reduction of the TPU 8t. It manages to provide these security benefits with almost no performance penalty. This balance is intended to make it easier for firms to deploy autonomous agents that require high-speed reasoning and strict privacy rules.

Solving the memory bottleneck

As AI agents become more capable, they often need to analyze vast amounts of data simultaneously. When an agent scans a long document, the memory bandwidth can become a bottleneck that slows down the entire process. The TPU 8i Max addresses this by doubling the SRAM and high-bandwidth memory of the standard 8i model.

This design also helps data centers use their physical space more efficiently. By doubling the amount of silicon within an existing rack footprint, Google allows for greater compute density. This is important for companies trying to scale their AI infrastructure while managing power and space constraints.

To support these large-scale tasks, Google also introduced TPU Sync. This software layer coordinates memory and storage across clusters to manage how data is shared. According to Google, this technology can reduce the time it takes for an AI to start generating its first token by more than 90% when using large context windows.

Open-source tools for developers

Google is making it easier for developers to use these new chips by releasing several open-source tools. One major addition is TorchTPU, which lets developers run PyTorch applications and Hugging Face models on TPUs without rewriting their code. This removes a significant technical barrier for many AI researchers.

The company also released a new TPU agent suite in preview. This suite integrates with popular coding environments like Cursor and Claude Code to help developers optimize their custom kernels. It allows users to identify performance bottlenecks and configure their environments using natural-language instructions.

Another tool, the Agent Compute Engine, focuses on cost efficiency for autonomous systems. It automatically suspends agent sessions when they are waiting for human feedback or database responses. By transferring working memory away from the accelerator during these pauses, the system reduces unnecessary compute usage for the customer.

Scaling and future infrastructure

To manage these large deployments, Google is expanding its Cloud Pathways through the Google Kubernetes Engine. This allows organizations to coordinate many TPU systems through a single controller for tasks like reinforcement learning or pretraining. It provides a unified way to manage complex, distributed AI workloads.

The company is also making dynamic slicing available for its Kubernetes engine. This feature lets organizations reconfigure TPU hardware resources on the fly to meet changing demands. Google says this can reduce the time it takes to start a new workload by up to five times and speed up fault recovery by 4.5 times.

Behind the scenes, Google is using AI to help design these very chips. Using tools like AlphaChip and Gemini-assisted design, the company can reduce certain chip layout tasks from weeks to just hours. This internal automation helps Google keep pace with the rapid demands of the AI industry.