Google launches two new AI chips with greater processing power and enhanced security for sensitive data
Two versions cater to different AI requirements
On Monday, October 9th, Google unveiled two new versions of its artificial intelligence chip systems. One is designed to run agents that process large volumes of data, such as hours of video footage or entire software codebases. The other targets businesses and government agencies that want to use AI with sensitive data, like financial records, without exposing it during processing.
The TPU 8i Max and TPU 8tx systems join the existing TPU 8t, used for training large models, and the TPU 8i, designed to speed up response times for already trained models. Both new versions reuse chips introduced in April, but Google has reconfigured how they integrate with memory, networking, and security features to suit different tasks.
Why AI agents need more memory from chips
An AI agent tasked with fixing a software bug might consult files, run tests, and then return to the code with the results. Each step adds information to the context window—the data the model can consider while working. If multiple agents do this simultaneously, processors may spend time waiting for the data they need to proceed.
To reduce this wait time, Google combined two TPU 8i chips into a single package. According to the company, the 8i Max has twice the computing power and internal SRAM memory of the 8i, and doubles the bandwidth of the HBM memory. This allows the system to transfer twice as much data per second between the HBM and the processor. The design also enables twice as many chips to fit into the same rack space, Google says.
How the 8tx protects data during processing
Encryption typically safeguards data while it’s stored or transmitted over a network. But during processing, the system needs to access this data in memory. That’s where the TPU 8tx comes in. It combines the TPU 8t chip with fifth-generation AMD EPYC processors to isolate both the task data and the model weights—the parameters that determine how the model works—directly in the hardware.
Google says the same system can run a model and fine-tune it with private data without transferring information between separate server groups. The company also claims the internal network can cut latency in half compared to the TPU 8t. For an agent that calls the model multiple times during a task, the waiting time for each step adds up.
Google aims to expand its chips beyond its own cloud
Tensor Processing Units (TPUs) were originally created for internal tasks like voice recognition and translation. Today, Google uses these chips to train and run Gemini, offers access to them in its cloud, and lets customers install them in their own data centers. The company says its TPU business now serves nearly 100 customers.
Growing this market also requires that companies’ software runs on these chips. In October, Google open-sourced TorchTPU, developed with Meta’s PyTorch team. The tool simplifies adaptation for those already using the library to develop models. The group also announced a deal to offer TPUs in Oracle’s cloud.
Anthropic, the company behind Claude, illustrates how major customers distribute this demand. They use Google’s TPUs, Amazon’s Trainium chips, and Nvidia’s GPUs. In April, they announced a partnership with Google and Broadcom to expand their computing capacity with next-generation TPUs starting in 2027. In Monday’s announcement, Google did not disclose pricing or availability dates for the 8tx and 8i Max.








