Google expands TPU family with chips for large context and secure AI

Expanding the TPU family

Google has officially introduced two new specialized silicon chips to its eighth-generation Tensor Processing Unit (TPU) architecture family. These new additions aim to provide higher compute density and more secure ways to run artificial intelligence. The company is expanding on the 8t and 8i chips it previously revealed at Google Cloud Next 2026.

The 8t and 8i chips were originally designed to handle the two primary tasks of AI: building large-scale models and running them in real-time. By adding the 8i Max and 8tx, Google is creating a more diverse toolkit for different types of AI workloads. This expansion helps customers choose specific hardware based on whether they need more memory, more security, or more efficiency.

As reported by SiliconANGLE, these chips are part of a broader push toward agentic AI. This type of AI is designed to perform complex tasks autonomously. To do this effectively, the underlying hardware must be able to handle specific demands like massive amounts of information or highly sensitive data environments.

Scaling massive context windows

The TPU 8i Max is specifically designed to handle massive context windows. A context window is the amount of data an AI model can consider at one time while answering a question. A larger window allows a model to work across long documents, audio files, video, or large collections of code without losing track of the information.

Expanding these windows is difficult because it puts a heavy demand on the chips to move data in and out. To solve this, the 8i Max doubles the compute, SRAM, and high-bandwidth memory of the standard TPU 8i. This provides the processors with enough on-chip memory to handle data without having to swap it out constantly. Reducing this movement lowers latency and decreases the power needed for the model to "think."

Google developed the 8i Max while pushing the limits of its Gemini frontier model. That model supports a 1 million-token industry standard, which is roughly equivalent to 1,500 pages of text. While flagship models use these large windows, the 8i Max also supports smaller models that typically use windows between 128,000 and 256,000 tokens.

Hardware-level security

For organizations that handle sensitive information, Google introduced the TPU 8tx. This chip is optimized for secure and confidential inference, which is the process of running a model to get an output. It is built for regulated enterprises and governments that need to run proprietary models alongside their most private data.

Normally, data is encrypted while it is stored or moving over a network. However, the data must be decoded for an AI model to process it. The TPU 8tx addresses this risk by building a secure enclave directly into the hardware layer. This keeps both the sensitive data and the model weights protected and isolated from other processes.

The 8tx is built on the same architecture as the TPU 8t and uses custom AMD central processing unit hosts. This combination pairs training-grade compute with fast inference. It allows organizations to run and customize models on a single confidential system, which reduces the risk of data leaks that can occur when moving information between separate clusters.

Impact on data centers

Efficiency in the data center is a primary goal for these new chips. The TPU 8i Max is designed to fit twice as much silicon into the same rack footprint as previous models. This reduces the amount of floor space required per square foot and per watt in a data center. It makes it more cost-effective to scale up for the industry's push toward ultra-large context windows.

For businesses, this means they can run larger models more efficiently. The 8i Max provides a purpose-built accelerator that handles the heavy lifting of memory-intensive tasks. By optimizing the hardware to match the software's needs, Google helps customers manage the high costs associated with running frontier models.

The 8tx offers a different kind of efficiency by streamlining workflows for regulated industries. Because it supports fine-tuning, distilling large models into smaller ones, and training agents through reinforcement learning on a single system, it removes the need for complex data migrations. This allows for faster development cycles in environments where security is the top priority.