Google is expanding its eighth-generation Tensor Processing Unit (TPU) lineup with two new specialized systems, TPU 8tx and TPU 8i Max, as the company prepares its AI infrastructure for increasingly demanding workloads involving autonomous agents, confidential computing, and massive context windows.

Announced on November 9, the two additions bring Google’s total number of eighth-generation TPU systems to four in a single year. The existing TPU 8t focuses on large-scale AI training, while TPU 8i is designed for low-latency reasoning. The new systems build on those foundations, addressing two increasingly important challenges for businesses deploying advanced AI models: protecting sensitive information and providing enough memory and computing capacity for complex AI workloads.

The announcement also highlights Google’s broader strategy to make its custom AI silicon more accessible to developers and enterprise customers, rather than limiting its use to Google’s own infrastructure.

TPU 8tx brings confidential computing to AI workloads

The first new addition, TPU 8tx, is designed primarily for secure and confidential AI inference, although it can also handle smaller-scale training workloads.

Its target audience includes governments and regulated enterprises that need to run proprietary AI models and autonomous agents alongside sensitive information. Traditional encryption protects data while it is stored or transferred across networks, but information generally needs to be decrypted while being actively processed. That creates a potential security exposure, particularly when organizations run AI workloads involving confidential business information.

Google says TPU 8tx addresses this problem by incorporating hardware-level security that protects sensitive data and model weights even during active processing.

The system uses the same underlying silicon as TPU 8t and pairs it with fifth-generation AMD EPYC CPU hosts. It also incorporates Google’s low-latency Boardfly networking technology, which was originally developed for real-time AI reasoning workloads.

According to Google, TPU 8tx can deliver up to two times lower latency compared with TPU 8t, with a near-zero performance penalty from its additional security capabilities.

Another important advantage is that organizations can perform several AI development tasks within the same system. They can fine-tune models using private datasets, distill larger models into smaller versions, and train autonomous agents through reinforcement learning without moving sensitive information between separate computing clusters.

That combination could make TPU 8tx particularly useful for organizations that want to bring advanced AI capabilities into highly regulated environments without compromising their data protection requirements.

TPU 8i Max targets massive AI context windows

The second addition, TPU 8i Max, tackles a different infrastructure challenge: the growing memory requirements of advanced AI models.

As AI agents become capable of handling longer tasks, working across large software codebases, and analyzing hours of video, the amount of information they need to keep accessible increases significantly. Processing power alone is not enough if the hardware cannot supply the required data quickly enough.

Google calls this challenge the memory wall, where memory capacity and bandwidth become bottlenecks that prevent processors from operating efficiently.

To address it, Google has combined two TPU 8i chips into a single package, creating a configuration designed to scale computing and memory resources together.

Compared with the standard TPU 8i, the TPU 8i Max delivers:

• 2x computing power: Double the compute capacity for demanding AI workloads.
• 2x SRAM: Twice the on-chip static random-access memory.
• 2x HBM bandwidth: Double the high-bandwidth memory throughput.
• Higher rack density: Twice the silicon within the existing rack footprint.

The increased memory bandwidth is particularly important for workloads involving massive context windows, where AI systems need to access large amounts of information without constantly waiting for data transfers.

The denser configuration also allows data centers to deploy more computing capacity per square foot and per watt, according to Google.

Four TPU systems for different stages of AI development

With TPU 8tx and TPU 8i Max joining the lineup, Google’s eighth-generation portfolio now covers four distinct workloads.

TPU 8t remains the primary option for large-scale training of foundation models, with an emphasis on maximizing useful AI computing output relative to energy consumption. TPU 8i focuses on low-latency reasoning and reduces inter-chip communication delays by up to half compared with the previous generation.

The two new systems extend those capabilities in different directions. TPU 8tx prioritizes confidential computing and secure inference, while TPU 8i Max is designed for massive context windows and higher computing density.

Google says the original TPU 8t and TPU 8i already offer up to 2.7 times better training performance and 80% better inference price-performance, respectively, compared with the previous generation.

The company has also revealed that its TPU business is growing, with nearly 100 customers across frontier AI laboratories, financial trading firms, retailers, universities, and other industries.

Google wants to make TPUs easier for developers to use

Beyond the hardware announcements, Google is also working to reduce the software barriers that have historically made specialized AI accelerators more difficult to adopt.

One significant development is TorchTPU, which Google open-sourced in October. It allows developers to run PyTorch code and Hugging Face models natively on TPUs without rewriting their existing code. Google also supports leading open-source inference engines, including vLLM and SGLang.

The company has introduced an open-source TPU agent suite, currently in preview, that integrates with development tools such as Claude Code, Cursor, and Google Antigravity. It is designed to automate TPU onboarding, identify performance bottlenecks, and optimize custom kernels through natural-language instructions.

Google is also introducing improvements across its software infrastructure. Its XLA Serving Engine provides a low-latency C++ runtime for deploying models trained in Python across TPUs, GPUs, and CPUs. Meanwhile, the Agent Compute Engine can automatically suspend and resume idle AI agent sessions, helping reduce infrastructure costs when agents are waiting for external tools, database queries, or human input.

Another important development is TPU Sync, an open-source model-state layer that combines on-chip memory, host memory, shared storage, and data center networking into a unified memory hierarchy. Google says its approach can reduce key-value cache transfer latency and time to first token by more than 90% in the relevant workloads.

These software developments are important because hardware performance is only one part of the equation. Developers also need familiar tools, accessible frameworks, and efficient deployment options to make custom silicon a practical alternative to more widely used AI accelerators.

Google’s latest TPU announcements reflect how quickly AI infrastructure requirements are changing. As businesses move beyond conventional chatbots toward autonomous agents that can work with larger datasets, longer context windows, and sensitive information, specialized computing systems are becoming increasingly important.

With TPU 8tx and TPU 8i Max, Google is extending its custom silicon portfolio to address those demands while continuing to open its infrastructure to a wider developer ecosystem.