Google has expanded its eighth-generation Tensor Processing Unit lineup with two systems aimed at distinct enterprise needs: protecting sensitive AI workloads and increasing memory capacity for demanding applications. The additions, called TPU 8tx and TPU 8i Max, bring Google to four TPU systems shipped within a single year for the first time, the company says.
The expansion broadens Google's challenge to Nvidia as it develops its custom AI accelerators into a platform for external customers. Rather than introducing another processor generation, Google is using existing silicon in new system configurations, changing packaging, networking and memory arrangements to serve different workloads.
That approach separates the development of a new chip from the delivery of a new product built around it. Google says this modular strategy allows it to introduce hardware variants in months rather than the years typically required to develop a new processor.
The TPU 8tx focuses on confidential computing for enterprises and government agencies handling sensitive information. Encryption commonly protects data while it is stored or moving across a network, but processing creates another security challenge: information must remain accessible to the hardware performing the computation.
Google says the 8tx protects and isolates data and model weights during active use. The capability addresses a category of security requirements that Nvidia also serves through its GPU confidential-computing architecture. For customers evaluating either platform, protection during computation is a separate consideration from encryption at rest or in transit.
The 8tx also uses Google's low-latency Boardfly network, which was developed for real-time inference. Google claims a latency improvement of up to twofold compared with the TPU 8t. The company says customers can use one system for model fine-tuning, distillation and reinforcement learning for AI agents, avoiding the need to transfer sensitive data among separate clusters for those tasks.
The TPU 8i Max addresses memory constraints instead. Google has placed two TPU 8i chips in a single package, providing twice the compute, memory capacity and memory bandwidth while retaining the same rack format. The configuration is intended to increase the resources available within a given amount of data-center space.
Memory capacity and bandwidth can become limiting factors when AI applications handle large inputs or many simultaneous tasks. Agents working across substantial codebases and systems processing hours of video are examples of workloads that can put pressure on the movement and storage of data, rather than computation alone.
Google is also expanding the software support needed to make its hardware accessible. It has released TorchTPU as open source, with the aim of letting developers run PyTorch code on TPUs with minimal changes. The platform also supports inference frameworks including vLLM and SGLang.
Software compatibility matters because moving an AI workload between accelerator platforms involves more than selecting a chip. Framework support, deployment tools and the amount of code that must change all affect the effort required to adopt new hardware.
Google now counts nearly 100 external TPU customers. Its broader lineup reflects an effort to serve those customers with systems tailored to different operational requirements, as inference—the routine execution of trained models—takes a larger role alongside model training. The latest additions give that effort two specific targets: workloads requiring protection during processing and applications constrained by memory.
Sources: The Decoder