Google adds TPU 8tx and TPU 8i Max, bringing four TPU systems in one year

Both new variants use existing silicon, with Google changing packaging, networking and memory setups around the same processor rather than introducing a new chip generation. The company says this modular approach lets it deliver new hardware in months rather than years.

The TPU 8tx targets enterprises and government agencies handling confidential data, keeping data and model weights isolated during active use. Google says a single system can be used to fine-tune models, distill large models and train AI agents with reinforcement learning without moving sensitive data between clusters. The TPU 8i Max places two TPU 8i chips on one package, which Google says doubles compute, memory and memory bandwidth in the same rack format.

Google is also releasing TorchTPU as open source so PyTorch code should run on TPUs with minimal changes, and says vLLM and SGLang are supported. The company reports nearly 100 external TPU customers as it pushes its chips toward inference workloads.

Agent developers running memory-heavy workloads, such as agents working across large codebases or long videos, may find the TPU 8i Max a new option on Google Cloud hardware.

Teams with confidential data may be able to run fine-tuning and reinforcement learning on one TPU system without splitting data across clusters.

If your code is written in PyTorch, Google's TorchTPU release suggests you can test inference on TPUs with fewer code changes, though you should verify compatibility for your own stack.