Our newest TPUs: 8tx and 8i Max, built for the agentic era

Google has added TPU 8tx and TPU 8i Max to its eighth-generation lineup, alongside TPU 8t for large-scale model training and TPU 8i for real-time reasoning. The expansion responds to businesses moving AI agents into production, where two needs are becoming especially pressing: protecting sensitive information during computation and supplying enough memory for agents to reason over very large contexts. Google says it has introduced four TPU systems in one year, using modular system design to adapt compute, networking, memory and security faster than developing entirely new chips would allow. It also points to support for PyTorch and open-source software as part of making its accelerators easier for developers and cloud customers to use.

TPU 8tx targets regulated organizations that need to run or customize models and agents alongside private data. Using TPU 8t silicon, hardware-level confidential computing and AMD EPYC host CPUs, it protects data and model weights while they are in use; its low-latency network is said to provide up to twice lower latency than TPU 8t with near-zero performance penalty. It can also support fine-tuning, distillation and reinforcement learning without moving sensitive data between clusters. TPU 8i Max addresses memory bottlenecks in workloads such as coordinating agents across codebases or analyzing long videos. Combining two TPU 8i chips in one package doubles compute, SRAM and HBM bandwidth, while fitting more computing power into the same rack footprint.