Google adds TPU 8tx and 8i Max, two new chips aimed at secure inference and memory-heavy agent work

The two systems target regulated buyers that need data protected during processing, and agent workloads that are limited by memory.

Google has added two eighth-generation TPU systems to its lineup: TPU 8tx, aimed at secure and confidential inference for regulated enterprises and governments, and TPU 8i Max, designed for very large context windows and high compute density per rack. Google says the 8tx pairs the same silicon as TPU 8t with 5th Generation AMD EPYC CPU hosts, building hardware-level isolation that keeps sensitive data and model weights protected during processing, and it also allows fine-tuning, distillation and reinforcement learning on one system. The 8tx reportedly delivers up to 2x lower latency than TPU 8t. The 8i Max combines two low-latency TPU 8i chips in one package to double compute, SRAM and HBM bandwidth relative to TPU 8i, targeting agent workloads that bottleneck on memory. Google notes TPUs now serve nearly 100 customers, and that this is the first year it has introduced four TPU systems.