> 8t for large-scale training and 8i for real-time reasoning

> TPU 8i Max for massive context windows and the highest compute density per rack, and TPU 8tx for secure and confidential inference

You had one job, Google