"Google has split its TPU architecture into specialized chips for training (8t) and inference (8i). The 8t chip focuses on throughput for large models, while the 8i focuses on low-latency inference. This strategy aims to address the evolving demands of agentic AI."