Amin Vahdat, head of AI infrastructure at Google, stated in a recent interview that electricity is the most fundamental bottleneck for AI infrastructure expansion, not chips or manufacturing capacity; software optimization's contribution to capacity improvement is "likely no less than hardware." He also revealed that Google's 8th generation TPUs, released this year, have been split into two independent chips for the first time: 8i (inference) and 8t (training), reflecting the rapid expansion of the inference market.