Google's 8th-gen TPUs split training and inference into two chips. Here's what it means for enterprise AI infrastructure strategy in 2026.