Alphabet’s custom Tensor Processing Units, first announced publicly in 2016 after already running inside Google data centers, now give Google a structural AI infrastructure advantage as cloud rivals race to buy scarce accelerators for model training and inference. That advantage is not absolute...