“The number of chips deployed for AI inference will be orders of magnitude greater than the number of chips used for training.”