“The amount of compute spent per inference query should be variable, potentially by a factor of 10,000 to 1,000,000, depending on the difficulty of the task.”