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