The calculation of hours is a little tough to imagine sometimes, is it the inference time itself, or the period of time used? Is there an average token cost per hour of use (average or explicit?)