Good point. Could usage patterns + inference costs give us proxy metrics? What would be a fair baseline?