how do we know they aren't making 9800$ profit on inference from 10k$ tokens at api pricing?