You seem to have misread the article (which is not mine by the way), which makes the point that inference costs and revenue seem to scale with each other.