I think the nuance here is what people consider the “cost” of “inference.” Purely on compute costs and not accounting for the cost of the model (which is where the article focuses) it’s not bad.