The inference is absolutely not sold at a loss, at least not when paying API prices (the subscriptions are less clear). The reason frontier model companies aren’t profitable is because training the models is so costly, not inference.