Last I heard, which I don't have any sources for so could be completely wrong for all I know, but some of the big LLM providers do make enough money off of their users to pay for inference and infrastructure. And the only reason they aren't profitable is because of the insane amount of money they spend on developing new models.