I don't think that's actually the case. A lot of people say that, but inference is absurdly cheap, even compared to "api-subsidized" subscription prices. Where the math doesn't work out is the fact that they're training new models all the time. Any individual generation of model, if they were to just stop on it, (I suspect) would be profitable. But when you factor in continuous R&D and training costs of chasing the next gen, they're just hemorrhaging money.

But that's the game they're playing because they don't actually care about selling a product to customers; they care about investors. Nobody gets rich anymore by making a product and selling it to people. They get rich by making a product and then pitching it to investors and then getting bought out by a bigger fish or going public and having a big IPO.

The way the economy works now is that the companies are the products. Or at least their shares are, for a time...