They can afford it because the market rightfully bets on such trained models being more useful than upstream sources.

In fact, at this point in time (it won't last), one of the most useful applications of LLMs is to have them deal with all the user-hostile crap that's bulk of the web today, so you don't have to suffer through it yourself. It's also the easiest way to get any kind of software interoperability at the moment (this will definitely not last long).