Complex queries should be optimized. The only time I have run into query issues is when: I was lazy, and didn’t want to optimize the query/schema/data layout and structures; the system was built by someone of questionable sanity and skillset; or the generated data is legitimately compute intensive, requiring lots of math i.e. more ops and cycles, but at that point it’s a report and the user can wait for it to gen (but in this case it is a lot of math that hasn’t taken advantage of parrallization).
In the wild, I’ve mostly seen people having zero clue what they’re doing (this includes management, who are as culpable for not caring about perf).
I’m not saying it should always or even most of the time be the case (quite the opposite - I’m with you and parent on that) but to say that there is no situation like that strikes me as pretty odd.
In ChatGPT's case, I'm assuming they're throttling responses to not hemorrhage money.