I kind of doubt they trained chatgpt on petabytes of application logs and web server logs. Is keeping all of this crap even useful for more than a small amount of time at this scale?

Actual good information will always be useful, most of this "big data" seems to be the equivalent of recording background static.