That's a misconception: you don't "use" relays (in the sense that you don't have to have a static list of relays you always use), you write to relays. When reading you connect to the relays of whatever the people you want to read from.

Some apps indeed use this method of selecting a static set of relays, and if that was the protocol you would be correct about centralization or bloat, but this is legacy from a naïve unfinished early implementation, most apps do the correct thing now and the rest is transitioning.