That's only true because most companies admit defeat before trying: they hit the disk when serving. The big Internet companies (Google, Facebook, LiveJournal, hell, even Hacker News and Plenty of Fish) all serve out of RAM: they keep everything a user is likely to hit in main memory so that a request never needs to perform I/O other than network. In this situation you're absolutely RAM-constrained.

I remember trying to optimize some financial software a couple jobs ago and hitting a brick wall because that's the speed the disk rotates at. We ended up buying an iRAM (battery-backed RAM disk) and sticking the DB on it. You can get this a lot cheaper by avoiding the DB and using a RAM-based architecture if you're willing to sacrifice fault-tolerance under power outages (or if you have some other architectural solution for fault-tolerance, like writing to multiple computers).