My vote begins with what Brian has suggested, but goes a bit further.
Consider a quad-cluster of SF25K's, with each node physically sited on a different continent.
You'd have superb throughput, scalability, fault tolerance,
and security against the off chance some "bad guys" invade your data center.
But then, if all that is a bit much, just go with a tricked-out SS20 running Solaris 2.6, and attached to a DSL line.
Well, the question is not really so dumb. If you have 5000 users, do you really know what any of them are doing all the time?
All you can do is try to somehow track gross violations of resource overuse, and steer those users to other, non interactive resources.
With a load leveled cluster, new users stay away from very busy nodes until they become less busy. It works very well. And creating an cron job that checks usage levels every so often and sending out a warning email to offending users works well too.