You're thinking too qualitatively; imagine the spikes they get... Also most apps are I/O restricted, not CPU...
Also, all these proposed messaging architectures are somewhat flawed. Twitter isn't really "sending" messages to other people. It's more like a person's message history is bound to their account, and appropriate privileges are applied. Then, when people try to "read," privileges are obeyed and information is produced...
Am I the only one who believes the solution to Twitter does not involve a massive distributed system? Twitter is inherently centralized... people don't see that.
Ok I can't edit the old post but they are basically just reinventing mailing lists. Anyway, poling is fine if it's initiated by a user action or it's a vary limited number of systems but if you have a few million users you can't uses it per user. As to reading old log's that's a trivially parallizable problem see database replication for some hints. So you separate the active messaging that automatically adds new Tweets to a user’s cell phone from the way users read old messages with the webpage and your home free.
PS: Text is cheep EX: Slashdot is running off of ~4 computers. And bandwidth is not a scaling issue until you need a single system to handle more than ~1GB of bandwidth per second. (Saturating an OC-3 line costs a lot of money but you don't need to change your architecture to do so.)
Also, all these proposed messaging architectures are somewhat flawed. Twitter isn't really "sending" messages to other people. It's more like a person's message history is bound to their account, and appropriate privileges are applied. Then, when people try to "read," privileges are obeyed and information is produced...
Am I the only one who believes the solution to Twitter does not involve a massive distributed system? Twitter is inherently centralized... people don't see that.