I suspect you're confusing the protocol (which can scale fine) with the Synapse implementation (which was never written for performance in mind). The protocol itself should be fine for twitter-size amount of messages however, and there are more performant implementations in the works.
