I certainly don't mean to throw any shade on the project -- if anything, the RAM utilization was less of a surprise for me than being bound to single-core performance by default. Sure, it's possible to split Synapse up into workers, but it definitely feels like there's some voodoo involved (my first attempt at setting just a Synchroton worker failed miserably this week)

Is it really necessary to maintain the state of the entire network on all federated nodes? (Since I'm guessing that's what you've described)

It isn't necessary, and we don't do it. What I've described is what happens if you type '/join #matrix:matrix.org' and your server starts participating in a large room which spans tens of thousands of users and hundreds of servers. Rooms are only replicated over the servers which participate in them. So if you don't join massive rooms, your server can be have fine. (And in the near future massive rooms should be usable too).