Hmm Im not convinced you'd be able to get the performance required to the DB even with the most exotic of caching/read and write tricks. You could maybe memdb it, but then you need a hefty amount of RAM, which again costs.
Has anyone actually tried to benchmark 10000 "nodes" reading and writing to a single DB on decent commodity hardware, let alone the high end stuff? I don't think it would be very pretty.
Plus, what about the pipe, thats a hefty pipe you'd need to serve up 10000 nodes.
In this scenario, you don't actually have 10000 x the DB load, you don't need 10000 DB writes to commit one transaction, you just need one. So, the load on the DB is equivalent to having 1 node. Reads can be cached and then served and farmed out to the 10000 ports.
In memory traffic is fast, a lot faster than DB traffic, so the load on the server would be high, but not untenable. It would just be like a busy website.