The other thing I don't understand is why people focus on the network overhead rather than the overhead to run a single node.  As far as I can see, this should always be roughly O(n).