I think the typical approach to the bottleneck issue with regard to traffic analysis attacks is that the machines on the edges can act as mixes. They can essentially launder traffic from within their respective meshes so that any intermediary between them can't do attribution. Of course, you have to trust the mix! So what then? Then communication to them has to be encrypted and onion routed, and moreover, continuously sent (even if what is encrypted is the message, "No data here dummy, this is just chaff") and then that has to be sent along, all so the mix doesn't know that you're actually communicating anything.

It's a whole category of research really. Papers like Herd at Sigcomm and Vuvuzela at SOSP are the two latest I've seen and following references there should be helpful. I think if you look at Herd there are a few tricks in there to lower the cost of all of the chaff with the superpeers (or whatever they call them, I read it a while ago). A hybrid system that mixes meshnet schemes for local peer to peer traffic with secret sharing schemes and mixnets for more disparate networks seems workable to me. The question is what benefits does the meshnet provide over the mixnet style schemes?