Well someone has to write the connection fabric, even in a content-addressed overlay. Writing a DHT that performs NAT traversal is just as difficult as for any other P2P application. See http://www.slideshare.net/vschiavoni/dht-and-nat for some examples.
CDNs have it easy because they control the network the servers are on. Writing p2p applications for the internet at large is much harder than it should be because of NAT.
What I'm suggesting is that if instead of a futile effort to convert the Internet to a new, Cisco-sponsored network protocol we invested in adoption of a few standard overlay networks, people could write P2P applications that targeted overlays instead of the Internet, and IP addresses would matter only as much as Ethernet addresses matter today.
The overhead of NAT is irrelevant on overlay networks; it scales linearly with the number of direct connections established (OSPF would call them "adjacencies"), but probably scales logarithmically with the number of relationships on the network.
Yes, assuming that the overlay algorithm actually establishes direct connections with a small number of peers. That's the way most overlays are designed because they need to reduce the cost of establishing a connection. By way of contrast, I recently worked on a gossip algorithm for producing uniformly distributed peer samples with no warmup time. This algorithm doesnt maintain an adjacency graph at all. Instead it sends a single udp datagram to a different peer at every tick (100 ms for the test system). Its a really simple algorithm that provides stronger guarantees than any existing peer sampling service but replace each datagram with a NAT traversal and the overhead kills it.
It works just fine for unicast, and even in strictly unicast situations, there are plenty of cases (conference calls, for instance) where "subscribe" turns out to be useful.
If this all sounds hand-wavy, well, it's already the mechanism used by the exchanges and ECNs to route stock information, and it's already a mechanism used to build CDNs. It's a considerably more straightforward evolutionary path for the Internet than forklifting out every IPv4-aware device or software package.