The privacy concern makes me realize something else : there's no incentive for users here. They lose privacy, what do they win?
It does not prevent totally the idea of using p2p to distribute content as a low level implementation, but since we (the ones who manage servers) are the only ones to benefit from it, we must first find a way for it to have no impact on users.
The biggest privacy problem is because of how p2p works currently : we have a list of IPs associated to a resource. How can we obfuscate this without going through proxies?
It works if we don't do naive P2P but rather Friend-2-Friend; this is what retroshare does, your downloads can go through your friends so that only they know what you download. But that requires a lot more steps than traditional bittorrent, so I'm not sure it could work in general.
Perfect Dark (a Japanese P2P system) is a direct implementation of that concept, where you automatically "maintain ratio" as with a private torrent tracker by your client just grabbing a bunch of (opaque, encrypted) stuff from the network and then serving it.
A more friendly example, I think—and probably closer to what the parent poster is picturing—is Freenet, which is literally an onion-routed P2P caching CDN/DHT. Peers select you as someone to query for the existence of a chunk; and rather than just referring them to where you think it is (as in Kademlia DHTs), you go get it yourself (by re-querying the DHT in Kademlia's "one step closer" fashion), cache it, and serve it. So a query for a chunk results in each onion-routing step—log2(network size) people—caching the chunk, and then taking responsibility for that chunk themselves, as if they had it all along.
Good ! Now I know you have JQuery 1.2.3, which is vulnerable to exploit XYZ, which I can now use to target you. This is one reason why apt-p2p and things like that can't be deployed in large; it's way too easy to know what version of what packages are installed on your machine.
This makes me think that an other feature could be to not disclose all peers available, but randomly select some. This would force someone wanting to look someone up to download a possible big amount of times the same list to check for an ip, instead of just pulling it once per ressource to know as a fact.
Both ideas are not exactly privacy shields, but steps to mitigate the problem.