I'd like to now revisit the earlier comment about IP multicast being inappropriate as the ``architectural foundation'' for distribution in the internet. IP multicast, in its native (non-tunneled) mode, is fast and uses the wires efficiently. This is good, and is true because it runs in boxes that can (usually) run at wire speeds, and that join multiple wires together (i.e., routers). IP multicast also has no buffering, and has a ridiculously small addressing space. This is bad, and is true because long ago it seemed (and probably was) appropriate to graft multicast functionality onto IP, which is a bufferless thing that today has not enough addresses.
Even if someday we manage to go to IPv6, IP multicast will still be bufferless, and so still inappropriate for most kinds of distribution. A yoid architecture, by which I mean one where 1) the distribution topology auto-configures and can be limited only to boxes that are consumers of the content, 2) addressing is domain-name based as described above, and 3) the links are always unicast IP tunnels, even if the nodes on either end share a wire, can have the advantages of IP multicast (fast boxes attached to and joining together multiple wires) without the limitations.