If they have more than one attachement, then they are (currently) using more than one AS - so with "one AS one policy", they should be fine. [If folks are violating the "one AS one policy" rule, then they are going to have problems that they have brought upon themselves. BGP does not really support more than one AS per policy.] If there are 2 ASs that touch in more than one point, then BGP4's inter-AS metric should take care of this (you will need to have some more care in what routes you pass between these 2 ASs and what routes they pass on to their neighbors), but as you need to coordinate to have 2 ASs touch in more than one point anyhow, you should just need to add a more to the coordination. In any case, downstream ASs should not know or care or need to do anything to make this work. You certainly should be able to do aggregation by site w/o any problem (even for ASs that touch in more than 1 place).
The second point I made was that a major role of the routing arbitrator will be to provide enough knowlege of topology to make more aggregation possible so we can approach the kind of figures Tony projected. Maybe that point got lost because of the way it was presented.
Actaully, the role of the RA is to get people to give it their (local) data so that the RA can get a database that has (nearly) complete global info. Projections of how much we are going to get from CIDR will be somewhat soft until we started getting a couple of real life examples. I have poked at what we would get under various CIDR strategeies, but as I have not done it for "real", I am not 100% sure yet. Here is some results from a run that I did of my data a while back: 1381 total nets 646 CIDR routes if CIDRize by site 619 CIDR routes if CIDRize by AUP (NSFNET vs non-NSFNET) 1026 NSFNET nets; 468 NSFNET CIDR routes 355 non-NSFNET nets; 151 non-NSFNET CIDR routes 500 CIDR routes if CIDRize everything together Even if we just CIDRized by site, it looks like we will get nearly all of the savings that we would get if we fully CIDRized. --asp@uunet.uu.net (Andrew Partan)