>(although 2^18 is still probably too big). The limit might have to move up
>if we fill the routing tables with /18's...

Let's say we did have an absolute limit of /18s and 2^18 entries.
2^18 entries of 32 bytes each is 16Mb, which is almost within the
capacity of a Cisco 2500. (Well, Ok, CISCO would do something clever
about not storing the complete net and mask given that it would never
be more than /18 for external networks.)

This is presuming that we can nuke the older Class C's in the swamp, and that
we only have one view to worry about.

I envision an eventual environment where there are 20 or more well used
exchange point across the United States. I suspect this is about what
the US Internet will look like a year to a year and half from now.

20 ways out means 19 alternate paths to store in memory. Despite the fact
that Tony says that alternate's don't take up as much room as the prefix itself,
this is not a trivial amount of memory consumption that can be ignored on
the large scale.