The model scales well, imho. Regionalize your network into
pieces. Apply each of the pieces into 1 or more proximities to a
NAP>MAE.
Been there, done that.... I've already stated my view on the
scalability of such arrangements.
Benefit: I gain low latency transit to most everyone.
Drawback: It is technically challenging to create an automate
system to regionalize and create appropriate filter lists.
It also complicates every peering relationship and multi-homed
customer connection, as you have to worry about both multiple
external AS's and your internal routing redistribution from all
of these regional routing clouds.
If you presume a fairly dense set of interconnects among transit
providers, then you're not going to get a significant improvement
in latency despite the added complexity.
/John