Number of users matters not at all. It is all about throughput.
An embedded, like Geode, or Via processor maxes out at about 45 meg. Less if doing VPN.
An Atom dual core, or other Dual core will do wire speed fast ethernet. It will NOT do full gig. Depending on IO, you will get 300meg to 500 meg.
For full gigabit, you will need AGP connected cards, and still you will hit the wall at 600meg to 800 meg.
Scale-ability is also all about the number of states the router can manage, which is a finite number in m0n0wall. That factor is often easier to ballpark calculate by counting users and what their general use profile will be. 100 people checking code in and out of a software repository is much different from 100 people heavily torrenting and/or playing games, from a state perspective, even if their bandwidth may be very similar.
This is what often brought down the old Linksys WRT54G routers, where their states were very long lasting, but not a very large state table, so users would often fill up the state table via usage with wide access patterns. It was common for a single user to fill up the state table on a WRT54G just by polling a few server lists on some very popular games. It can be a concern in m0n0wall too since the state table size is fairly hard coded.
I'll assume your mention of "AGP connected cards" is a label swap mistake, I would imagine you're going for PCI-Express. I just want to make sure no one would get thrown off with the idea. (I won't go so far to say that there are
no AGP network cards, because I don't like to state assumptions as facts, but I would certainly bet that there aren't, and losing that bet would almost be worth it to see such a beast.)