So, all credit for this idea goes to Enarcy who mentioned the idea over lunch to me.

Basically, using the Desert map, I've made a mission that creates groups of 12 AI randomly. As I create each group, I measure the average server FPS, RAM and outbound bandwidth. You can see the sweet spot - I'm hoping it will assist me in keeping mission sizes right for our server.

Below is the graph for Firefly (2.2GHz AMD on Linux), the CTB server. You can see as the number of AI Units reaches the 100 mark, server FPS drops and then plateaus at 150 units, when the rapidity of FPS decline is reduced. At 200 units, the decline of FPS is maintained, but server outbound network traffic goes through the roof up too 330 AI Units.

Below is the graph of my laptop (2.6GHz on Windows XP) running a dedicated server. Up to 240 units, server FPS is stable, after which, network outbound traffic plateaus and server FPS drops dramatically to 330 units.

 Below is the mission file I used to perform the tests.