For example, how much processing power do you think it takes to have a computer fly a real helicopter upside down, and perform other acrobatic maneuvers that no human could possibly perform? It has been done, and its a lot less than you'd expect. It does take a day before hand to train it though in a simulator.

A proper design would have the centralized server (more like a server farm, no need to use expensive supercomputers like enahs suggests) taking inputs from all users and combining it to create a robust policy for specific observations that the client-based agents generate. Each additional user is actually a simulator to train the centralized AI. And a policy query to the central server farm would just be a database query.