So, if each user is a simulator training your A.I., how does it remove the processing burden from the end user? You have just added more overhead on the CPU with regards to communicating the information and bookkeeping with the master A.I. server.
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.
You are now contradicting yourself from earlier, this is no-longer real time with regards to gameplay.
You are not talking about running the A.I. on a central server. You are talking about sending gameplay information and learning, and the updating A.I. and then updating the client.
That is different from the A.I. running remotely on a computer with more processing power. But you have not freed up any processing power from the end user.
Completely different thing then you talking about earlier.