"We are going to emulate all computers and physics even when players aren't logged in." I built a prototype trying to do something similar, inspired by a text-based RTS/MMO that kept running while you were asleep. For me this was an unrealistic goal, if the simulations are complex enough you can't operate at a price point users will pay for. Even if you can, because the game is now CPU bound, far less players fit on…
Really? Given what Notch plans (2,000 DCPUs per machine - which is certainly doable), let's assume he's running a high-cpu instance (large) at Amazon for those 2,000 DCPUs. That's $500 per month for the machine. If he can't get that much money out of 2,000 players, he's doing it wrong. Assuming you allow for eventual instead of atomic consistency, and allow for loss of CPU for several seconds if an instance goes kabl…
Assume for a moment a team of developers spend several months if not years doing infrastructure build-out to support the imagined CPU simulation system. Now they just have to implement the rest of the engine, the part powering the actual game, in which all players, CPUs, and other elements are active and interacting with one another at all times.
I'm hoping Notch succeeds, but lets be realistic -- the closest anyone has got to offline simulation is effectively "event queue and timer" and not for lack of trying.