Cloud flexibility allows to run the service at lower costs.
But if the bandwidth is enough, the limits of what the cloud can do depend on the user's connection lag, each task has a maximum acceptable lag, and when it comes to physics, close objects, potentially interacting with the player, require a lower lag than far ones, for which the physics engine can run in the servers at high sampling rate and communicate with the client at a lower rate, as long as it is high enough to ensure fluid motion 8BTW this already happens locally in racers, where the physics engine runs at several hundreds Hz sampling rate, while the video refresh is in the tens Hz order of magnitude. Anyway when the user's connection lag is barely below the maximum acceptable for a task, the cloud can help keeping real world lag closer to the connection's best case performance. Remote AI could accept an even higher lag, but still low enough to ensure a rate in the tens Hz.
This means that for most real time tasks, a ping below 100ms is required, for things that must be synchronised with the refresh rate it must be below 33ms for 30Hz refresh rate and below 16ms for 60Hz.
Before posting this I did several speed tests, the worst, under heavy load, gave me 495ms ping, while the best, under light connection load (just the test page) gave me 36ms ping, so I'd fall short by a few ms to use the cloud for far away objects physics in games with 30Hz refresh rate, but it could be able to run AI, at least of the NPCs not directly and fast interacting with me in the game.
NB: whatever the PR spin, forget running the graphics on the cloud (except precooking some non-dynamic tasks), it would require very low ping and monster bandwidth.
Stwike him, Centuwion. Stwike him vewy wuffly! (Pontius Pilate, "Life of Brian")
A fart without stink is like a sky without stars.
TGS, Third Grade Shooter: brand new genre invented by Kevin Butler exclusively for Natal WiiToo Kinect. PEW! PEW-PEW-PEW!

