The current approach is as follows:
1) User sends next action to OnLive.
2) OnLive renders the frame and sends it to the user.
3) Repeat.
The "faster than light" approach is:
1) OnLive renders several frames that could be the outcome of all the possible actions by the user, and sends all of them to the user.
2) User chooses the next action.
3) The frame is already available and is rendered ASAP.
4) The action is sent to OnLive.
5) Repeat.
It is of course not actually faster than light, but it's faster than the time required to send light to OnLive and back. Although, if you're going to this trouble, you may as well just render the frames locally and not bother with OnLive at all... :-S
Well, going faster than the speed of light isn't possible, but if compressed really well, you could theoretically send massive quantities of data really quickly.
For example, if you want to play a game with someone who's on the other side of the world (or the galaxy) without that annoying lag imposed by relativity, you can simply have them send you a full scan of their brain, and then you can simulate it locally (not necessarily in your computer, it can run in the nearest Google MindSharing Center™).
Heck, you don't even need to play the game yourself. Send Google a copy of your brain too, and they will simulate the encounter in their servers and insert the memory in your brain when they're done. EVERYTHING can run in the cloud!
Ok, how is that possible? I genuinely am curious about the science. Feel free to provide links.