If you want < 10ms ping, you need to:
1) Be within 10ms of target datacenter (1,860 miles roughly)
2) rip out the TCP/IP handlers in all of the hardware along the route and replace with highly optimized, workload specific IP handlers (you'll want to do this in the kernel, OR in an ASIC if you're really serious about speed).
3) own all of the fiber along the route.
Basically, if you can eliminate all of the finger pointing and just own the network end to end, you could probably implement this today, it's just that no one is at that scale except for maybe Google (note: I have no idea how Google runs their internal network, but if I had to guess, this is probably what they're doing).
Edit: You can get this now, it's just hard and not worth the cost or effort to do (since no one will bear the cost of building a network like this for gaming, or really for any application I can think of... there just isn't a need for this kind of network performance).
I haven't tried OnLive, but the scenarios are very different so I don't think the comparison makes sense.