The input lag on this would be horrible, even worse than your typical online game as you can't do any client-side prediction as its all server-side. So, if you had a 80ms latency to their server, it would take 80ms from when you pushed a key/moved the mouse before the server processes it and 80ms + rendering time (at 60fps, add on another 16.7ms) back again before you can visually see the results of your action. At least, in your typical fps game, you can start rendering the result of your action immediately and predict the movements of the other players and you will be right most of the time.
To get latencies anywhere near as fast as 80ms, they would need to have server farms all over the place. And these wouldn't be small server farms either, since, one person playing an intensive game would require a modern computer of their own.
Then factor in the bandwidth involved, and its prohibitive for anyone with bandwidth caps or who pays for the amount of bandwidth used. You also certainly wouldn't be able to have multiple people playing a game in your house either, as even if your internet connection could support 1 connection, would it be able to handle 2? 3? LAN parties would be impossible with this tech.
Then from the hands-on previews that I've read (for example:
http://www.shacknews.com/featuredarticle.x?id=1090 or
http://i.gizmodo.com/5184502/onlive-streaming-games-hands+on-impressions), there are noticeable compression artifacts, and even some complaints about input lag (the server was only 50 miles away).
But then again, I don't think we are the target market.