Well that's like asking me what I think the next blockbuster movie would be, who will it star and what will it be about. If I could answer you then I'd be a billionaire
The quick answer is I don't really know. I could throw ideas out there I guess.
Maybe use cloud to make games more variable without having to patch the game every single time you play. So use those servers to generate either manually or procedurally new looks for the graphics, new paths for ai people, new missions, or whatever just so that the game doesn't always look the same every time I play it. All games could benefit from this because right now after you have installed a game it's stuck with the same looks, the same towns, the same weather conditions, the same people walking the same paths, etc which is bland unless they patch the game everyday which would be a bad user experience and not to mention tedious/expensive. Instead use the cloud machines to do it dynamically. That's a latency tolerant idea that all games could benefit from.
Or maybe there is a way to incorporate other peoples gameplay to enhance mine. So take data from other peoples games like perhaps missions then have done, dramatic landscape changes that they made, or crazy plays they made in sports games send all that data to the cloud machines and let them sort out which would be cool to incorporate with other peoples games, package it all up and send it all out as needed. That could be cool especially for sports games because you could use the things human players do in a sports game and turn them into ai into my game. Think about the biggest complain of sports games, it's the ai which becomes drab and predictable after a while. You can use the cloud servers to keep the ai fresh because you can farm human ai into cpu ai, so the ai would never become predictable as it's now always changing on a daily basis. That's also latency tolerant and would let other peoples gameplay help randomize mine automatically.
Or perhaps use the servers to pre-compute major world changes. If you script pre-computed stuff then it can look cool but will always look the same which is boring, or if you rely only on the host console then it can be dynamic but you can only do so much computationally speaking so it will look limited. Maybe you can combine both? So if there's 100 buildings in the level you are on then the cloud servers can pre-compute new ways for each of these buildings to be destroyed, damaged or decayed, maybe refreshing that data every few days. Let it do the heavy computing then package it all up into a script and send it out to peoples games as they play. On the users end it will still be playing back a script so it will be computationally light since the cloud servers did all the heavy lifting ahead of time, but it will look unique every time they play and would look more elaborate than the console alone could computationally handle. This would also be latency tolerant.
Anyways none of those are oh wow ideas though that will revolutionize things, I have no idea what those killer ideas will be anymore than I could have predicted Tetris, Super Mario Brothers, or GTA3. That's where we have to wait and see what happens and who will have that boom pow idea that makes cloud the next big thing.
Just to add to what TrungGap said, we don't know if they would be limiting themselves doublefold either. We don't have the data to know what percentage of console users don't have any online access. We also don't know how many users they would gain by going this route as well. Take my case, a console like the ps4 is of zero interest to me because it's just a console, I don't need one anymore than I need an mp3 player. I hadn't planned on getting an xb1 either for the same reason except that unique things like standard Kinect, vm's, and standard cloud supported games could make me a return customer because they are something new. So while they lose a customer like you, they gain a customer like me. So it's not cut and dry that they are losing 50% of their users, to me that number seems extremely high. We also don't know the value applied to each customer. So for example maybe online customers spend more money hence are a more desirable target compared to customers that are not online. If that's true then from a business perspective it can be worth losing 2 not online customers to get 1 online customer.
