Time to show a little bit of knowledge is dangerous
I think the PS3 AAA games will be the bottleneck. PS1 and PS2 you can emulate on a PC, so no need for custom servers and it should scale in amazonian cloud fashion way.
PS3 I assume you can not emulate, so you need enough PS3 servers that handle the load/request for playing PS3 games on the devices that Game Now supports. The reason I am guessing that AAA is the bottleneck, is that if a game is made with Phyre Engine for instance. Which I think is a multiplatform SDK/Middleware/Something, it should be possible to recompile games based on that suit of software for the PC platform.
PS4 is "close" to PC as a hardware platform, so it probably is not a big issue to compile a version for PC ie Game Now, if it becomes very popular. Basically add Game Now Server as another output from Phyre Engine for smaller titles. AAA could get a requirement to take care of that to pass certification etc. And another assumption is that a Game Now server (virtualised most likely) should be easily more powerful than a PS4 if needed. And you can tailoer the virtualized hardware so that it only reserves what is needed to run for example Resogun...
Economic viability, well, I assume the capex now will be high, but it should have decent opex going forward, if you get to the point where the servers are virtual machines in a cloud. When I do not want to play Resogun anymore, just shutdown my VM and fire up another one for the guy that wants to play Killzone Shadowfall. You need less hardware and because that flexibility gives higher utilisation of the hardware.