There's the obvious DRM win. Accessibility will be higher with way lower or non-existent hardware costs. These probably look very enticing on a cooperate slide deck. Not to mention the subscription model.
On a more technical side game streaming is only really desirable for games that don't run well on commodity hardware (see AAA 3D action titles). Latency is very important for these titles and the general rule of thumb is that 50ms of input latency is readily noticeable to humans. Average home connections have ~20-30ms round trip latencies. These speeds are not a reality everywhere or consistent. Combined with hardware input, world simulation, and rendering latencies it can be difficult to get latencies low enough to be comfortable.
Modern games just have the client run its own world simulation and rectify with the server later to hide the network latency. This strategy won't be possible with streaming. Is there some alternate optimization that can be made for streaming? The architecture will definitely be simpler with the main-frame paradigm. Maybe if there's only one client it could be feasible to send some batteries included chunk of frames that can be easily constructed based on the next set of inputs. Is there any hint that progress can be made in this direction?
If the latency problem can't be solved, I'm definitely bearish or game streaming. Latency insensitive games for the most part are not difficult to run on commodity hardware.