Because cloud compute does not help with graphics, it does not free GPU ressources to render better images and will not be much good for anything interactive (unless it's very static... thus not very interactive).
Things you could pre-compute may as well be defered locally (you can renter certain parts of the image only once every 2 or 3 frames locally...) it will probably be better and much more reliable.
But more importantly, MS azure is built on x86 CPUs, not GPUs, that's not a very good use of ressources (think energy cost) to render high definition, high details graphics on those compared to a decent GPU.
Now things that are indirectly related to graphics like physics, they would be even worse, imagine having a million objects falling around a scene because of some explosion, now try to send coordinates for all of these all over your internet connection with the colision data etc. for every frames... that's even worse than rendering the whole game remotely (like gaikai does by example) because there is just way more data to move around... if there are not that many objects, well the simulation can run locally... same goes for dynamic shadows, in a decent resolution...
Could compute is great for things you can render asynchronously, then send over the often huge datasets associated with the results, for the uses people think of when they mention using it in games they forget that often generating the information on the local machine is often faster than generating it remotely then sending it back to the user... that all may change as internet connections become faster, but so will consoles...
