well said, Haraldzz. Regarding the jaggies when moving, though, a lot of people say that if you want MSAA it'll require more samples than it would otherwise. Up to several times more samples, in fact. Each sample is a ray cast, which is expensive. You're already doing one for every pixel on the screen.
I contend that MSAA is only good for an antialiasing factor of 2x. Beyond that, use some weak FXAA for a post-process filter. It tends to slightly blur some stuff, which I almost find desirable in many cases.
Another issue that people don't tend to bring up is lighting. The "big" thing about voxels is that they're inherently able to do ray cast lighting and shadows, which is cool, but also stupid expensive. Someone advocating for voxels is nearly as absurd as someone advocating for realtime Mental Ray renders with super max settings. It's just simply not going to happen, at least anytime soon.
So the solution is to just use static lighting for voxels, and I guess deferred lighting for dynamic lights. The big
(and I mean BIG) issue with doing static lighting on a per-voxel basis is that it would disallow you from reusing voxels. All voxel lighting information is unique, which means that all voxels must be unique. There are some ways around this, but they're all kind of hacky and still consume tons of storage space and memory.
Once again, we have a better chance of adopting REYES.
Now that my voxel rant is over
I will admit now that I am having trouble determining whether I want 100% realtime or to go with megatextures. I am getting pretty disappointed with the artistic quality of shadows when using shadow maps. One of my friends even told me he prefers baked lighting because it gives him more artistic control.