Now, now, now, let's clarify a few things.
1. There will not be a massive overhaul of hardware within the next 2 years. Why? Because we have reached the absolute limit of silicone technology. We can make our chipsets smaller, but it's a long process. We can add cores, but that increases power consumption and heat exhaustion. We can no longer use higher frequencies. Gradually processors will be more and more like today's GPU's - composed of minature, highly-specialized cores and SPU's, much like the PS3's multi-core system, but on a microscale on a single chip, which will improve performance, but not by "a whole lot" in the sense of "revolution". We're about 7-10 years away from the first widely-available processors with Graphene-based SPU's, since this particular material is no longer a rarity and can be produced in labs. In fact, since last year, a technology to produce it cheaply is widely available. THOSE chips allow frequencies as high as (theoretical) 1000Ghz, and current Graphene "CPU's" are clocked 100Ghz and up, HOWEVER they have troubles with operations on floats, so it's still a matter of "the future". Another option would be nano-tubes of carbon, but those are expensive to produce, fragile and generally "stuff of science fiction" as far as a desktop is concerned for at least the next 15-20 years.
2. There WILL be graphic overhauls. Engines are enhanced every year, the Polycount rises steadily and it will rise, however significantly slower now than, say, 5 years ago, due to the previously mentioned reason.
3. As far as Voxels are concerned - yes, it's possible to use them in video games. Yes, it's possible to use them as atom-like molecules. Yes, it is possible to animate groups of them rather than individual voxels, contrary to the common opinion, as long as they are dynamically attached to a standard skeleton made of Bones. In fact, I could even procure readily-available solutions for animating Voxel models efficiently, but I trust in your Google abilities. No, it's not cutting the model into pieces and making it look unnatural - I'm talking about full flex-like animation.
4. Notch's arguments are true as long as we assume that the ammount of Voxels is fixed and that they are affixed to certain positions and treated as "solid" points. What Euclideon does is dynamically producing more or less Voxels "around" a fixed "mesh-like" grid of a given model depending on how far it is from the viewer for the best viewing experience at the least CPU/GPU resource usage. Yes, it will be memory monging, but only when given models are treated as "individual" ones. If each type will be saved as an entity in RAM, rather than treating each stone individualy, you could save tons of space and clone said stone almost infinitely. Of course said "stone" would have to become an individual entity once it's damaged, animated etc. etc., but those are "marginal peculiarities". I'll support Carmack on this one - it is possible, but not "now", maybe not so much because of hardware limitations, although more RAM would always be welcome in the scenario of Voxel usage, but because all the big players in the market invested billions upon billions of dollars on enhancing the Polycount, thus pushing the GFX Chip and CPU industry forward. If all of a sudden we'll loose the need to have top-notch hardware, companies lose on their investment, Intel and AMD lose on their research into better, faster CPU's and AMD (again) and NVidia loose money on developing such innovative structures as CUDA cores or new types of Pixel Shaders, which would be rendered relatively useless since "mid-tier" hardware is "theoretically" capable of rendering "some blocks". This is definatelly not something either of those will vouch for.
"Unlimited Detail" sounds "fishy" - it's obviously limited by hardware, there's only so much you can stuff into an Array, even if it's dynamic, so let's not treat it as "Unlimited", rather as "Really Close to Unlimited". Not just that, let's remember that you cannot "texture" Voxel models - instead, each Voxel has its own colour values, so I do hope there's a smart way to transfer these RGB values onto the "external" Voxels or else this is pretty pointless.