This looks a lot like
sparse voxel octrees. As a concept, SVO is nothing new at this point, and id has been considering using it as part of their id Tech 6 engine.
A sparse voxel octree is basically a hierarchical structure for points in 3D space. The advantage of using a hierarchical structure is that you can stop looking at any time, and so zooming works very well: you just traverse the tree until you get so far down that further detail won't be visible, then you render. If the player moves closer, that simply means you'll go further down the tree, but you'll cover a smaller space so the load doesn't really change.
Now, the sparse voxel octree both gives and takes away. It enables detail at all scales, but since the structure is hierarchical (log n insertion and deletion, AFAIK), moving large numbers of points about is going to be
really hard, not to mention actual deformation or changes of the objects themselves. One would probably use an SVO to show world detail: landscapes and "3D textures" - and then use simple polygon skeletons for collision detection, moving critters around, etc.
Sparse voxel octrees work, and they work very well on the kind of static-but-detailed data shown in the demos, so it seems quite likely that's what they're doing. But sparse voxel octrees aren't New And Revolutionary Unique Technology only available to the Australians here.
Here's an nVidia demo,
another demo (not by nV), and
an example of the kind of tricks you can do by combining SVO and procedural generation.