Sparse Voxel Octree (SVO) Demo by Jon Olick

As others have said looks like some sort of voxel raycaster. I wrote one a couple years back that got to about 1FPS w/o too much optimization so it's pretty conceivable that with better optimization and faster hardware you could go much faster. The big problem with building scenes like this is that the datasets get really big really fast. There's a lot of interesting work happening in the area, in particular id's upcoming stuff and the GigaVoxels project. [LINK] [VIDEO] [LINK]

Nitpick: As the name implies, megatexturing solves the problem of texture memory management and rendering, allowing essentially unlimited texture detail on any object. The problem of unlimited geometry detail is still unsolved. Carmack has hinted that id Tech 6 may tackle the geometry problem using voxels instead of polygons, and id actually demoed a voxel rendering system at SIGGRAPH 2008: [VIDEO]

He, and his company, have done research into the field of SVOs to compliment megatexture. I don't know enough to say if it's the same approach, but it's certainly the same field, so you'd expect him to know a fair bit about it. [VIDEO]

I think the biggest problem that people aren't talking about is that LOD for geometry simply isn't as important as texture quality when getting close to a surface. I'm pretty sure the reason they have so many duplicate objects in the original presentation is because they simply ran out of memory to store all the texture data for the scene, because you need really high rez textures to be able to zoom in so close to each object. Notice that most demos which zoom in on the model don't use textures [VIDEO] , because they don't have textures with high enough resolution.