To mirror what the above posters said and add a few, imo, three important factors:
1) Information density is much, much better. You can have a (full quality) normal in 2 to 3 bytes, while vertices may easily reach 32+ bytes.
2) New techniques, such as parallax occlusion mapping, will remove most of the traditional normalmaps' limitations.
3a) Normalmaps and similar approaches have "perfect" LODing, and it's called mipmapping
3b) The transitions are always smooth, while it is hard to create good-looking geometry LOD transitions. While rendering a 1M polygons model at a close distance is feasible, rendering 100 of them when they're 250m away is not, and continuous LOD at those polycounts is stupidly expensive for geometry.
To illustrate these points and others, imagine a 2048x2048 16-bit parallax/displacement map with in-shader normal computation and parallax occlusion mapping. Besides for certain "external" edges, you'll get similar quality to that of having 4 *MILLIONS* Polygons. And the full texture only takes 8MB, and no real further LOD scheme is required but perhaps imposters. I'd love to see REYES beating that! *grins*
Uttar