We're talking about a whole lot of 3D rendering, many polygons, and a whole crapload of textures. Easy pickings for a smartphone native app, where just about every major phone on the market has obscene amounts of hardware acceleration under the hood.
Not so easy for desktops - whose graphical performance is often nearly non-existent. Not to mention you take another huge performance hit from WebGL... which also has poor penetration. So not only would most users be unable to use your newfangled 3D maps, many of the ones that can, will have an unusably slow experience.
This is one of those places where native apps really shine.
Current JS and WebGL performance is well within the range needed to make this work. MapsGL is already rendering a quite large workload, in fact, my intuition is that the 2D rendering, especially with rasterizing label text all over the place, is likely more expensive than navigating an octree and rendering a mesh.
2000 - "Run this on a phone? You're crazy" 2012 - "Run this on a desktop? Insane!"