I just hope that they didn't remove the octree voxel cone tracing from the engine like everyone seems to interpret now - which was their keystone technique that they seemed to have spent lots of time (including 10min of their only video to demonstrate the engine itself) promoting this idea.
In the techniques list on the nvidia website, they seem to mention tile-based deferred rendering as the method used.
I guess the octree structure was too costly to maintain. However, release of a gk110 card should be able to eliminate this problem with their dynamic parallelism, which has been demonstrated to make octree construction super-efficient.