We need to know the maximal possible bounds of models to know if they should be rendered.
However, computing that requires processing the animation and checking every frame, which is quite time consuming.
This computation is done in "model-space", and depends only on the mesh and the animation. As such, it can easily be cached.
Further, skeletal animation have fixed addresses in memory, so can be used to identify an animation efficiently.
This is basically P223 but not terrible.
On Combat Demo Huge, here is a simple profiling difference (red is svn):
You can see the lag spikes from all units recomputing bounds when they change animations. It actually takes over 500 ms the first time on my computer, which is crazy. By caching these results, the lag completely disappears.
This is unlikely to have a huge impact in a regular game, but it can't really hurt.