We would like to support opensubdiv in some form in the future. Subdividing arbitrary triangle meshes with Catmull-Clark subdivision isn’t useful. Really, a subd surface should be treated like a different input data type, a quad mesh with edge weights specifically built in a modelling package as a subd cage.
Once this data can be imported in the engine, Pixar’s runtime library to dynamically tessellate is definitely cool but is poorly suited to the strengths of hardware given game performance demands. HW tessellation in general is poor for performance per poly. It is also a pretty complex system which would be a lot of work to integrate.
My expectation is that the best bang per buck for game performance is to import subd data, tessellate and mesh simplify as a preprocess to generate optimal LOD meshes.
All that said, this isn’t on our schedule at the moment.