Good point! It's not just the LOD, it may be also the fact that the parallelization makes GPUs a better fit for subdivision than CPUs, and surely there's a matter of connection bandwith to the GPUs: the vertex coordinates, as well as lots of other vertex-specific data, needs only to be sent for the base vertices, the new vertices get their new values from interpolating old values, and the textures like displacement map control the difference from interpolated value to desired value. Of course a texture would have been just some weird compromise of resolution of the sent data, except you don't have to provide a texture for every attribute, and more importantly, such a texture might be static (e.g. if encoded in normal space it may work throughout an animation of an object).