I think the argument here is that creating quality assets with those levels of detail is expensive, not that you can create large and unwieldy meshes by not understanding how to model efficiently... that goes without saying.
A 10 mio poly mesh with a typical XYZW+RGBA+UV+UV2 = 12 float stream format would require 12x4x10 mio = 480 MB of GPU memory per copy. Skinning needs weights (200 MB) and at least a second copy, so one single 10 mio poly mesh would eat up about 1.3 GB of GPU memory.
So even with ray-traycing, optimizing models for a manageable polycount will remain crucially important.