I can see how it's slightly more beneficial for the artist, but I would hold my reservations whether Blender can properly triangulate the mesh afterwards. I've had experiences as shown above that required effort on my part to correct. Perhaps I missed the important part where he describes the intelligent algorithm Blender uses. He tends to babble in the video.
Indeed combining textures into one will render fast. If lighting remains constant, it's the best way to go. If you're developing for mobile devices, it's the only way to go Baking normal maps is fun too. I'm not sure how to do it in Blender, but I wrote my own program that does it for me. If you check out the Stanford 3d scanning repository, you can download and test with these models.