This seems like a bad argument when talking about anything other than tech demos. It assumes that the only things in memory will be seen by the in-game virtual camera, so that if you turn your view to look at an object with textures or geometry that wasn't in the original view, you'll have to stream in from the hard drive. I'm pretty sure that even the most aggressive streaming engines save some memory for off-camera content (and little things like audio), at least at a low level of detail so there's something to show while the hard drive catches up. In short, games need memory beyond what's on screen.
On the other hand, that argument also implies the GPU is doing nothing but texturing for the entire frame, since shaders aren't as bandwidth dependent. But if the GPU is just texturing objects for the entire time the frame is being rendered, that doesn't leave any time to apply shaders after the GPU is done; since many shaders depend on a textured object for color data, those shaders can't run before the object is rendered. So I doubt any modern game would be able all of that memory per frame anyway.