Just use render-to-vertex-buffer, no big deal. I think they already do that on appropriate platforms, and the “restrict number of bones per vertex” checkbox is the checkbox that moves it off the GPU and onto the CPU. I could be wrong, though, haven’t read that code in detail.
That being said, perhaps they just do recast to normal/displacement maps on simplified geometry in real time. As long as there’s 5 or fewer pixels per triangle (and ideally, not more than 2 triangles per pixel) it’ll look good enough! It’s actually a problem if you have TOO many triangles for the view size of the object.
Regarding running out of VRAM, that’s what the Quadros are for. But, of course, you can still run out. Then there’s the DGX boxes, but sadly, they don’t take a display out, just running gradient back propagation for neural net training :-/