Not sure about the conclusion regarding NVidia value capture. I imagine the context for many applications will come from a physical simulation environment running in dramatically more GPUs than the AI part.