As the sibling comment points out, gp is clearly talking about the source mesh from which the normal map is generated. But that there is never a higher-detail intermediate mesh during rendering is also wrong since the engine could use tesselation or geometry shaders to do just that. Doing it all in the fragment shaders means you either need to limit yourself to small displacements, have visibly wrong rendering at oblique viewing angles or add a lot of complexity to the fragment shader to simulate self-occlusion via raymarching and still end up with visibly wrong rendering in some cases.
