Except the article is saying the same as I’ve written above:
Especially DLSS v2.0 offers significant performance improvements. NVIDIA however is utilized dedicated hardware or DLSS through the Tensor cores, whereas AMD would need to run it over the compute engine.
In this case, the actual CAS Shader does the job as well and probably in taking less resources than DirectML.
The value proposition to me of DLSS is not just rendering as-if 4K from a 1K image, it is doing so with silicon which does nothing otherwise. In other words, you get not only higher res picture with lower GPU use (1K render instead of 4K), but you also therefore free GPU CU which you can use to enhance the rendering (more pixel Shader effects per pixel).
Now honestly it is really possible NVidia drivers implement the DirectML API via their existing DLSS technology in which case it would be indeed a huge win.