It's a hardware level feature, though. The reason they didn't support hardware prior to RTX was because they didn't have the tensor cores to do the right math.
FSR is substantially less capable because it can't assume it has the correct hardware to get the throughput DLSS needs to work. I know the "corporations suck" talking point is fun and there's some truth to it, but most of the proprietary stuff nvidia does is either first or better by a significant bit. They use the marriage of hardware and software to do things you can't do effectively with broad compatibility, because they use the architecture of the cards it's designed for (and going forward) extremely effectively.