DLSS is largely run on tensor cores that do many more, but simpler operations than a conventional GPU is capable of. A GPU without them might be able to do the same job but it would be so much slower it would be useless.
Intel tried to bring XeSS to their iGPUs that lack XMX cores (Intel's tensor version) but had to drastically tone it down so much that it looks like junk. It's a whole different version that can work on other hardware, but is inferior.
Maybe Intel Arc and Nvidia RTX could both use each other's matrix computed upscaling, because they both have significant matrix math processors, if they shared, but the other graphics cards without enough matrix units can't use the upscaling methods in a timely fashion because they simply lack the hardware.
The reason DLSS, XeSS and FSR2 can all work on the same games is because they all take a previous frame and motion data to upscale the current one.
https://docs.unrealengine.com/5.2/en-US/temporal-upscalers-in-unreal-engine/ They are all temporal upscalers that can be all used in a game if any can. But they use different hardware so you can't just use an upscaler on mismatched hardware.