At which point you'd basically just going temporal AA which already exists. DLSS works because it uses specific tensor cores to work, whereas a move to temporal would have the same hit as TAA does now.
Which is probably part of why AMD went this route.
They had a specific objective in mind and surely they talked with partners before implementing such solution.
Implementing a temporal upscaler might not have the same value, other than didactic, because a lot of developers already have a working solution for their needs.