While it could theoretically be done on device, it would require the device to have dedicated hardware that is capable of doing the processing, so it would only work on a limited number of devices. It would be pretty easy to test this if a known modified video were available.

You are right that nvidia cards can do it for games using DLSS. Nvidia also has a version called RTX video that works for video. But are they could to be dedicating hardware for playback every single time a user requests to play a short? That is significantly different than just serving a file to the viewer. If they had all of these Nvidia cards laying around, they surely have better things that they could use them for. To be clear here, the ONLY thing I am taking issue with is a comment that it seems that youtube may be upscaling videos on the fly (as opposed to upscaling them once when they are uploaded, and then serving that file 1 million times). I'm simply saying that it makes a hell of a lot more sense any day of the week to upscale a file one time than to upscale it 1 million times.

They could do that without upscaling. Upscaling every video only fly would cost an absolute shit ton of money, probably more than they would be making from the ad. There is no scenario where they wouldn't just upscale it one time and store it.