I would imagine it would be computationally untenable because it would mean the videos have to be re-encoded every time the ads need to be changed, which would be frequent enough to be very expensive. Plus, I'm not sure if they keep the original videos stored somewhere to re-use as a source to prevent degradation from re-encoding.
i would imagine the video stream is spliced, such that the ad video stream is spliced into an existing pre-encoded video stream, may be between some keyframes or some such.
This is how most livestreams works today - you have separate chunks of video streams that are spliced together.
The only real problem is if the end user can detect (reliably) where a splice _could_ happen, they can just ignore the spliced video. This is what some old twitch adblocking methods did iirc - and just show a blank screen.
