most likely many videos are precalculated already, but what they do is making screenshot of current tab like https://developer.mozilla.org/en-US/docs/Mozilla/Add-ons/WebExtensions/API/tabs/captureVisibleTab, and send it for processing to their ai, so it far from frame by frame
About vercel I doubt that they will allow to store and process your ai model on their runtimes.
nextjs also can't help with this in any way, you need your custom backend that will work directly with your model.