I thought i read somewhere that they don’t, which is what makes them more performant.

It's possible that the server just forwards a selection of packets from each participant (hopefully dropping some from those who aren't speaking or moving) and the client stitches them into the view.

Sure, that would be the SFU doing it's job, but my point is that you aren't maintaining a connection per participant, but one connection with Zoom itself where you get all the video data. It's true that could be in the form of a single video stream of the entire gallery, or separate streams of only the participants visible to you.

Google Meet does send each video participants video stream seperately. You can tell this by disconnecting internet and then clicking a (now frozen) participant - they become fullscreen but low-res.

That means the client must have had that data locally rather than needing to wait for the server to send it.