I don't think anyone has yet trained on all videos on the Internet. Plenty of petabytes left there to pretrain on, and likely just as useful once the text/audio/image pretraining is done.