> But where is your startup going to find millions of users for extra training data?

The (big) community. There are existing decentralized approaches to accomplish this. And I am not mentioning the data that could be consumed beyond DMCAs.