They still have (2), right? An image classification startup that provides services to other companies will see more images than competing firms, which lets them train their models on more data, which gives them better quality models than competitors. Ditto fraud detection, ad targeting, or product recommendation (within a specific vertical, at least).

It seems like it'd be hard to build a durable competitive advantage without either owning the user relationships or owning the data, correct? Such a company would be just an algorithm and some tools around it, which are relatively easy for a competitor to copy.