Well even the free models aren't free at scale. And if you're paying for the inference anyways you might as well pay for one in a model which was specifically truncated to your needs.
To provide some perspective here, SD can be retrained on comparably small samples to generate images for very specific needs (example see [1]). So there is certainly a bunch of $ to be harvested by generating very good training sets for industry needs, say for instance mockups in architecture, creating logos and brand styles or just nice looking designs for consumer products.
sure, stock photography may just disappear as a business model.
while SD is freely downloadable, it isn't energetically free to run inference,
so for example (and as an off the cuff speculation), opportunities could arise to provide execution of these models from places where the cost of energy is lower
we can never confidently predict what the business models will be when such an impressive new technology arises.
If everyone can just run their own prompt against free models, wouldn't it be more likely that no one pays for it anymore?