[throwaway675309]

They won't. They have both the financial resources and honestly probably a legitimate claim that the model produces a sufficiently transformative result that would be considered fair use.

Even if the ruling did go against them, it would likely still be acceptable to train the models since it would be the usage that would be legally suspect.

Which means that in five years anyone will be capable of running these models entirely off-line on their personal machines and no one will ever know the difference.