Also, not a single provider can pull the rug if I'd like to use a particular model.
It's trivial to rent all of the above components in a competitive market under different periods, you may simply rent the token output from someone who put in the effort on this as well.
An open frontier model induces margin compression on training and inference prices across the industry.
Worst case you can still archive the weights and hope for capable hardware to get affordable. At the same time this archive is the base line for a new frontier lab to start over if required. I see open weights as a pure upside even though I can't run the bigger models in my home lab.
There's something to be said for making these technologies equitable to those with less resources, but that's due to the cost of hardware and the nature of how transformer models scale... not really anything to do with "open"ness.
You don't need much capital up front, you can just rent servers to run it. It's going to be more expensive than openrouter though because you probably don't have economies of scale if you're the only customer.
I'm hoping that all software will be getting opensource clones on par with original software 5 months after the closed source software release.