If you believe a few assumptions then foundation model companies like Mistral look very valuable even if they open their models: 1) few companies are be able to produce top performing foundation models at any given time, 2) this will continue to be true for a long time due to the required expertise, infrastructure, and cost, 3) this somehow translates into a capability commercial customers will need and pay for at scale.

That third step is a bit of a leap. I tend to think that as the open knowledge about how to train models improves the training "recipes" needed to get close to state of the art will get easier to replicate, not harder, with "good enough models" requiring less data and less hardware.