There is no moat. https://newsletter.semianalysis.com/p/google-we-have-no-moat...

What will close the way back?

How do you reconcile these ideas with the fact that cheap open weight models are only slightly behind the state of the art?

If anything, I would bet that next year you could get today’s flagship performance for significantly cheaper via an open-weights model.