I don't think it's possible to build a moat around models at all. The model architectures are public, and there are already distributed group training projects so the compute isn't a barrier. The only moat is data.