Do you see proprietary datasets holding up as a moat for AI based applications?

With foundational models being easily available to teams where do you think defensibility and moats in AI based software and applications will arise? Do you see synthetic data and proprietary data will have similar impact in terms of model customization .