How to build moats for AI products?

Been thinking a lot about how to build defensibilities for AI-based products. A lot of AI founders seem to think proprietary data is the biggest moat. Another way to look at this is through the lens of economies of scale - traditionally one of the most common moats: Economies of scale work best when you have high fixed costs, so that scale can help you create a cost advantage. It takes a lot of time/money/investment to obtain high quality proprietary data. But I think most people are forgetting the ‘scale’ part of the equation. High fixed-cost data without scale won’t be an effective moat. Would love to hear if anyone else agrees or disagrees?