It's hard to say. Good moats, at least mid-term, are rarely algorithms. It's usually data and infrastructure.

However, for some applications I am interested in, I think that robust representation learning solutions could give a significant edge.

But that is mostly an open problem in high-dimensional spaces.