I think one could also imagine a world where there will be a cannibalization/consolidation on the lower end of the market because of the democratization (which I am not saying is bad or good because I am not sure about the effects ML has on people and democracies, but that is another discussion).
Rather, I observe that nobody today would hire someone to (in essence) build numpy functionality. Yet there are a lot of jobs in ML/AI currently which amount to building things that will be commoditized within years (and be used by everyone like numpy today). So in my opinion learning AI basics will not necessarily lead to more AI jobs or startups because these basic features are commoditized via big tech open source.
Second, on applications:
My research is actually in the applied domain and that's where my argument mostly stems from. Doing novel, valuable applications in AI requires concentrated resources and expertise. For example, if you think about self-driving cars as an extremely high-value application, they require highly specialized domain expertise in various domains and machine learning combined with massive compute resources.
Simple applications (e.g. applying basic machine learning tools to common workflows and processes) are prone to undergo commoditization precisely because they are obvious. So I do not think the gatekeeping argument holds because there is no way around obtaining deep and wide expertise to generate valuable application IP.
So for a country wanting to focus on applications, I still think there is a good argument that concentrating resources is a more successful approach.