I feel like all the useful stuff you have listed here is more like advanced machine learning and different than the AI that is being advertised to the public and being mostly invested in. These are mostly stuff we can already do relatively easily with the available AI (i.e highly sophisticated regression) for relatively low compute power and low energy requirements (not counting more outlier stuff like alpha fold which still requires quite a lot of compute power). It is not related to why the AI companies will need to own most of the major computational and energy innovations in the future.

It is the image/text generative part of AI that looks more sexy to the public and that is therefore mostly being hyped/advertised/invested on by big AI companies. It is also this generative part of AI that will require much bigger computation and energy innovations to keep delivering significantly more than it can now. The situation is very akin to the moon race. It feels like trying to develop the AI on this "brute force" course will deliver relatively low benefits for the cost it will incur.