look at Andrew Trasks work on crypto+ai if you haven’t already. He wrote Pysift which is a framework for federatedml (paradigm here crypto and ai coexist and take us forward in the age of edge compute and chaos)
Federated ML is awesome but doesn’t seem like a hard necessity. Seems more like an optional, nice to have. The most powerful models will prob still be the ones trained on the sum total of public human knowledge