Nice article, though the analysis is a bit simplistic (Facebook and Google's motivations for making Pytorch and Tensorflow open source are much more strategic than reduced training costs and revenue from hardware).

It's also strange that having made some effort to investigate the space of open-source licensing models, the OP speculates on creating a license that allows for dual licensing but doesn't have the GPL's virality, but is seemingly not aware of the LGPL, or of the relatively common practice of giving a blanket exception for add-ons through some specific interface (such as plugins) to a GPL'd application.