It's not, it's a statistical model of existing text. There's no "genAI", there is "genMS" - generative machine statistics.
When you look at it that way, it's obvious the models are build on enormous amounts of work done by people publishing the the internet. This amount of work is many orders of magnitude more hours than it took to produce the training algorithms.
As a result, most of the money people pay to access these models should go to the original authors.
And that even ignores the fact that a model trained on AGPL code should be licensed under AGPL as well as its output - even if no single training input can be identified in the output, it's quite straightforwardly _derived_ from enormous amounts of trainings data input and a tiny (barely relevant) bit of prompt input. It's _derivative_ work.