First of all, even if the models never reproduce any of the copyrighted content verbatim, they will soon be good enough that they will be able to replace the work of the people [1] that produced the content the AI was trained on, like writers and programmers.
Second of all, the nature of the use of the copyrighted content is purely for-profit, and the ability to squeeze the profit out of the best models will stay with the largest corporations -- effectively transferring wealth from the people that produced the content, to the corporations.
We need to find way to make sure the access to this data the models were trained on, and the resulting models, is open and fair. Who wrote which paper [2], or who supplied the initial GPU training, should not really matter that much in the grand scheme of things.
[1] or at least cause downward pressure on their earning potential
[2] you could use similar arguments here they use for using the copyrighted content -- each paper (except maybe the Attention is all you need paper) contributed only marginally