Nobody knew how well models were going to scale up in usefulness, by (mostly) just scaling up training data, training times and architecture sizes.
Despite already using high data, high computation, low complexity solvers (I.e. deep learning), the Bitter Lesson continued to surprise!
At the point it became clear scale of data, parameters & compute were the top three (!) high bits for success, they would have realized (correctly) that staying a pure non-profit would rapidly make them irrelevant.
Setting aside any personal greed by anyone, the problem is still there.
Staying or reverting to a pure non-profit doesn’t provide the resources the non-profit’s mission requires.
Which was to have real safety impact at the cutting edge, not just publish safety related papers from the gallery.
TLDR; They absolutely need profits as part of their structure, or their non-profit mission is unattainable. But all the outside pressure they are getting as they adapt, is necessary to ensure the new structure remains committed to the mission.