I expect that they aren’t wasting these resources, either. The first useful machine learning models I ever used were from Facebook research. There were other options at the time, but what they created was accessible, fast, and had very usable outputs.
I suspect there is still more useful work in the area happening there.
The problem is that, given past Facebook/Meta reputation wrt privacy, who would trust their "AI cloud" if they monetize it like Microsoft is doing right now?
I suspect there is still more useful work in the area happening there.