I suspect it's likely. It's really hard to argue that there's nothing "sufficiently transformative" about turning some source code for a web api into a machine learning model capable of presenting code suggestions for a wide variety of different types of problems.

The actual issue isn't the training of the models, but the use of the resulting output: Just because training the models was fair use, doesn't mean that a case where you use some code the model regurgitates from its training data verbatim is.