For machine learning (un)fortunately, lots of the stack runs on Python. Lots of ceremony was done to circumvent GIL (e.g. PyTorch data loader). “Reasonable performance Python” I imagine is actually something in huge demand for lots of ML shops