That's the thing with Python: A lot of things should be bound by all kinds of limitations, but are in practice often limited by the Python interpreter if not done carefully.

Fundamentally for example, if you're doing some operations on numpy arrays like: c = a + b * c, interpreted numpy will be slower than compiled numba or C++ just because an eager interpreter will never fuse those operations into an FMA.

Numpy is partly written in C but includes a lot of Python code. If you include scipy or scikit learn or pandas, most of the code is python calling primitive numpy C operations. I'd expect that many semi-complex data science programs to benefit from improvement in the python interpreter, especially if they weren't written in super tight numpy code.