I think the idea is that blas + lapack are still at the foundation of basically the whole stack. Back in the day (or at least when I last looked), numpy was very much built on of them, and the rest of the data science stack was built on top of numpy. No idea how much these things are still the case.

At this point it has been translated to C and now they patch the C directly but: https://github.com/numpy/numpy/tree/main/numpy/linalg/lapack...