They could probably have gotten similar results by using a combination of numpy and numba. They could also have just written a C extension which they basically did. The key is to get the final code to run both in parallel and vectorize on your exact hardware. So there are compiler flag choices too if your using C. Nice though.

I usually read news like this so this is my opinion about this subject: python is a good language for prototype, PoC, and general tasks where performance and security is not really important. For other uses languages like C, C++, Rust, Fortran, and libs like MIPS are better option. I think every language has its pros and cons, so the optimal is select the best language for the particular task, no try to change the language for be the best in all aspects (it is impossible).