I would guess compiled Fortran accounts for the majority of processor cycles on supercomputers worldwide, given how common it is in applications modeling weather, fluid mechanics, nuclear physics and quantum chemistry.

Yes, NumPy depends on BLAS/LAPACK libraries which are written on Fortran. In turn TensorFlow and Theano which used for implementation of most state of the art deep neural networks depend on NumPy, so Fortran is still alive and gearing such modern research.

All of this is wrong. Numpy does not depend on fortran code and TF and theano implement their arithmetic without fortran code.

From http://docs.scipy.org/doc/numpy/user/building.html

> Various NumPy modules use FORTRAN 77 libraries

    $ pip show tensorflow theano
    
    Name: tensorflow
    Version: 0.7.1
    ...
    Requires: six, protobuf, wheel, numpy
    ---
    Name: Theano
    Version: 0.8.2
    ...
    Requires: six, scipy, numpy

You are right that arithmetic in both of them is not written on Fortran, but the packages do use NumPy as you can see in the pip output.