This is often very overstated. For example in government research labs there are often huge, critical pieces of software in both C and FORTRAN. I worked in one of these labs in the 00s and about 80% of my team’s work required running reports and simulations through a huge system written in C between the late 80s and mid 90s. Even generating plots for conference presentations was done in pure C with a special DSL tool that took in a static plot conf file and wrote raw postscript outputs.

For the nuclear industry, this is true as well. A lot of it is because of the original codes were written in Fortran, and thus there is some institutional lock-in even for new stuff. Everybody in the organization already knows Fortran, it's quite fast, and there are good libraries for doing all of the standard number crunching.

However, my unscientific and unsupported observation is that in fields that have emerged after Fortran went out of vogue as a general-purpose language, this same process happens again, but just with a more modern language. SciPy etc. are quite common in fields like genomic analysis, and I suspect that in 20 years, this will still be the case even if almost nobody else is writing new Python code.

From what I see, majority of scientific computing code is now in "Python", i.e. C foreign libraries it uses.

Many of those libraries are actually C++, like eigen or Tensorflow.