It was already in wide use for scientific computing by 2000, due to the comparative ease of writing interfaces to C code. The main idea was to use Python as a glue language to "steer" high-performance computing.
The Python/C API was easy to learn and use, Python's reference counts worked well for C-based objects, and it was easier to build non-trivial data structures than Perl or Tcl, which were its two main competitors at the time.
(Tcl extensions required manual garbage cleanup, I remember Perl's extension API as being rather complex, and I had to read the Advanced Perl manual to understand something as simple as having a list of dictionaries.)
I’d say also it was more at war with node until data science took off.