Many parts of tensorflow required Python- at least when I worked there a few years ago, it was nearly impossible to compile XLA into a saved model and execute it from pure C++ code.

The claim was that Python is only used for prototyping at Google, not that people are writing the AI frameworks themselves in pure Python.

JAX is obviously implemented in C++, but the scientists running the training runs which cost millions of dollars are writing lots of python.

Rather than repeat myself: https://news.ycombinator.com/item?id=40182940