Nowadays, installing pytorch or tensorflow with pip also installs CUDA drivers, see e.g. https://www.tensorflow.org/install/pip
Since when / how is that possible? My understanding for a long time was that pip could only install python packages and that one of the most important benefits of conda was that it could also install other stuff.
With pip you can install anything that was packed as a Python wheel. For example you can install Cmake with pip install cmake; it will be installed from https://pypi.org/project/cmake/#files where it has been packaged.
pip install tensorflow[and-cuda] installs CUDA drivers packaged as wheels: https://pypi.org/project/nvidia-cublas-cu12/ , https://pypi.org/project/nvidia-cuda-runtime-cu12/ and so on…