But the question is moot because GPUs have proprietary firmware, and NVidia drivers that give you CUDA are also closed-source.
Maybe air-gapping your sensitive machines is a more viable approach for ML.