Once GPGPU got more commonplace, researchers also doing AI, finding suitable GPGPU tasks, used the more user friendly and advanced tooling, which was CUDA.
So, "NVIDIA inventing AI on GPU" suggests you do not know much about the history of AI work on GPUs. But, feel free to correct me.
https://en.wikipedia.org/wiki/Shader
https://www.khronos.org/opengl/wiki/History_of_Programmabili...
"The GeForce 3, the first NV20 part, contained the first example of true programmability. Despite NVIDIA being a pioneer of highly configurable fragment processing, its programmability was in its vertex processing. The GeForce 3 was the first GPU that brought programmabilitiy to consumer hardware."
The AI on GPGPU was possible only because Nvidia provided a library with matrix and activation functions running on a GPU at speeds far surpassing CPUs and designed a fairly nice API that anybody could understand. There was nothing like that before and that's why they had such a foothold with academic institutions. Of course they didn't invent AI but made it possible to run on their GPUs and actively helped researchers to do that while Intel and AMD slept (well, Intel at least tried to do that on CPU with MKL).
Tool makers do not get attributed inventions made with said tools.