[alfalfasprout]

No one stops you but it's a huge time investment. It's not only CUDA... it's the massive amount of optimized libraries NVIDIA provides. For example, CuBLAS (linear algebra libraries), CuDNN (libraries for deep learning). These things really need massive engineering teams to pull off and are very tied to the hardware.

[shmerl]

So there should be combined effort to make such libraries that are cross hardware. If many people need it, why don't they pool resources and create open libraries like that?