Then explain why everyone buys Nvidia cards for GPGPU?

Because GPGPU software stacks do a very poor job of supporting either of OpenCL or Vulkan Compute, which are the main free alternatives to CUDA. This is especially problematic in the ML space.

Yes that was my point which contradicts what foxbluff wrote >> And IME it is the proprietary drivers that usually offer least value and most problems <<

CUDA does not inherently offer "more value" than Vulkan Compute, though. It was simply developed earlier.