> With OpenCL, you can develop on laptop, test on workstation and deploy on actual servers. With CUDA, you're trapped.
GPGPU developers usually prefer hardware with Nvidia GPUs inside, so you can do all this too. I wouldn't call Nvidia anti competitive, just like I wouldn't use that word for Apple. They saw a niche (HPC/smartphones) when no one thought that market to be attractive, jumped in with proprietary technology that others later tried to reproduce (OpenCL/Android) and they kept the market leadership in terms of profits by reiterating their product.
I always get disappointed when I see something that CUDA specific. "Hey that's interesting machine learning package! Oh wait, that's Nvidia specific."
With OpenCL, you can develop on laptop, test on workstation and deploy on actual servers. With CUDA, you're trapped.