I wish there was a more open adoption of CUDA like functionality. I use an astrophography imaging tool called PixInsight and it's very parallel math heavy. It supports CUDA, and you can offload some tasks to the GPU and it makes a world of difference (computing a mask of stars takes seconds instead of 5-15 minutes). It's certainly got me locked into NVIDIA.
Kind of. Vendor lock-in is only a part of why it's the only game in the town. Another, larger part is that AMD's software support is awful, and also that ATI/AMD were ignoring GPGPU/ML stuff for a while, letting Nvidia indoctrinate academia into using their stack.