With regards to your thoughts about the hardware being expensive, I found the following [1] article interesting. I'm mostly referring to the vendor lock-in part with CUDA.

[1]: https://streamcomputing.eu/blog/2014-08-05/7-things-nvidia-d...