I'll use it for web since there is no alternative, but for desktop I'll stick with an OpenGL+CUDA interop framework until a sane, modern graphics API shows up. I.e., a graphics API that gets rid of render pases, static pipelines, mandatory explizit syncing, bindings and descriptor sets (simply use buffers and pointers), and all the other nonsense.
If allocating and populating a buffer takes more effort than a simple cuMemAlloc and cuMemcpy, and calling a shader with arguments takes more than simply passing the shader pointers to the data, then I'm out.