This would be a really nice move to have CUDA independent ML stacks.
Computer shaders are available in Vulkan, DX12 and Metal, with potentially better performance than OpenGL. This covers all the platforms they mentioned, except for web browsers.
A better approach, with higher performance, and less platform variation (than OpenGL), is to use WebGPU, which is available on all native platforms, plus web browsers. The limitation is that WebGPU is still a work in progress, still being changed, only available on browsers behind a feature flag. There is supposed to be release candidate later this year, but a number of people are already writing code for the new platform. So I figure that is the future for this kind of project.