What exactly does Valve stand to gain at all from funding a CUDA compatibility layer targetting mainly machine learning software? They’re a video game company. Arguably the most gaming-centric thing CUDA is used for was explicitly discarded in the blog post (“Raytracing is gone”).

Machine learning is massive now and there are many companies who could be interested in funding this kind of project. I’m pretty skeptical it’s possible to make any good guesses with what little info we have.