Support for all the CUDA code that's already out there. Not that I've used Metal, but that's what seems to be the reason behind OpenCL's lack of adoption
Sure. But there is also a lot of Metal code out there in order to make iPads fly. And they really do fly, I was amazed what kind of performance I could get out of them. For example, I had an algorithm running on my iMac Pro, programmed in Swift, utilising all of its 18 cores fully. It took about 5 minutes on a typical example. Then I recoded the algorithm for Metal, and it runs on the same example on my iPad Pro in under 10 seconds.
So, my bold prediction: Apple is going to shred Nvidia to pieces within the next 10 years.