I think that's how they are preparing for the imminent commodification of their Deep Learning space - by milking the cow as much as they can before they are forced to compete on price. Once TF/Torch/MXNet can reliably run on AMD/Intel hardware without changing a line of code and with comparable performance, they will feel both consumer and data center pressure and their high margin business might evaporate quickly. Still, with CUDA/cuDNN they were the pioneers and the only relevant party so they deserve some ice cream for that.