CUDA: NVIDIA's Moat
CUDA is Nvidia’s proprietary software layer that is optimized to work with its GPUs. Many data libraries including various open source libraries and software stacks are optimized for CUDA. Thus, if you are AMD or another chip player, you are boxed out from those libraries and stacks unless you have a method of translating your software layer such that it is interoperable with CUDA. AMD apparently has such a translation effort, but it is a work-in-progress (per former Google CEO Eric Schmidt).