What needs to happen for GPUs to be commoditized so that an infrastructure builder is relatively indifferent to using GPUs from Nvidia, AMD or any other provider?
Either Nvidia has to open up CUDA to third party providers, or someone (OpenCL?) has to finally manage to create a high level cross hardware abstraction that has all the features of CUDA, can match CUDAs performance on Nvidia hardware and is as easy to use as CUDA. Honestly I'm not sure which is more unrealistic.
The CPU analogy for the GPU case isn't AMD and Intel, it's x86 and ARM.