NVIDIA are easily the ones holding it back at this point. Both AMD and Intel have OpenCL 2.0 support on windows, and Nvidia hasn't even shipped support for OpenCL 1.2. Nvidia obviously want OpenCL to die off and CUDA to do well.

Well, there's also Intel who decided to ship their accelerator product (MIC / Xeon Phi) without OpenCL support initially. Their big marketing angle was that HPC programmers could just use their existing OpenMP x86 code. This shows how Intel had no idea in what they're getting into, because accelerator programming needs to deal with different problems than on the CPU, regardless of whether you have a simple CUDA core or a MIC x86 [1]. Results have so far been rather poor because of that. Kepler GPUs with CUDA support are for that reason still the best tooled accelerators for HPC programming.

[1] First of all, when you need on the order of thousands to tens of thousands of threads to saturate your cores, memory becomes a big issue again. Your threads have maybe 400k usable memory. Memory bandwidth is usually even a more important limitation - these systems have 2-3 times less bandwidth per physical thread than CPUs. This can be mitigated using a programming model that uses bandwidth very efficiently (e.g. stream programming like in CUDA, OpenCL) - an area where GPUs are traditionally more advanced, since they deal with this limitation since a long time. Don't get me wrong though, Intel will probably catch up at some point.