Again GPU vs CPU, many people ignore one very important fact: Big companies dont care about money, and with extremely fast Xeons and now als the AMD Maple Architecture (32 Cores!) GPUs are getting behind flexibility/price (not performance/price) ratio.
Not to mention if you have 32K multilayer textures like in Gravity.
Now, not so fast. The future of GPUs is (still) just on the rise. Performance wise they ARE the future. The problem is if you want to make a huge renderfarm that NEEDS I/O coherency. We are not there, yet. Vega and Volta will change that. There were already Vega prototypes at CES with an 1 TERAbyte (yes you heard it right) SSD, beside 16GB HBM VRAM on GPU! They showed off 250 GB Texture PT in realtime!
So it come, but the big renderers will not jump on the GPU wagon UNLESS you have the same Memory (nUMA or UMA cluster) and FAST interfrabrics from GPU to GPU. Both AMD and NVIDIA are still infants regarding that now.
Never forget, a renderfarm is high-performance computing. And they dont care about money (primarily), but “can we 500 to 5.000 people at this or that VFX/3D company do that?”