Not all results need rendering, ford.
If an AI does nothing in a frame, it presents as a bug. Sure, macro AI functions can span seconds or even minutes. The problem is that AIs have to branch: solve for most situations to find the best option. It's CPU, memory, and time intensive. The more possible outcomes, the worse it gets.
Remember that AI's just now have figured out how to reliably win Chess. Now imagine 3D games with thousands of pieces and an infinite number of moves for each. It might be possible if you had a core for each piece but now imagine trying to run the same software on a quad-core processor. A single frame on the quad-core will take 4-5 minutes to process AI for.
Hell, to make AIs truly how they need to be, x86 won't work because of the overhead that comes with every new thread started. There would have to be global, hardware register mutexs so that software only has to check that one variable to know if the AI is done for the frame.
Developers have to code for the lowest common denominator. Choking the CPU/RAM leads to far worse outcomes than choking the GPU.
...and yes, still only grazing the surface of the topic. TL;DR: everyone will go to console gaming before there's a monumental shift to CPU focus.