I agree. Seems like the OP is wedded to the idea of ARM desktop CPUs replacing x86. I find that to be highly unlikely as the x86 ISA is well entrenched.
I don't think that's the only reason, though it is a very important one.
OP seems to focus on how Sandy @ 32nm to Coffee @ 14nm hasn't shown as much as 28nm GPU's that went to 14/16nm. As has already been pointed out, graphics are very parallel so with a die shrink you can just add on more compute units and call it a day. CPU's are entirely different, being general purpose and also having to deal with a ton of dependencies.
Also, remember that 28nm was a fairly crappy process. OK, so maybe it wasn't bad, but it was not ideal for high power/high performance. You cannot compare TSMC 28nm to Intel 32nm or GF 32nm SOI. Both performed better at the cost of density. Therefore, a GPU going from 28nm to 14/16nm gained a lot more than an AMD/Intel CPU at 32nm going to 14nm.
ARM is not some magical solution, or else it would have taken over, as power usage is critical. Let's say you do have an A11 at 4W matching an Intel at 15W (which is dubious at best). Expand on that. Power obviously isn't going to scale at a 1:1 ratio, but if you believe those numbers, an A11 at 40W should be in the ballpark as an Intel at 150W. That is HUGE. There is no way companies would sit there and not take advantage of that.
There is no magic ISA. If it were possible to make a RISC CPU at 35W that performed the same as an 8700k at 91W, it would have been done. To date ARM has focused on low power, low (relative) performance areas. I would love to see a company try to design a "balls to the wall" RISC CPU, but none exist that I know of.