I really want to switch away from Intel across the board but we are still far from there. And in fact you're not really disagreeing with my point: I said Intel & Nvidia are leading for raw power, but when you are looking at performance/watt, they are not as far ahead or even (depending) trail disastrously.

Intel doesn't seem to have it in their DNA to really focus on mobile across the board. They excel on process (Si node), no question. They have the highest performance CPUs with lots of superscalar, speculative, and multicore performance, and a lot of work going into internal bus architecture. If you don't care about power consumption they seem to be pretty much the only game in town.

And even on laptops, they have the rest of the infrastructure (advanced IO controllers etc) where ARM still noticeably lags. The numbers you cited were mobile chipsets vs the A9. Intel more than held its own, and these didn't get into multicore performance where most of the ARMs' performance rolls off dramatically. Yes, even Apple's. Those Intel mobile chips, much as I hate them, are more general case (see below).

And of course these days most of us care about power (not enough in the datacenter, but that day is coming soon as well). On that front ARM is great. But as I noticed above they don't get the multicore advantage Intel does. Now that doesn't matter as much on a tablet since you actually want to spend as little power energizing silicon as possible. Look at the asymmetry of the A9: a huge amount of the area goes to servicing graphics (kind of the weird descendent of the Alto in this way), less on dealing with multicore bus contention. Which is what you want on a tablet! But it's less useful when, say, you're running a CAD app.

And other design houses have found other points in the space: Q'comm is the only other performance leader in ARM (mediatek etc are going for lower power/fast design) and qcomm parts are notoriously power hungry too (oops). Apple is the only one who has high performance and good power consumption -- but for an execution mix that they control closely, not for the general case!

Interestingly NVIDIA suffers from the same thing, but has managed to cleverly switch into a parallel track (ML with GPGPUs) which is currently also power consumption insensitive and can pay not just a higher W for performance but $ too.

I can't wait for the day to arrive, but no, I don't believe that day is yet nigh.

PS: this post reflects some work I did specifically in this area, unfortunately proprietary to the organization who paid for it.