Power is not always the only cost. I work in an industry where you pay per core for software license fees far in excess of the cost of power. There, perf/core is what matters.
I would imagine the same is true for e.g. server response latency when that latency correlates to user experience and thus growth/popularity of your service. People might not notice a few ms, but 100ms vs 300ms is definitely visible.
I think the real value is new things like TSX/HLE.
(And if you say "thin client, distcc, big build farm", then you just need the latest CPUs in the build farm instead.)
Of course this is as much an argument for a real module system in C/C++ as anything else...
(And, I completely agree that TSX and other architectural stuff is totally awesome, too. Perhaps that'll drive adoption more than incremental perf, but I'm just arguing perf is still often really important.)
Example: I recently built a Linux video transcoding machine to reencode H.264 BluRay movies to smaller H.264 files [1]. A $110 AMD FX-8300 is able to reencode an average BluRay movie around 47-50 fps, while you would need to spend more than twice that amount of money to match this performance with Intel (the performance of the FX-8300 at this tasks falls between a $200 Intel i5-4590 and a $250 Xeon E3-1231 v3).
[1] Using: avconv -i input.mkv -threads auto -s hd720 -c:v libx264 -c:a libmp3lame -sn -b:v 1400k -b:a 128k -ac 2 output.mkv