Reduced power comes from reduced complexity, and that has a performance cost, yes -- but it takes far more than a 10% increase in power to get a 10% increase in performance.

Intel's designs, being tuned for high sequential performance, are significantly better for that purpose; but if you look at VPS services which currently only get a fraction of a CPU each, I'm sure they could run more power-efficiently with a larger fraction of a punier CPU.