Again, efficiency does not matter.
Look at it this way:
Assume every block is worth $50,000.
The amount of power that can be consumed to produce that block is $50,000 minus some overhead number (real estate, cost of hardware, labor, profit, etc.) The higher the price the more power can be consumed. Power consumption is directly proportional to price regardless of efficiency.
The more efficiently the power is used the higher the hash rate and difficulty of the network can and will be. Efficiency drives difficulty and hash rate, price drives power consumption.
Sure some miners are better situated (cheaper power, cheaper cooling, better hardware, etc.) so they are more profitable than other miners. This extra profit allows them to afford to buy even more hardware - pushing out the less efficient miners who will be losing money. As new hardware comes on line and old hardware is retired the power consumption remains proportional to price.
Also, I am looking at the overall averages for the entire network and looking at estimating the power consumption of the entire network. Local variations do not matter when looking at overall averages.
"Allows them to buy more hardware" and "necessarily leads them to spend all new profits into recycling and building more efficient hardware in a never-ending always forward moving cycle" are two different things. I agree that what you're proposing is a possible scenario, but not the most likely.
We've seen this same scenario play out with the move from dedicated servers to multi-core virtualized cloud servers - the question people were asking with datacenters in 2007 was "how the hell are we gonna get enough power for this?!" and now in 2014 that question seems absurd.
Purchasing, deploying, running, and managing the cap-ex and op-ex of all this new hardware at scale is *costly*. It's not frictionless in terms of time and effort. Having the money to buy new hardware and being able to do so are different things at scale.