[jononor]

We observe larger and more power hungry at the high end, because those models are selected for maximum performance - at the expense of everything else.

But GPU power efficiency (gflops/watt) has been increasing at an exponential pace as well - over the last decades it has been consistent doubling every 2-3 years. So over 10 years we have had 10-30x improvements in power efficiency. The same is true for cost efficiency (gflops/dollar).