Jens,
you make some very relevant points.
Also let me add choice of precision of the mathematics in the problem.
If I am not wrong, an arbitrary limit on 30MW has been proposed for an Exascale machine.
By tthat I mean that at the moment an exascale machine is perfectly possible - but it would just consume a huge amount of electrical power.
I read a very good paper a few years ago which discussed the choice of arithmetic precision for problems: "picowatts per flop"
I believe software authors will have to pay more attention to the coice of presicion (or integer) computation.
For instance we see that Machine Learnign workloads on GPUs quite happily use single precision.
I have lost the reference to this paper (sorry)
On anther note, I believe that Formula 1 teams were consideting a limit of the power budget for CFD simulations.
Of course there is much dancign on pins going on there - do you consider the power to the rack (which you can measure easilty at the PDU)
But how about the power used for cooling? I guess you have to consider your entire data centre as an adiabatic box and meter the overall power in.
checking the 2017 FIA regulations though there is still a FLOPS cap.