IEEE had a good article a few months ago about the impending slowdown in the growth of supercomputer power. The reasoning is sound, and boils down to a simple idea. The energy required per floating point operation is not dropping exponentially. Therefore the energy required to support an exponential growth curve is growing exponentially. In other words, an unsustainable amount of energy will be needed to support the current growth curve.
A floating point operation currently requires about 70 picojoules of energy. This number is not anticipated to drop using currently available CMOS technology. Thus. the total energy needed to sustain the supercomputer growth curve will soon match the output of an entire nuclear power plant.
We certainly could build a supercomputer with its own dedicated nuclear power plant. My guess is that we won't. So as long as we use CMOS transistors for computation there is a maximum supercomputer we will build. And that will happen in roughly 10 years.
A word repeated several times in this analysis is "current". Current CMOS technology operating with current paradigms and our current data processing technology has a power limit which will end the supercomputer growth curve.
But as Ray Kurzweil has said over and over again, computing hardware has changed many times in the past 2 centuries and it will probably change again in our lifetime. Graphene quantum molecular switches coupled by plasmon interconnects to photonic modulators could offer a dramatic reduction in the power per operation. Yesterday this was science fiction. Today it is cutting edge research. In 40 years it will probably be embedded in our wrists.