A picojoule sounds like such a trivial amount of energy.  But when you multiply it by a petaflop you realize that the Singularity will require huge amounts of electric power.

Intel, IBM, and other computing companies have shown no ability to create a "Moore's Law" for power.  The power required to switch a transistor is not dropping geometrically.  And since the number of transistors per CPU is rising geometrically that means the power required to drive future CPU's is growing exponentially.  This graph of the Top500 computers implies no end to Moore's Law.  But if the y axis were replaced by $ we would see that at some point in about 20 years the entire gross national product of the USA will be required to run a single computing platform.  (Note that the fastest computer in the world has been a 33 petaflop system for 2 consecutive years.  From 2002-2005 the fasted computer stalled at about 50 teraflops before resuming the exponential growth curve.  So there is precedence for a pause in the growth.)

In 2019 this projection indicates the fastest computer in the world will be a 1 exaflop computer, 10^18 flops/second.  At 10 picojoules per flop (an extremely optimistic number) that would require 10 million joules per second.....10 MegaWatts.

Ray Kurzweil has a high opinion of the scientists working to solve the picojoule problem.  But his laudable goals are not reflected in any current state of the art systems and there is no evidence that we will ever see mass production of an exascale computing platform with femtojoule transistors.  Will we ever see a single exascale computer?   Probably, we'll see one.  And it will cost a billion $/month to supply it with the gigawatts of power it needs to operate.  Who is going to pay those billions of $? 

Realistic economic considerations drive one and only one conclusion.  The Singularity will not happen because the picojoule roadblock will not be overcome.