Maybe - but then again - maybe not.
http://www.theregister.co.uk/Design/graphics/icons/comment/stop_32.png
All predictions in HPC hardware develpment point towards massive concurrency in 2-5 years.
Wether this comes in the shape of GPGPU, 20 millions embedded processor cores in an n-dimensional torus, FPGA or any other devlopment remains to be seen.
In any case it will be a development which predictably will kill of 'HPC in the COTS cloud'.
Massive concurrency - possibly with an OS which runs a single thread per node -- is exactly what the Amazon SW/HW stack is not good at.
This holds true for even an ordinary off-the-mill ISV code. Running the same code on 1000 embedded processors instead of 100 x86 cores will save you a factor of ~10 in terms of power.
In 5 years time this will sum up to a factor of 5 or more in total cost savings.
So the expensive bits in 5 years time will be power, power, software development to adress the power issue (massive concurrency), and - power. Not many incentives for HPC/Cloud.