The US Departement of Energy Magellan project, a pilot project for using cloud computing for science, has issued its final report:http://science.energy.gov/~/media/ascr/pdf/program-documents/docs/Magellan_Final_Report.pdf
As a cloud computing cynic (or, more generally, a Latest Big Buzzword cynic), it's hard not to feel a little vindicated. It turns out cloud computing, while having some advantages for some workflows, is not magic, and current cloud computing stacks are designed for random web-based stuff, not scientific computing. A not entirely unfair summary of the high points:
* Current big research computing centres already often have the economy-of-scale advantages of cloud computing, and in fact are generally cheaper (and come with real scientist support).
* Current public cloud offerings work fine as long as performance doesn't matter.
* Being able to roll your own image can be helpful for some scientists: but then, either
(a) you're offloading a lot of sysadmin duties off to the scientists, who now have to know enough about sysadmining their OS of choice to be able to roll a correct and complete OS image from scratch just to run the couple of programs they want to run; and/or
(b) you encourage sharing of images, which are then
i. "kitchen sink" huge images which contain tonnes of stuff unnecessary for any particular job, and
ii. prone to different types of security issues. ("Use my new compute image for quantum bogodynamics! I promise it hardly ever emails copies of all your private data to me!")