Below are the comparitave results of some test we ran on various instances of AWS clusters. We also ran identical jobs on our 3 different series. One should keep in mind this these tests were primarily CPU based and there was very little disk IO and most memory operations would have occured in cache. The ICTS 2800MHz cores with fast cache outperform Amazon's best offering by far (a lower value on the graph is better).
We ran an instance of Amazon's fastests offering, the m2.4xlarge, as a 2 node cluster with a total of 16 cores. We bid at $0.14 per hour and the longest job took just over 1/2 an hour, hence we paid $0.14 total. Had the job run for 100 hours this would have cost us R224 excluding data transfer costs. 50 similar runs would have cost just over R11,000. This would not be an unusual number of runs for a researcher to complete in a week.