Matei Ripeanu has an interesting brief article in IEEE Distributed Systems Online in which he analyzes the shape of the by-now-(in)famous Top 500 list of supercomputers, released every six months since 1993.
He notes first that a plot of performance vs. rank gives a power law. Not in itself surprising. But then he notes that the power law coefficient is getting smaller over time: in other words, the bigger machines are, on average, getting faster more slowly than the slower machines. Thus, for example, the bottom 25 machines in the Top500, if aggregated together, would match only the #30 machine in 1993, but match the #5 machine in 2005.
Why this change? Alex Szalay attributes it to the top500 spurring people to buy bigger computers. (I.e., the act of measuring supercomputer evolution perturbs that evolution!) A provocative thesis, but hard to evaluate. Matei attributes it simply to the increasing ease with which one can aggregate systems.