DRAM prices (in current milli-cents i.e., 10^-5$ per bit) for each successive DRAM generation and for all DRAMs sold taken together. Instead of time, cumulative sales (in bits) are shown on the x-axis. Time can be inferred by the connected data points, each representing a discrete year (see the year labels for the 64M  DRAM history between 1994 and 1999). Costs for each DRAM generation follow a  typical life-cycle pattern: very high at the moment of market introduction, followed by a rapid decline as production experience accumulates, finally hitting a "floor price", after which prices may even increase again. Typically, DRAM floor prices are US $1 per unit DRAM  (irrespective of memory size); hence floor prices per bit are lower for each successive DRAM generation. Price declines follow a classic learning and experience curve pattern, driven in this case by progressive elimination of defects during production and improvements in lithography. Like all experience curves in manufacturing, progress down the learning curve is not indefinite. Thus, to sustain further price decreases radical technological change (or what  we have termed "learning via R&D" elsewhere) is required. Interestingly, combining experience gained in manufacturing and pushing the technological frontier via innovation, results in an aggregate linear (double-logarithmic) trend where price per bit drops by some 30 percent for each doubling of cumulative bits sold. The result is a single curve driven by two entirely different and distinct phenomena. "Market pull" is hence an entirely inappropriate single  metaphor to describe the history of DRAMs (and as we will see later on for  other technologies as well). The figure updates and confirms earlier analysis by Pete  Klenow and Doug Irwin (Journal of Political Economy, 102(6):1200 - 1227, 1994)  that found significant learning effects for each successive DRAM generation, but only limited spillover effects between individual generations or between  different countries and manufacturers.