When network effects are present, the owners of a “platform” can generate significant rents: think of DVDs, computer operating systems, or various protocols like TCP/IP. Once a platform becomes dominant, the owner (generally but not always the firm controlling some crucial IP) can extract profit from firms that produce for the platform, like movie studios or computer software companies. Because network externalities are present, the firms pay these rents rather than defect to alternative platforms.
Nonetheless, we do see platforms switch on occasion; famously, Internet Explorer replaced Netscape, a story discussed in a well-cited paper by Tim Bresnahan. What, empirically, might cause such a switch? Cantillon and Yin, in a paper that the grapevine suggests is on its way to acceptance at AER, consider an example from finance. In the 1990s, trading in German treasury futures migrated almost entirely from a London based exchange to one based in Frankfurt. The network externality in this case is liquidity: the more firms on an exchange, the more liquid an investment on that exchange is. The authors show that German government pressure on firms to switch, and easier remote access of an electronic exchange in Frankfurt versus an open outcry floor in London, were relatively unimportant. What was important is that the firms on both exchanges showed relatively little overlap during the years in which the switch took place. Firms that valued liquidity less, and valued complementary products sold on the Frankfurt exchange more, were more likely to trade on the upstart. This suggests that, in general, product tie-in and complementarity can overcome the devastating combination of first-mover advantage and network effects.
One last note: look at the regression output reported in this paper. I admit to my econometric incompetence, but nonetheless I am skeptical when I see panel data regression output with many variables significant to 5, 6 or more standard errors. What is going on is that, month-over-month, the overwhelming majority of firms do not switch exchanges. So the data is very good at predicting that firms will do exactly what they did the month before. Surely there are better ways to estimate sparse discrete choice in panel data? What I mean is, why isn’t the dependent variable here “switched exchanges” rather than “is on exchange X”? Wouldn’t that make the results easier to understand?
http://www.ecares.org/ecare/personal/cantillon/web/bund.pdf (2009 working paper version)