Sean McGrath has posted a nice set of slides (ppt) that he used for his presentation on XML pipelining at XML 2004. Sean is more or less the pipe guy of the XML community, which is an honorable position. Most interesting is his slide (nr 4) on API components versus data-flow components. Unfortunately, I don't know exactly what he said about this slide (I wasn't there), but he must have raised some interesting points.
The last few decades have made clear that data-flow components never go out of fashion. In particular, the only really successful component model is the idea of pipes and filters of Unix. Unix components can easily be composed; they can be implemented in any language and they are suitable for anticipated (e.g. shell scripts) as well ad-hoc (the command line) composition. API-based component models tend to be language, or at least language family, specific. Hence, components relying on such a component model are coupled to a language or platform. As a result, they go out of fashion. This doesn't make API-based component models useless, but components based on such a mechanism just don't last. Over time, components will have to survive in a heterogenous environment, since fashion changes while time passes by. In this way, the Internet is comparable to time. If you want to connect components in an heterogenous environment, then language specific solutions just don't work. Your current environment is the Internet, which is quite heterogenous.