Charles put together a great post related to the benefits of getting to an SOA based architecture. It reflects some of the pro's I outlined in Point, Counterpoint - but adds a lot more. However, I think one major topic is still missing from the debate landscape - the compexity of managing an N to the Nth grid of small components.
Ironically, the same "network effect" argument that Charles uses to show the correlation between a SOA's value and the commitment you have made to it, is essentially the same argument I would use to show the correlation between SOA maintenance cost and brittleness.
I am reminded of an experiment that was run in the final days of Exodus Communications (at the time, the world's largest hosting company - to whom I sold my first company). Interested in getting into the Application Hosting business to complete with the likes of LoudCloud at the time, Exodus modeled what it would take to make the business work. What we found was the only way to make Application Hosting viable was to limit the number of versions of each ISV in any layer of the stack that we would support. You can use BEA, but only versions x.y.z and a.b.c. You can run Oracle, but only 220.127.116.11.1 and 18.104.22.168.2. See, the problem is, each ISV acted as a lously coupled part of a complete system - but the testing matrix across an N-component system was N to the Nth. There is no way that all permutations can be tested. In fact, LoudCloud later split their business, sold the application hosting stuff to EDS and left OpsWare, a software company focused on trying to make software to manage this nasty deployment problem.
SOA is no different - every new component has to effectively work with each client that calls it. Want to change a version of a component, do all of its client support this new version? You see, the network effect of dependency is as strong as the network effect of value. And this, I am predicting will be the major wrench in SOAs cogs.
More on this later. I have been scanning for early signs of this discussion and I thank Charles for at least giving me the context to start the debate.