Ah, so it's a way to sell more machines to run more infrastructure software (also sold) which companies think will increase their scalability, which they don't really need because most of them are never going to have the amount of business that would force them to scale, where simple client-server software would suffice while they're going down the tubes.
Guess you've never been to the other side of that. The other side of that is a set of applications that are good enough to win your company the business, but that don't work together at all.
You can say "you should have thought about that in the first place, good design would have cleared that up" but good, thorough design that attempts to make everything work together flawlessly results in long development cycles and lost business.
Our company won, we built the best products and got the market share, and now we've got a set of applications that our customers expect and want to work together, and we are struggling to deliver it.
SOA is one way we can help with this problem. We can add SOA interfaces to each application, and start constructing the one integrated product our customers want, in an orderly way, quickly, without re-writing all our apps, and piecemeal. We can add SOA interfaces to each application's back end, one at a time, prove it works, and then work on meta-applications to combine the results.
We built much of the software to handle this ourselves. There are OSS options for most pieces of this architecture if we wanted to use an ESB engine (check out Mule for example), and with our VM environment we should not need significant investment in infrastructure. We just need time to build it, and hence corporate wherewithal.
SOA (and ESB and the like) in-and-of themselves will not provide a solution to enterprise integration, any more than the EAI engines of 10 years ago, but at least they provide a common technology to build around so that other developers can tap into the functionality of our applications.