Andrew Turner's act of data liberation reminded me that I'd made a similar point at THATCamp. Web applications are often coupled directly to a database as shown on the left of the diagram below, and other applications on the Web that can't access the database must scrape data from the primary app (illustrated by a dashed scrape-scrape-scape line). A better architectural design pattern is shown on the right of the diagram: use Atom or KML (especially for geographic apps) as a general purpose service layer to which many apps (including the cool apps of the future 2 and 3) can connect. The New York Public Library is one institution using this design, and my sense from attending the mashup session at THATCamp is that there will soon be others.
This is not entirely new to GIS architects. If you subscribe exclusively to the OGC service architecture, you would of course use a WFS instead, but Atom has the advantages of being more generic and more attuned to the architecture of the Web itself.