I mentioned ways to share music between multiple computers, so I did an experiment to see how storing music on Amazon's S3, with JungleDisk. I copied a bunch of music on to the disk, and then added the tracks to iTunes. When those tracks were played back, there was a significant delay before playback.
When an file is accessed from S3, with JungleDisk, the whole file is downloaded to a local cache, before the file is returned to the caller. It should be possible to still use Amazon's S3 for this purpose, but the client on the computer would need to not do any caching. It should be possible to pass the file descriptor directly from the http client to the file system.