I'm not sure I'd be too worried about bandwidth issues - once content is replicated it could be served up from people who already have it rather than always going back to the original source. In fact, I suspect things would have to be done this way to provide some kind of resilience.

I think getting the security model for distributed content would be the crux of this kind of architecture, and I suspect that is probably solvable and would really benefit from an open architecture with multiple implementations.