As the social networking site Twitter has risen like a hot air balloon in a gale force updraft it's become completely snarled in its own guy wires and functions are getting dropped over the side like so much dangerous ballast (archives, IM access, tweets "with others", etc.)
JonW sent me a link to this article which discusses Twitter's scalability issues. (Meaning how it will it grow.)
I have no opinion on whether relational databases or cloud computing are the answer...
...but Robert O'Brien's comment seems likely:
Your assertion that Twitter can't be decentralised conflates the notion of discovery and service. Discovery can be centralised while the service provision can be decentralised. DNS is a perfect example of a centralised namespace with a decentralised service that maps names to ip addresses. Bit Torrent is another. And to over simplify, what you describe in terms of an architecture is typical of many P2P systems - it is just P2P inside the wall.
In a perfect world there is no technical reason why, for example, a twitter type discovery and friend management service couldn't map all message delivery on to a network of decentralised jabber end-points.
The scaling problem comes because Twitter is trying to "capture" all our tweets, control the UX and API, control the namespace to insert themselves as a centralised message utility in order to extract value. i.e. it is the constrains of the business model that makes Twitter have a bundled service and therefore means Twitter (the business/service) can't be decentralised.
Will Twitter "succeed" so that all our "tweets" are owned by Rupert Murdoch or Bill Gates or will the balloon pop in midair? Pass the cotton candy.