Ars Technica reports on a company called Cycle Computing that has built a 30k-core cluster on Amazon EC2 cloud for a client. It costs $1279 per hour to run, including fees to Amazon and Cycle Computing.
The cluster ran for about 7 hours, doing a computation for a pharmaceutical company. It had 3809 compute instances, each with 7GB RAM making 30472 cores, 26.7 TB RAM, and 2PB of disk space.
In the past, Amazon has built a virtual cluster that got onto the TOP500 list.
O’Reilly radar has an interesting article by DJ Patil about building teams to tackle data science. He talks about data-centric computing in modern corporations: online retailers, social networking sites, advertising companies. The common thread is that all have tools to analyze and form decisions based on the massive amounts of data that they accumulate from their customers/users.