My copy of ;Login: arrived yesterday last week. As I was reading it over a quick breakfast of oatmeal and coffee the article "Benchmarking Amazon EC2 for High-Performance Scientific Computing (PDF)" by Edward Walker caught my eye. Having worked with our own grid services at Network.com I understand that there is a tradeoff in terms of cost/value and performance and didn't find the gist of the results to be surprising.
In my opinion, a purpose built high bandwidth, low latency grid/cluster is generally going to outperform the lower cost more generic and I believe possibly more flexible solution. On the other hand, in the utility model the consumer of the service/resources doesn't really have to worry about the TCO of the service or the cost of management and up front implementation.
In short: if you have the time, money, talent and desire feel free to build your own. If not it may be worth while to trade off some of that performance for convenience.
