Or even just having your provider distribute the load of a number of servers. The same FastCGI handlers will startup on many actual machines, providing a measure of scaling.