One of the core principles in service oriented computing is that services are always on and available. There are however domains where running services all the time is not suitable, for example when applying simulations workflows in the eScience domain. The simulation services orchestrated by these workflows are typically used only rarely and irregularly, keeping them running all the time would result in a significant waste of resources. As a consequence, we developed the approach of on-demand provisioning of workflow middleware and services. In this paper we will give an overview about our work. We will present the motivation and main idea of our solution approach and will also provide details about some of the results of our work. The overview about our previous and current work is then complemented by a detailed discussion and comparison of the roles involved in both concepts, traditional service oriented computing as well as our newly developed on-demand provisioning approach.