IP-layer multicast (IP multicast) is a sophisticated approach for delivering data to multiple receiver hosts distributed in networks, but several issues have prevented its wide deployment. To solve the issues, several application-layer multicasts (ALM) have been proposed as architectural alternatives to IP multicast. ALMs depend on the reliability of the end hosts, but end hosts are less stable than network devices, such as switches and routers, which IP multicast depends on, because their computing resources are shared by other, unpredictable, processes. The article presents a new ALM protocol on an overlay of such undependable end hosts. It exploits a group hierarchy for scalable data delivery, for example, to thousands of receivers, as well as conventional ALMs. No leader hosts are fixed in the hierarchy, but they are dynamically and randomly assigned for each forwarding operation at the end hosts. Thus, the multicast trees change very frequently. This distributes and balances the stress in the delivery networks, avoiding congestion. We analyse the link stress and show that it is up to double the data bit rate when the receivers are equally distributed in the group hierarchy. This also makes the system stable and robust, not only to dynamic membership changes when receivers join and leave, but also to temporal performance changes in the receiver hosts, which would cause errors such as a packet loss when they fail to receive or send some packets in time. We also discuss scalability issues and show the results of demonstration experiments.