Video on demand (VOD) is becoming one of the most important services for many network operators. VOD needs a huge bandwidth to deliver data in a real-time network. Ethernet passive optical networks (EPONs) provide a low-cost and high-speed solution for multimedia services such as VOD. In this paper, we propose a scalable architecture based on an EPON, using single copy broadcast in the downstream direction and a special caching mechanism responsible for VOD requests, to improve VOD services. Our architecture consists of an optical network unit (ONU) with an associated storage unit and optical line terminal (OLT) components. We set a “broadcast” logical link identifier in the video packets so that all of the ONUs can accept them. Therefore, when the movie is sent from the OLT to the ONUs after it is requested by the user, each ONU accepts the movie packet and saves it in its own storage. In this architecture, the OLT just needs to send each video to the ONUs once, and the attached storage is responsible for future requests. With this architecture, at least 60% of VOD traffic can be saved in the feeder fiber. Simulation results show that our architecture can improve the overall quality of service in terms of packet loss, mean packet delays, jitter, and system throughput.