A method (and corresponding system and computer program product) for efficiently routing a multi-person conference to reduce bandwidth usage and latency is described. Computing devices used by the participants to participate in the conference are separated into proximity pools based on direct accessibility. For each proximity pool, a computing device is selected as the key node. Only the key node communicates directly with a server, and redistributes data received from the server to its pool peers. Communications between computing devices within a same pool are routed directed to one another. Proximity pools can be refined to factor in distances among the computing devices. Multiple servers may be used to reduce latency. Proximity pools are assigned to nearby servers. Communications between computing devices in different proximity pools are routed through the server(s) associated with the different proximity pool.