[foresto]

Also, STUN/TURN is not sufficient to handle scalable voice and video conferencing. It would take care of NAT traversal, but can't do anything about the bandwidth problem: Each participant would have to maintain a bidirectional A/V channel to each other participant. Many residential internet connections would struggle quickly, and mobile phone bandwidth even more quickly.

Adding a selective forwarding unit could solve the outbound bandwidth problem (each participant would then only need one outgoing A/V channel) but can't help with the incoming bandwidth problem (each participant would still need to receive the A/V stream of each other participant). It can't compete with the Matrix design as I understand it.

A sufficiently powerful multipoint control unit could solve the bandwidth problem in both directions. But as far as I know, MCUs capable of handling big conferences are not cheap.

In other words, I think pmlnr's XMPP endorsement underestimates what it takes to do scalable audio/video chat.