There is disclosed a system and a method of real time stream provisioning. The method includes dynamically allocating one of a plurality of servers in a server cluster for use in a chat, the chat including an audio component for each of n chat participants and encoding the audio component for each of the n chat participants, using at least one graphical processing unit, into n audio streams, where the n audio streams each include audio components for n−1 of the n chat participants where each of the n audio streams does not include the audio component for a different one of the n chat participants. The method further comprises transmitting one of the n audio streams to the n chat participants such that a chat participant of each of the n chat participants receives an audio stream excluding the audio component associated with the chat participant.