If people are avatars, then the amount of information exchanged in a conversation between them is the streaming audio and vectors describing the position of the avatar's bodies; I feel like this would be less bandwidth than is required for video chat, although the expectation of low latency will likely be greater due to the immersion.