fal reports faster-than-playback generation under its reported conditions. That is not a guarantee of an uninterrupted stream: queue variation, safety checks and delivery still need buffering and failure handling.
A shared generated scene can serve many viewers without another model call per viewer. Delivery bandwidth and CDN cost still grow with audience size; generation is only one part of the bill.
The strongest live channels are not "anything-goes" streams: they wrap infinite generation in a fixed format (a show, a classroom, a game) so the audience keeps coming back for the characters and the story, not just the novelty.