In a mesh topology, every participant sends their stream to every other participant — N participants means each one uploads N-1 copies of their video, which collapses beyond 4–5 people. An SFU fixes this: each participant uploads one stream to the SFU, and the SFU forwards copies to everyone else. From the participant's perspective the upload is constant; the SFU absorbs the fan-out. For 50-person conferences, SFU is mandatory; for two-person calls, mesh is fine.
The SFU also makes simulcast and SVC useful. Each participant uploads multiple resolution layers (simulcast) or one scalable layer set (SVC). The SFU forwards the appropriate layer to each receiver based on receiver bandwidth — a viewer on a fast connection gets 1080p, a viewer on cellular gets 360p, from the same upload. Without an SFU, you'd need to encode each viewer-specific stream at the source.
Open-source SFUs in 2026 include mediasoup, LiveKit, Janus, Jitsi Videobridge, Pion (a Go library that can build an SFU), and OpenVidu. Commercial managed SFU services include LiveKit Cloud, Twilio Video, Daily.co, Vonage Video API (formerly TokBox) and Cloudflare Calls. The choice usually depends on language stack (mediasoup for Node.js, Janus for C, LiveKit for Go), scale needs and DRM/recording requirements.
