SFU, MCU or mesh: how group video calls actually route media
Mesh costs grow per peer, MCUs mix, SFUs forward. The three group-video topologies from RFC 7667, what each costs in bandwidth and servers, and who can see your media - from the IETF specs.
One-to-one video calls can run peer-to-peer, as our STUN and TURN guide explains. Group calls strain that design: every added participant costs every sender another upload. This guide explains the three ways group video routes media - mesh, MCU and SFU - what each costs, and who can see your media in each design, grounded in the IETF specifications that define them.
Mesh: everyone sends to everyone
The simplest design has no server in the media path at all: RFC 7667 calls this topology Topo-Mesh. Every participant sends their audio and video directly to every other participant. There is no fixed headcount where it breaks: the cost grows per peer, and where it bites depends on bitrate, devices and each sender's uplink. The arithmetic is unforgiving though - a five-person call already means each participant uploads four copies of their stream, and encode/decode CPU grows the same way. Upload links are usually the first constraint, which is why real group-calling systems move media through a server.
MCU: a server that mixes
The traditional answer is the Multipoint Control Unit. In RFC 7667's terms, an MCU receives everyone's streams and sends back a single combined stream: a mixer decodes all video, composites it into one layout, re-encodes it, and each participant downloads just that one stream. Upload and download per participant stay flat no matter how large the call gets. The price is server-side transcoding - CPU-heavy and adding latency - plus a fixed layout: everyone sees the same composed picture. RFC 7667 documents this family (Topo-Mixer and the simpler video-switching MCU, which forwards one selected stream instead of compositing).
SFU: a server that forwards
The design modern video APIs converge on is the Selective Forwarding Unit. An SFU receives each participant's stream once and forwards the streams each receiver actually needs - full resolution for the speaker on screen, thumbnails or nothing for the rest, simulcast or scalable-codec layers matched to each receiver's bandwidth. RFC 9605's introduction describes exactly this role: SFUs "efficiently route media streams to call endpoints based on factors such as available bandwidth, desired video size, codec support". Each participant uploads once (plus simulcast layers) and downloads only what they watch, and the SFU never decodes video, so it scales far cheaper than an MCU. The trade-off: receiver download still grows with visible participants, and layout flexibility moves to the client.
Who can see your media?
WebRTC requires media encryption, but "encrypted" needs a second question: encrypted to whom?
- Mesh (peer-to-peer): RFC 8827 requires DTLS-SRTP between the two browsers, so media is encrypted end to end by construction.
- MCU or SFU: the DTLS-SRTP association terminates at the server. RFC 9605 names the two layers plainly: hop-by-hop encryption between each endpoint and the SFU (the SRTP WebRTC mandates), and optional end-to-end encryption of media between the endpoints themselves. Without that second layer, the media server - and whoever operates it - can access call content.
- End-to-end encrypted group calls add that second layer (SFrame, RFC 9605, is the IETF mechanism): the SFU still sees the routing metadata it needs, but not the media.
So when a provider says calls are "encrypted", the meaningful question is whether group media is encrypted only hop-by-hop to their SFU or end to end between participants.
Choosing a topology for your product
1:1 calls: mesh is right - no media server at all. Small group calls: mesh can work while per-peer uploads stay within the senders' links - the practical ceiling depends on bitrate and devices, not a fixed headcount. Real group calling: SFU is the default answer in 2026, whether self-hosted open source or a managed video API (our build-vs-buy guide covers that choice). MCU mixing survives where a single composite stream is the point - SIP/PSTN dial-in, recording a fixed layout, or feeding legacy systems.
FAQ
Why does mesh struggle as calls grow?
Upload bandwidth and CPU. Each participant sends a full copy of their stream to every other participant (RFC 7667's Topo-Mesh), so N people means N-1 uploads per person. How soon that hurts depends on bitrate, device and uplink; upload capacity is usually the first wall, not download.
Is an SFU the same as a TURN server?
No. A TURN server (RFC 8656) blindly relays one call's packets between two peers when a direct path fails; it forwards encrypted traffic without keys. An SFU is an application-layer router that terminates each participant's WebRTC session and selectively forwards many participants' streams - and unless the app adds end-to-end encryption, the SFU can access the media.
Can a group call be end-to-end encrypted?
Yes, with a second encryption layer between the endpoints on top of the hop-by-hop SRTP - SFrame (RFC 9605) is the IETF mechanism designed for exactly this, leaving the SFU only the metadata it needs to route. Whether a given product actually does this is a vendor claim worth verifying.
Do I need an SFU for a 4-person call?
Not strictly: 4-person mesh means 3 uploads per sender, which can be fine at modest bitrates on good uplinks. There is no universal headcount threshold - the trigger is when per-peer upload and CPU growth stops fitting your users' devices and networks. Plan the SFU before quality forces the issue.
Sources
- RFC 7667, RTP Topologies (mesh, mixer/MCU, video-switching MCU): https://www.rfc-editor.org/rfc/rfc7667
- RFC 9605, Secure Frame (SFrame) - SFU routing role and the two-layer encryption model: https://www.rfc-editor.org/rfc/rfc9605
- RFC 8827, WebRTC Security Architecture (mandatory SRTP/DTLS-SRTP): https://www.rfc-editor.org/rfc/rfc8827
- RFC 8656, TURN (relay behavior, contrasted above): https://www.rfc-editor.org/rfc/rfc8656
Drafted with AI assistance from these sources.