Choosing between WebRTC mesh, an SFU, and an MCU affects call quality, infrastructure cost, recording, moderation, and the amount of application logic your team must maintain. This guide compares the three group-call architectures, provides a repeatable capacity-planning worksheet, and outlines a proof-of-concept process you can rerun as room sizes, codecs, SDKs, and infrastructure assumptions change.
Overview
A WebRTC group call has three broad architectural options:
- Mesh: Each participant sends media directly to every other participant.
- SFU, or selective forwarding unit: Each participant sends media to a server, which selectively forwards chosen streams to other participants without normally mixing them.
- MCU, or multipoint control unit: A server receives media, decodes and composes it into one or more layouts, then sends composed output to participants.
The right choice depends less on a universal participant limit than on your interaction model, device mix, network conditions, moderation requirements, recording design, and operating budget. A two-person support call, a six-person creator room, and a broadcast-style panel may all use WebRTC while needing different media paths.
Mesh
Participant A <----> Participant B
\ /
\ /
Participant C
SFU
A ----\
B -----> SFU ----> selected streams to A, B, C
C ----/
MCU
A ----\
B -----> MCU ----> composed layout ----> A, B, C
C ----/Mesh is simple to prototype, but every sender must upload a separate copy of its media to each peer. As the room grows, upload bandwidth and connection management can become the limiting factors, particularly on mobile networks. An SFU centralizes routing while preserving separate participant streams, making it a common starting point for interactive group calls. An MCU centralizes composition and can simplify a fixed layout or single-output workflow, but server-side media processing generally increases compute and operational complexity.
| Concern | Mesh | SFU | MCU |
|---|---|---|---|
| Practical participant range | Small rooms; validate with real devices | Small to larger interactive rooms, subject to server and client capacity | Useful when a composed output or controlled layout is central |
| Bandwidth usage | High sender upload as peers increase | Usually one primary upload per sender; downloads depend on subscriptions | Usually one composed download, with server-side media processing |
| Server cost | Lowest media-server requirement, but clients carry the load | Routing, egress, signaling, monitoring, and recording costs | Routing plus decode, composition, encode, and often recording costs |
| Recording support | Requires a separate capture or recording design | Flexible: record tracks, subscriptions, or composites | Natural fit for a composed recording, subject to processing capacity |
| Simulcast | Limited value because peers connect directly | Strong fit: the SFU can forward an appropriate layer | Possible, but the composition pipeline must handle input layers and output encoding |
| Layout control | Mostly client-side | Usually client-side, with server-assisted selection | Centralized and consistent |
| Moderation | More distributed and difficult to enforce consistently | Server-side permissions and subscription controls are practical | Centralized control is strong, but media policy must be implemented carefully |
| Mobile performance | Can degrade as peer count and uploads increase | Often easier to control through subscriptions and layers | Can reduce client composition work, while adding server processing |
| Implementation complexity | Low initial complexity; scaling and edge cases arrive early | Moderate; requires signaling, routing, subscriptions, and operations | High; adds media composition and processing decisions |
These are architectural tendencies, not guarantees. Browser behavior, SDK implementation, codec support, simulcast behavior, network paths, and device thermals can change the result. Treat the table as a decision aid and confirm it with a controlled test.
How to estimate
Start with the room model rather than a provider’s headline capacity. Define how many people publish video, how many publish audio only, how many watch, and whether each viewer needs every active stream.
1. Define media roles
Record the expected number of publishers, viewers, moderators, and automated participants. A room with 20 attendees and four active speakers has a different media profile from a room where all 20 cameras are continuously published.
2. Estimate media relationships
For mesh, a participant sending one encoded stream to n − 1 peers has approximately n − 1 outgoing media relationships. Across the room, the number of directed sender-to-peer relationships is approximately n × (n − 1). This is a planning model, not a bandwidth guarantee, because adaptive quality and audio-only participants change the actual load.
For an SFU, begin with one publisher upload per published track and calculate downloads from subscription policy. If a client subscribes to every active video, its receive load rises with active publishers. If the application shows only a stage, grid subset, or dominant speaker, the SFU can reduce that load.
For an MCU, estimate input tracks, output layouts, target resolutions, frame rates, recording variants, and concurrent compositions. The important variables are decode count, composition count, encode count, and output duration—not only participant count.
3. Convert the model into cost inputs
Use provider-specific rates or internal infrastructure estimates for signaling, media relay, bandwidth or egress, recording, storage, transcoding, observability, and TURN usage. Keep the calculation modular:
Monthly media cost = relay or processing cost
+ egress cost
+ recording and storage cost
+ signaling and ancillary service cost
+ monitoring and support allowanceDo not insert generic prices into this formula. Prices vary by vendor, region, commitment, codec, recording mode, and billing unit. Instead, place current quotes in a worksheet and preserve the date and assumptions beside each value.
Inputs and assumptions
Use the following worksheet before selecting a WebRTC platform or building a media service:
| Input | Your value | Why it matters |
|---|---|---|
| Peak concurrent rooms | ___ | Determines simultaneous signaling and media allocation |
| Peak participants per room | ___ | Sets the room-size test points |
| Active video publishers | ___ | Drives encoding, subscriptions, and composition |
| Audio-only participants | ___ | Changes bandwidth without adding equivalent video load |
| Expected watch-only viewers | ___ | May require a broadcast or fan-out path |
| Target video layers | ___ | Influences simulcast and client download selection |
| Recording mode | None / tracks / composite | Changes processing, storage, and privacy requirements |
| Call duration | ___ minutes | Supports monthly usage and storage estimates |
| Device mix | Desktop / mobile / embedded | Reveals CPU, thermal, and network constraints |
| Regions and network types | ___ | Helps test latency, relay paths, and regional placement |
| Quality targets | ___ | Defines acceptable freezes, drops, joins, and recovery |
Separate signaling from media. Signaling handles room membership, permissions, offers, answers, ICE candidates, and state changes; it does not automatically solve media delivery. Use short-lived, scoped credentials and explicit authorization for room actions. The Video API Authentication Guide covers token claims, lifetimes, webhooks, and key rotation that are relevant to this boundary.
Privacy assumptions also belong in the worksheet. Decide whether recordings are enabled by default, who can subscribe to a participant, where media is processed, how long recordings remain available, and how deletion requests affect derived files. Avoid treating an MCU composite as a privacy shortcut: it can reduce exposed tracks in one output while still requiring careful control of source media.
Worked examples
Example A: small collaborative room
Suppose a product begins with rooms of two to four people, each expected to publish audio and video. Mesh may be a reasonable proof-of-concept because the topology is easy to explain and the initial server footprint is small. The worksheet should still test weak uplinks, background tabs, mobile devices, and a participant leaving or rejoining.
If the product roadmap includes larger rooms, do not optimize only for the first release. Measure the application’s signaling and UI separately from the media topology so an SFU can be introduced without rewriting room permissions and participant state.
Example B: moderated panel
For a panel with several speakers and a larger audience, an SFU can keep speaker media interactive while controlling which streams each client receives. Simulcast or layered encoding may allow the SFU to forward a lower-quality layer to a constrained device and a higher layer to a larger display, provided the selected SDK and browser paths support the required behavior.
If viewers do not need two-way interaction, consider a separate broadcast delivery path instead of subscribing every viewer to the interactive room. The appropriate choice may combine an SFU for the stage with a low-latency streaming workflow for the audience. Compare that design with the requirements in the low-latency streaming stack guide.
Example C: fixed-layout recording
If the primary requirement is a consistent grid, branded composition, or single downloadable file, an MCU can be attractive. Model the number of simultaneous layouts, output resolutions, recording duration, and whether participants need individual-track recovery. If editors later need to reframe or remove one speaker, separate track recording may be more useful than a composite alone.
Proof-of-concept workflow
- Implement the smallest room with the same signaling, permissions, codec settings, and SDK version intended for production.
- Test room sizes in steps, such as two, four, eight, and the next planned launch size. Treat these as test points rather than universal limits.
- Use a repeatable participant mix: cameras on, cameras off, screen sharing, mobile devices, poor uplinks, and background CPU load.
- For an SFU, test subscription policies, simulcast layers, dominant-speaker changes, and rapid join or leave events.
- For an MCU, test layout changes, source loss, composition recovery, recording start and stop, and output quality.
- Capture join time, connection success, round-trip delay, packet loss, jitter, bitrate, freezes, audio interruptions, CPU, memory, battery impact, and server utilization.
- Repeat each scenario enough to distinguish a consistent failure from a one-off network event, then record the SDK, browser, operating system, region, and infrastructure configuration.
- Set acceptance thresholds before reviewing results. A useful threshold should describe user impact, such as unacceptable audio interruption or video freeze, rather than relying only on server utilization.
For a broader operational framework, connect these measurements to the streaming reliability metrics and incident runbook guidance.
When to recalculate
Revisit the architecture and worksheet whenever a major input changes. At minimum, recalculate before increasing the maximum room size, adding recording, enabling screen sharing, expanding to new regions, changing codec or simulcast settings, moving between SDK versions, or introducing a new mobile device class. Recheck after infrastructure pricing or egress terms change, because the cheapest topology at one usage pattern may not remain the cheapest at another.
Set a scheduled review at least twice a year for a production system, and an immediate review after a significant WebRTC, browser, SDK, codec, or media-server change. Preserve the previous worksheet so capacity decisions remain auditable.
The practical next step is to choose the smallest architecture that satisfies today’s interaction model, then test the next two planned growth points. Use mesh when its simplicity genuinely fits the room size and device profile; use an SFU when selective delivery, moderation, or interactive growth matters; use an MCU when centralized composition is worth its processing and operational cost. Keep the assumptions visible, measure user experience alongside infrastructure load, and recalculate before growth turns an architectural compromise into a reliability problem.