arrow_back All posts

Engineering · 12 min

WebRTC at Scale: SFU, MCU, and the Architectures Nobody Documents

Building a 100-participant video room is easy. Building one that handles 10,000 simultaneous rooms across three regions is a different story.

MC
Marcus Chen
Feb 04, 2026 · 12 min read
WebRTC at Scale: SFU, MCU, and the Architectures Nobody Documents

We have shipped production WebRTC for telehealth, fintech webinars, and consumer social. Here is what the docs never tell you.

SFU is the default — but not for every workload

For 95% of use cases, an SFU (mediasoup, LiveKit, Janus) wins on cost and quality. But recording-heavy or transcoding-heavy workloads still benefit from MCU pipelines.

Simulcast is your scale lever

Encode 3 layers at the publisher, let the SFU pick per-subscriber. This single decision unlocks rooms 4-5x larger.

TURN is 60% of your bill

Self-host TURN with coturn behind an autoscaling group. Save 70% over managed offerings.

Measure MOS, not bitrate

Bitrate is an input. MOS (Mean Opinion Score) is the output your users actually feel. Build dashboards around MOS.

Share this post