How to Build an App Like Zoom: Video Infrastructure Reality

Why mesh fails past three people

The simplest WebRTC topology has everyone send their stream to everyone else. With N participants each device uploads N-1 streams.

At four people that's three uploads per device: already beyond typical home upstream bandwidth, and punishing on mobile CPU and battery. Mesh is viable for one-to-one and nothing more.

SFU is what production uses

A Selective Forwarding Unit means each participant uploads once to a server that forwards streams to everyone else. Upload cost becomes constant regardless of participant count.

What the SFU then has to do well is where the difficulty lives:

• Simulcast: clients publish several quality layers so the server can send each receiver an appropriate one
• Bandwidth estimation per receiver, adapting continuously as networks change
• Active speaker detection, so you forward the streams that matter rather than all of them
• Packet loss recovery through retransmission and forward error correction
• Geographic distribution, because a call routed through the wrong region adds latency nobody can fix client-side

Each of those is a specialist domain. Together they're why video infrastructure companies exist.

The parts people forget

TURN servers. Roughly a fifth of connections can't establish peer-to-peer through NAT and firewalls, and must relay. TURN relay bandwidth is a real, ongoing cost.

Recording means server-side composition of multiple streams into one file. CPU-heavy, storage-heavy, and it creates retention and consent obligations, especially in regulated settings.

Echo cancellation and noise suppression. Browsers provide baseline processing; the perceived quality gap between adequate and excellent audio is larger than for video, and audio problems end calls faster than video problems.

What to demand from a provider

Since you should almost certainly buy, Daily, LiveKit, Twilio, Agora, Vonage, evaluate on: SFU with simulcast, geographic coverage matching your users, recording controls and where recordings are stored, graceful degradation to audio-only, a BAA if healthcare is in scope, and transparent per-minute pricing.

LiveKit is worth noting as open source, so self-hosting stays available if per-minute economics eventually stop working.

When building is defensible

Only when video is the product and you have differentiated requirements, unusual codecs, very large broadcasts, or specialised low-latency needs, plus the team to own it. For a telemedicine, education, or support product, video is a component and buying is correct.

Related builds

• How to build a telemedicine app. BAA-covered video in a clinical flow
• Best video SDK for telemedicine apps
• How to build an eLearning platform