Zoom did not win because it waited for browsers to become a complete video OS. Native clients could pick codecs, tune congestion control, and fall back across UDP, TCP, and proprietary tricks that a tab cannot always match. The browser still matters: a join link that requires a 200 MB download loses the meeting that starts in ninety seconds. Zoom's engineering story is a dual stack — WebRTC where the web is the right container, and a native media engine where packet loss, CPU, and screenshare demand more.
WebRTC is a contract, not a quality ceiling
WebRTC gives you ICE, DTLS-SRTP, and a peer connection model that NAT traversal people already understand. For Zoom, a meeting is rarely a single peer connection to another person. It is a connection to a media server (SFU-style forwarding or mixing, depending on the era and the meeting type). Simulcast and layered coding let the server forward a quality the receiver can handle without decoding every sender at full rate in the cloud. Browsers have gotten better at this; they still vary by OS, GPU, and what hardware encoder is actually exposed.
Custom or highly tuned codecs on native clients exist because video is a CPU and bandwidth budget. A codec that looks great in a lab at 2 Mbps may melt a 2018 laptop when there are twenty tiles. Zoom's native stack historically invested in its own optimizations around H.264 and later more efficient codecs, plus screenshare modes that treat text slides differently from camera noise. The browser path has to pick from the codec menu WebRTC and CDNs will actually interoperate with.
Congestion, recovery, and ugly networks
The real Zoom network is hotel Wi-Fi, a corporate proxy that blocks UDP, and a mobile handoff. Native clients can implement custom retransmission, forward error correction, and bandwidth estimation that react in tens of milliseconds. WebRTC's congestion control has improved (GCC and successors), but product-level decisions — which participant's video to drop first, when to freeze a tile rather than scramble — live in Zoom's media engine and SFU, not in the W3C spec.
Audio is unforgiving. Opus in WebRTC is excellent; Zoom still has to keep AEC, noise suppression, and jitter buffers coherent across mixed clients. A meeting where the browser user sounds like a tunnel and the native user is fine is a support ticket that looks like "Zoom is broken" rather than "two DSP pipelines."
Failure modes of dual media stacks
The concrete failure is a feature that works on desktop and silently degrades in WebRTC — virtual backgrounds, certain stereo layouts, a codec — without the UI saying so. Users compare screenshots. Mid-size steal: a capability matrix per client, and tests that join a meeting with one browser and one native client on a lossy link.
Operational gotcha: forcing TCP or WS fallback for everyone behind a firewall "to make it work" will make it work poorly. Keep UDP as the primary path and instrument fallback rate. Another is hardware encoder bugs: a GPU that advertises H.264 and produces artifacts will look like a Zoom defect. Offer a software fallback and a way to disable hardware encode. Screenshare at 4K into a dozen mobile viewers is a denial-of-service you will inflict on yourselves; cap and layer. If you build a meeting product, do not assume WebRTC peer-to-peer scales to a webinar. Put an SFU in the middle early. Version skew between clients during a codec rollout is a split brain: some participants cannot decode. Gate the codec on a meeting-wide minimum, not per sender optimism.
What you can borrow
- Use WebRTC for reach; keep a native media engine for the networks and CPU budgets browsers cannot fully own.
- Put an SFU in the middle before you need webinar scale; P2P mesh dies at modest N.
- Instrument UDP vs TCP fallback and treat high fallback as an incident, not a success.
- Roll out codecs meeting-wide with a minimum client version, not sender by sender.