1. Executive Summary & Scale Metrics
During the 2020 pandemic, Zoom grew from 10 million daily meeting participants to over 300 million daily meeting participants almost overnight. While legacy enterprise tools (Cisco WebEx, Skype for Business) suffered massive freezing, degraded audio, and dropped calls, Zoom became the ubiquitous standard because it consistently worked smoothly, even on terrible home Wi-Fi connections.
Real-Time Video Conferencing Scale
ZoomMetrics for sub-150ms video conferencing across millions of concurrent participants
Achieving this required abandoning traditional Peer-to-Peer (P2P) WebRTC meshes and heavy Multipoint Control Units (MCU), building instead a globally distributed, custom-routed Selective Forwarding Unit (SFU) architecture optimized around the physical constraints of home internet bandwidth.
2. Requirements & Production Constraints
Functional Requirements
- Many-to-Many Video & Audio: Support meetings from 2 people up to 1,000 simultaneous video participants.
- Dynamic UI Layouts: Seamless switching between Active Speaker view (large 720p/1080p center stage) and Gallery Grid view (49 video thumbnail tiles).
- Screen Sharing: High-resolution, frame-rate prioritized screen sharing with real-time cursor annotations.
Non-Functional Requirements & The Upstream Bandwidth Limit
- The Upstream Asymmetry Bottleneck: Most residential broadband connections are asymmetric (e.g. 100 Mbps download, but only 5–10 Mbps upload). A participant cannot upload separate video streams to 49 other participants.
- Conversational Latency Threshold: Human speech patterns break down when audio lag exceeds 150 milliseconds (people start interrupting and talking over each other).
- Client CPU & Thermal Throttling: Decoding 49 separate video streams simultaneously cannot melt the user's laptop battery or spin cooling fans to 100%.
3. The Naive Design & Why It Collapses
NAIVE APPROACH 1: PEER-TO-PEER (P2P) MESH
[Alice] ─── (720p Stream) ───> [Bob]
│ ▲
└──────── (720p Stream) ───┴──> [Charlie] ... [49 Participants]
NAIVE APPROACH 2: MULTIPOINT CONTROL UNIT (MCU)
[Alice, Bob, Charlie...] ── (1 Stream Each) ──> [Central MCU Server]
│
▼ (Decodes all streams, stitches into grid)
[All Clients] <──────── (Single Grid Stream) ──────────┘Why P2P Mesh & MCU Transcoding Architectures Fail in Group Calls
ZoomTwo fundamental barriers that break video conferencing above 5 participants
P2P Mesh Upstream Bandwidth Collapse
criticalIn a 20-person P2P call, each user must upload 19 separate video streams (19 × 1.5 Mbps = 28.5 Mbps upload). Residential home Wi-Fi uplinks saturate instantly, freezing all video feeds.
MCU Transcoding Latency & Server CPU Meltdown
criticalDecoding, composing, and re-encoding 50 incoming video streams on a central server adds 300–400ms of transcoding latency, destroying natural conversational flow and requiring massive server supercomputers.
4. Deep Architecture: Layer-by-Layer Walkthrough
Zoom solves this with Selective Forwarding Units (SFUs), Client Simulcasting, and UDP Forward Error Correction (FEC).