Per-stream planning numbers (practical, not lab-best)
| Modality | Plan per concurrent stream |
|---|---|
| Audio | ~100 kbps |
| Video (group meeting, typical) | 0.5-1.5 Mbps |
| Screenshare | 0.2-1 Mbps (content-dependent, bursty) |
| Town hall viewer | 1.5-2.5 Mbps DOWN only |
Teams adapts downward under pressure (SVC layers, resolution drops) — which is why 'it works' and 'it looks good' are different budgets. Plan for the experience you want, not the minimum that connects. Policy can cap the ceiling: media bit rate in meeting policy (in your scan) trades quality for predictability on thin sites.
The concurrency model (per site)
Peak concurrent meeting users ≈ headcount × meeting-hours-per-day factor — but measure instead: CQD tells you actual concurrent streams per subnet at peak. Then the shape matters more than the sum:
- Symmetry: meetings are up AND down (one uplink each + downlinks); town halls are down-only floods (see eCDN); a site's asymmetric DSL breaks in the UP direction first — the classic branch complaint.
- The event exception: normal planning is irrelevant during all-hands; that budget is the eCDN concept's arithmetic.
- Headroom, not averages: voice quality lives at the 95th percentile of utilisation; a link 'only 60% average' can be saturated at every meeting's first five minutes. QoS (earlier concept) protects the audio when you're wrong.
Microsoft's Network Planner (TAC) gives per-persona estimates — fine as a first draft, blind to your real meeting culture; treat its output as a hypothesis for CQD to confirm.
What to watch (proofs)
- Actual per-site load: link utilisation graphs at meeting peaks (not daily averages) — 95th percentile against capacity.
- Actual stream mix: CQD streams by subnet/hour — your empirical concurrency factor per site type; recalibrate the model quarterly with it.
- The cap in force: media bit-rate settings in scanned meeting policies — deliberate ceilings documented, accidental ones caught as drift.
- Complaint correlation: poor-call % per site vs that site's 95th-percentile utilisation — the scatter plot that ends 'buy more bandwidth' debates with a yes or a no per site.