TL;DR
- Site count comes from the busier of two constraints: the coverage-limited radius and the capacity-limited throughput per square kilometre.
- Every traffic model needs three numbers per service: busy-hour session attempts, mean throughput per session, and concurrency.
- Average cell throughput is the wrong design KPI. Cellular capacity is scheduled, so design to a 5th-percentile user rate under the forecast load.
What inputs does a capacity model need?
Every capacity model reduces to three numbers per service: busy-hour session attempts per subscriber, mean throughput per session, and the concurrency factor that converts sessions into simultaneous active users. Multiply by the subscriber base and the penetration assumption, and you have busy-hour traffic volume for the area. Everything else — spectral efficiency, carrier bandwidth, MIMO layers — converts that volume into cells.
The forecast horizon matters as much as the numbers. A three-year horizon with a documented growth assumption is defensible; a ten-year horizon with the same method produces a site count nobody will fund. Plan the radio for three years and the civil works for ten, because towers are hard to add and carriers are not.
| Service | Busy-hour throughput per session | Typical concurrency | Design note |
|---|---|---|---|
| Voice over LTE or NR | 0.03–0.05 Mbit/s | 0.02–0.05 | Small volume, but sensitive to scheduling delay |
| Web and social | 1–3 Mbit/s | 0.10–0.20 | Dominated by session setup, not by sustained rate |
| Video streaming | 3–8 Mbit/s | 0.15–0.30 | The dominant capacity driver in most markets |
| Fixed wireless access | 10–50 Mbit/s | 0.30–0.60 | Behaves like wireline; concurrency far higher than mobile |
| Industrial uplink | 2–20 Mbit/s uplink | 0.50–0.90 | Uplink-limited, not downlink-limited; changes the whole design |
How do you convert traffic into a site count?
Compute the area traffic density in the busy hour: total busy-hour volume divided by the area served. Divide that by the sustainable throughput of one cell at the target load, and you have the number of cells the capacity constraint demands. Separately compute the number of cells the coverage constraint demands from the link budget and the target cell-edge rate. The site count is the larger of the two, and knowing which one binds is the single most useful output of the exercise.
If coverage binds, the answer is more sites or higher antennas. If capacity binds, the answer is more spectrum, more sectors, or more MIMO layers — and adding sites is the most expensive of the three.
Why is average cell throughput the wrong KPI?
Cellular capacity is scheduled, not pooled. Users share the cell in time and frequency, so the meaningful figure is the throughput a user at the cell edge or at the 5th percentile receives under the design load, not the average across the cell. A cell with 300 Mbit/s average throughput can easily deliver 2 Mbit/s to its worst-served tenth of users under load.
Designing to an average figure systematically under-provisions. The correct target is a percentile: for example, 90% of users above a defined rate in the busy hour at the forecast load. That target translates directly into a required spectral efficiency per unit area, which is what the capacity calculation actually needs.
Where does capacity planning usually go wrong?
The most common error is ignoring the uplink. Downlink capacity in a 5G NR cell with 100 MHz and 64T64R massive MIMO is large; the uplink budget is not, and uplink interference from neighbouring cells limits the cell-edge rate long before the downlink runs out. Designs that size on downlink alone produce networks where the downlink looks underloaded and users still complain.
The second error is treating indoor traffic as outdoor traffic. In dense urban areas a large share of traffic originates indoors, where penetration loss means the serving cell is doing more work per megabit than the outdoor model assumes. The third is ignoring control-channel overhead at high load: at 80% PRB utilisation the practical throughput is well below the linear extrapolation, and the last 20% of capacity is much more expensive than the first 80%.
How much headroom should you plan?
Plan the radio layer to reach 70–80% busy-hour PRB utilisation at the forecast horizon, not 100%. The remaining headroom absorbs forecast error, traffic shifts between cells, and the fact that real traffic is burstier than any model. Above 80% utilisation, scheduling delay rises and the user experience degrades faster than the throughput numbers suggest.