Two servers move exactly the same number of terabytes in a month. One is billed for 50 Mbps, the other for 900 Mbps. Neither invoice is wrong, and neither provider is cheating. That is 95th percentile bandwidth billing, and once you understand it you can usually predict your own bill to within a few percent — or decide you would rather not have one.
It is the dominant pricing model for internet transit and it leaks into dedicated server and colocation contracts, where it is often called burstable billing. It does not measure how much data you moved. It measures how hard you pushed, consistently.
What is actually being measured
The provider's router counts bytes on your port and records a sample every five minutes. Each sample is the average rate over that interval — bits transferred divided by 300 seconds — not an instantaneous reading. Inbound and outbound are counted separately, and you are billed on the higher of the two.
At the end of the billing month:
- All the samples are sorted from highest to lowest.
- The top 5% are discarded.
- The highest remaining sample is your billable rate.
That rate, in Mbps, is multiplied by your per-megabit price. The number of gigabytes you moved never enters the calculation.
The arithmetic behind the 5%
A 30-day month is 720 hours, which at one sample every five minutes is 8,640 samples.
Five percent of 8,640 is 432 samples, and 432 × 5 minutes is 2,160 minutes — exactly 36 hours.
So the model hands you 36 hours a month, spread however you like, in which your traffic is invisible to the invoice. Everything else counts.
Why a spike is free and a plateau is not
Three sites, all moving very nearly the same volume over a 30-day month.
| Site | Traffic shape | Spike hours/month | Volume moved | Billable rate |
|---|---|---|---|---|
| A | Steady 80 Mbps, all month | 0 | 25.9 TB | ~80 Mbps |
| B | 50 Mbps baseline, 900 Mbps for 6 hours once a week | 24 | 25.4 TB | ~50 Mbps |
| C | 30 Mbps baseline, 900 Mbps for 10 hours once a week | 40 | 25.4 TB | ~900 Mbps |
Site B's four weekly spikes total 24 hours, comfortably inside the 36-hour allowance, so every one of them lands in the discarded top 5%. It is billed at its baseline.
Site C moves the identical 25.4 TB, but concentrates more of it into spikes: 40 hours a month. Discarding the top 36 hours still leaves four hours of spike in the sample set, so the 95th percentile sample falls inside a spike. The bill goes to roughly 900 Mbps.
Same terabytes, same month, and Site C pays about eighteen times what Site B pays. Nothing about the underlying business is different — only the shape. That cliff at 36 hours is the defining property of the model, and it is why 95th percentile invoices surprise people who thought they understood their own traffic.
Site A is the third lesson: it moved marginally more than B and pays 60% more, because it pushed consistently instead of in bursts. Under this model a flat, well-behaved workload is the expensive one.
The three models, side by side
| 95th percentile | Transfer quota | Unmetered | |
|---|---|---|---|
| Billed on | Sustained rate, Mbps | Total volume, TB | Nothing — flat fee |
| Short bursts | Free, up to 36 hours a month | Counted in full | Free |
| Sustained load | Expensive | Counted in full | Free |
| Predictable in advance | Only if you model it | Yes, if you forecast volume | Yes, exactly |
| Worst surprise | A spike that ran too long | An overage line | Hitting the port ceiling |
| Suits | Spiky traffic, low baseline | Low, steady egress | Anything heavy or unpredictable |
There is no universally cheaper option. There is an option that matches the shape of your traffic, and the shape is something you can measure.
Where the model quietly bites
- Backups at a fixed hour. A nightly job that saturates the link for two hours runs 60 hours a month. That is well past 36, so your backup window sets your bandwidth bill. Stagger it, throttle it, or move it off the billed port.
- Cron convergence. Every job at the top of the hour produces a repeating spike at exactly the sample boundary. Spread them.
- CDN origin fill. A cache purge makes every edge come back to you at once. Purge selectively.
- Inbound counts too. Ingesting logs or accepting large uploads is billed the same as serving. People forget this until an ingest pipeline turns up on the invoice.
- A DDoS attack is traffic. Some contracts exclude attack traffic; many do not. Read the clause before you need it, and filter at the network edge so the flood never reaches your port.
Working out which model is cheapest for you
You need two numbers from your own graphs, not from an estimate:
- Your 95th percentile rate. Most monitoring tools can compute it directly; if not, export five-minute samples for a month, sort, and take the value 5% down from the top.
- Your monthly volume in terabytes.
Then compare three quotes on the same traffic: rate × per-Mbps price, volume × per-TB price after the included quota, and the flat unmetered fee. Whichever is lowest is lowest for your shape, and it will change if your shape changes. A plateau-shaped workload almost always does better on quota or unmetered pricing; a spiky, low-baseline workload is where 95th percentile earns its reputation for being cheap.
The metered-versus-unmetered comparison is worked through in more detail in Unmetered Bandwidth vs Metered: Working Out Which Is Cheaper, and sizing your requirement from first principles is in How Much Bandwidth Does a Server Need?
To measure rather than guess, monitor bandwidth and traffic on the server for a full month before you commit to a contract shaped around a number you have not seen.
The fourth option: no meter at all
NexonHost does not use 95th percentile billing on dedicated servers. Every build ships with an unmetered 1 Gbps port: no monthly transfer quota, no per-megabit rate, no overage line, and inbound is not metered either. The only ceiling is the physical port — about 324 TB in a 30-day month at line rate on 1 Gbps — and that ceiling moves when you choose a faster port.
That trade is worth stating plainly rather than selling. If your traffic is genuinely spiky with a low baseline, a well-priced 95th percentile contract can beat a flat fee, and you should take it. If your traffic is a plateau, or if the thing you actually want is an invoice that does not move, unmetered dedicated servers remove the variable entirely — and 10 Gbps dedicated servers move the ceiling to 3.24 PB a month, with up to 200 Gbps available on the larger AMD EPYC builds. The full configuration list is on the dedicated servers page.
Specifications and pricing are as of September 2026 — check the product pages for current figures.
Sources
- Chapter 2 — Internet Transit, The Internet Peering Playbook, DrPeering. Describes per-Mbps transit pricing metered by 95th percentile sampling. Retrieved 21 September 2026.
- What is Network 95th Percentile?, bunny.net Academy. Five-minute sampling and top-5% discard. Retrieved 21 September 2026.
- Sample and hour counts computed from a 30-day month: 720 hours ÷ 5 minutes = 8,640 samples; 5% = 432 samples = 36 hours.
- NexonHost unmetered dedicated servers. Retrieved 21 September 2026.
