95th Percentile vs Fixed Bandwidth: Billing Math for High-Traffic Servers

Bandwidth on a dedicated server is sold four different ways, and three of them are designed to look cheaper than they are. A 1 Gbps port quoted at $89 and another at $349 can carry the same traffic pattern; what differs is when the meter starts and how spike seconds are counted. This breakdown converts each billing model into a monthly dollar figure using the same traffic sample, so the comparison is arithmetic rather than marketing.

The four models in plain terms

Fixed bandwidth (rate-limited). The port is capped at a speed, and usage is unlimited up to that cap. Predictable, and usually the most expensive per Mbps.
Metered transfer. A monthly TB allowance with per-TB overage. Simple to understand, brutal during launch weeks.
95th percentile. The provider samples your throughput every 5 minutes (288 samples per day, ~8,640 per month), discards the top 5%, and bills the highest remaining sample as your sustained rate.
“Unmetered” with fair use. No stated cap, but a clause permits throttling or surcharging above an undisclosed threshold.

Worked example: one month of real traffic

Take a game network with a launch spike and a steady evening peak. Traffic over 30 days:

Period	Duration	Average throughput
Baseline	600 hours	180 Mbps
Evening peaks	108 hours	520 Mbps
Launch spike (DDoS + players)	6 hours	1,850 Mbps
Three short bursts	40 minutes total	3,200 Mbps

Total transfer: (600 × 180 + 108 × 520 + 6 × 1,850 + 0.67 × 3,200) Mbps-hours ≈ 108,000 + 56,160 + 11,100 + 2,144 = 177,404 Mbps-hours. Multiply by 3600 to get Mbit, divide by 8 for bytes: roughly 79.8 TB for the month.

Cost under each model

Model	Calculation	Monthly cost
Fixed 1 Gbps, $180/mo	Flat, spike absorbed by port cap	$180
Fixed 500 Mbps, $120/mo	Port caps at 500; spike traffic drops or queues	$120 (with packet loss)
Metered, 20 TB included, $2.50/TB over	59.8 TB over × $2.50	$149 + $120 base = $269
95th percentile, $6/Mbps	~520 Mbps billable × $6	$3,120
Unmetered, fair use at $110	Flat until provider decides otherwise	$110 (risk-priced)

The 95th percentile line looks catastrophic — but only because this traffic pattern is the worst case for it. The model bills the 95th-percentile rate, not the volume, so a workload with a low, flat rate pays very little. For a steady 300 Mbps streaming or CDN origin, 95th percentile at $6/Mbps costs $1,800 and the metered model at the same volume would cost far more. The billing model is not inherently fair or unfair; it either matches your traffic shape or it does not.

How 95th percentile actually computes

The sampling detail decides everything:

Sample interval. Five minutes is standard. A 1-minute interval makes short spikes far more expensive, because 5% of 43,200 samples is 2,160 minutes (36 hours) rather than 432 minutes (7.2 hours).
Inbound vs outbound. Most providers bill the larger of inbound and outbound, some bill only outbound, and a few bill the sum. Ask explicitly.
Burst allowance. Some contracts allow two spike windows per month to be excluded. This is worth more than a 10% rate discount for a games business.
Averaging window. A “95th of daily 95ths” averages 30 values and is much more forgiving than the straight monthly 95th.

Under the 5-minute model, the breach threshold is 432 minutes of high-rate traffic per month. Your 6-hour launch spike alone (360 minutes) consumes 83% of that allowance. Add three bursts of 13 minutes and you are over, and the billed rate lands at the spike level.

The real trap: DDoS traffic counts

On metered and 95th-percentile plans, attack traffic that reaches your port is billable unless it is dropped upstream. A 20 Gbps flood that gets filtered at the provider’s edge but still transits your committed link can produce a five-figure overage in one afternoon. Before signing, get three answers in writing:

Is scrubbed attack traffic billed to my account?
Is there an automatic surge-protection clause that pauses billing during a declared attack?
At what Gbps does the provider null-route the destination IP, and does that trigger a plan suspension?

A provider that answers all three clearly is worth a premium. One that says “don’t worry about it” is transferring the financial risk of their own filtering gaps to you.

Choosing by traffic shape

Your traffic shape	Best model	Why
Constant, predictable bitrate	Fixed bandwidth	Cheapest per Mbps at high commit
Low rate, huge monthly volume	Metered or unmetered	Rate-based billing punishes volume
Sharp peaks, quiet troughs	95th percentile with burst allowance	Peaks are partly discarded
Marketing launches, seasonal traffic	Fixed or metered with surge clause	Avoid rate-based overage on spikes

Model your own numbers before you buy

Export 30 days of interface counters from your current host at 5-minute granularity — sar -n DEV -f /var/log/sa/saXX or your provider’s graph export — then compute the 95th percentile and total volume yourself. That gives you a billable rate and a TB figure you can put in front of a sales team. Providers rarely volunteer both numbers, and the negotiation usually favours whichever one you did not ask about. For servers where port speed and filtering are the actual constraint rather than the meter, dedicated server hosting with clearly stated bandwidth terms removes most of this ambiguity at the contract stage.

Ask these five questions before you sign

What is the sampling interval, is inbound counted, is scrubbed attack traffic billable, what is the overage rate per TB, and is the port rate-limited or burstable? Five answers, five lines in a spreadsheet — and a monthly bandwidth bill you can predict to within 10% instead of discovering at invoice time.