95th Percentile Billing: Why Your Bandwidth Bill Spikes

Your burstable circuit does not bill on your average. It bills on your peaks. And one bad day can set the rate for the whole month.

Most people try to fix this by using less bandwidth. Wrong move. The number that sets your bill is one specific sample out of thousands. You lower the bill by moving about 36 hours of peaks, not by cutting total traffic. Here is the math.

The 37-sample rule

Your carrier takes a reading every five minutes. Over a 30-day month, that is about 8,640 readings. At the end of the month they throw out the top 5 percent. That is 432 readings, or 36 hours of your busiest moments. The next one down, the 433rd-highest, sets your billable speed.

Think of it this way. Your bill is set by your 37th busiest hour, not your average. You get 36 hours of free spikes. Hour 37 is the one you pay for.

How a 40 Mbps average becomes a 180 Mbps bill

Say you have a 100 Mbps commit at $800 a month on a 1 Gbps port. Burst runs $9 per Mbps over commit.

Your nightly backup runs 2.5 hours at 180 Mbps. It runs 22 nights a month. That is 55 hours of traffic at 180 Mbps. Your free discard window is only 36 hours. So the 95th percentile lands at 180 Mbps.

Now the math. Overage is 80 Mbps over commit. At $9 each, that is $720. Your bill is $1,520. You expected $800.

Now compress the same backup to 1.5 hours a night. That is 33 hours a month. Under the 36-hour window. The spike gets discarded. Your 95th drops back to commit. Your bill is $800 again. Same total data. Half the bill.

Reading your invoice and your graph

Pull your invoice and find these lines: "Committed Data Rate" or "CIR," "95th Percentile In," "95th Percentile Out," "Burstable Usage," "Overage Rate per Mbps," and "Port Speed."

Then pull your carrier portal graph or your own SNMP or MRTG export. Match the interval. Compare the two.

One thing to check in your contract. Most carriers bill on the higher of inbound or outbound, so an asymmetric spike counts in full. A few bill on the sum of both directions, which roughly doubles your exposure. Find that language before you renew.

The four peak sources that set most bills

Four things cause most spikes. Each shows up as a tall, repeating block on the graph.

Backups and offsite replication. Usually overnight. Fix: cap the speed in Veeam or on your Synology.

OS and patch distribution. Shows up on Patch Tuesday. Fix: throttle in SCCM or stagger the rollout.

Security camera and video uploads. A steady wall all day. Fix: QoS shape it at the firewall to 90 percent of commit.

Cloud sync. OneDrive, Dropbox, Google Drive, all hammering at 9 a.m. Monday. Fix: cap the client or reschedule the big syncs.

Setting the commit: the $800 vs $1,520 decision

Pull three months of 95th values. Set your commit at the median, not the max.

Here is why. On a 1 Gbps port, moving your commit from 100 Mbps to 200 Mbps often costs under $300 more a month. But bursting 100 Mbps over commit at $9 each costs $900. Same 100 Mbps. Triple the price.

Overage rates usually run 1.5x to 3x the per-Mbps commit rate. That ratio is the number to check. If you burst often, buying the commit is almost always cheaper.

Clauses to change at renewal

Fight for these at renewal:

Sampling interval. Ask for five minute. Reject one minute. A shorter interval catches more peaks.

Measurement method. Max of in and out, not sum.

Overage rate cap. No more than 1.25x your committed per-Mbps rate.

Right to adjust commit down each quarter without penalty. Most contracts only let you go up.

Burst forgiveness or a 90th percentile option. Lumen, Zayo, and Cogent will sometimes grant this on request.

And one more. Once your 95th runs over 60 percent of your port speed, flat-rate dedicated internet at the same port is often cheaper. Run both numbers.

The call to make this week

Call your account rep. Ask for three things.

First, 12 months of 95th percentile history, in and out. Second, the sampling interval in writing. Third, a flat-rate quote at your current 95th level and at commit plus 50 percent.

Then ask for a one-time overage credit this month if a single event caused the spike. Carriers grant this more often than people expect. It never hurts to ask.

Stop cutting total bandwidth. Start managing the 36 hours of peaks. Pull the graph, find the recurring spike, throttle or reschedule it under the 5 percent line, then reset your commit to the real median.

Not sure what your 95th is costing you? Upload your bill and we will read the graph with you.