How Datacenter Bandwidth Billing Actually Works (Without the Hand-Waving)

Bandwidth is the most misunderstood line item in hosting. Compute is easy to compare — cores are cores, gigabytes are gigabytes. But network billing has three competing dialects, and quoting two providers who use different ones is how people end up convinced the datacenter is overcharging them when the real problem is the units.

Committed rate: the floor you pay for

A committed rate (you will see CIR — committed information rate) is capacity the provider guarantees is yours whether you use it or not. If your plan says 25 Mbps committed, that 25 Mbps is carved out of the uplink for you alone; congestion elsewhere on the segment cannot take it away.

Committed capacity is cheap to guarantee and cheap to buy. The catch is the opposite of a data cap: it is a floor, not a ceiling. The question that matters is what happens above the commit — and that is where the billing models diverge.

The 95th percentile: bursting without a scare

The industry-standard answer is 95th-percentile billing. Your port is sampled every few minutes around the clock. At the end of the month, the top 5% of samples — about 36 hours' worth — is thrown away, and you are billed on the highest remaining sample. In plain terms: you pay for your sustained ceiling, and your bursts are free.

This is why it is the standard: it matches how networks are actually built. A provider provisions uplinks for the aggregate of everyone's sustained use, and individual customers' short bursts statistically cancel each other out. Backups run overnight, a release spike hits on Tuesday, a validator catches a burst of traffic — 95th percentile absorbs all of it without a support ticket or a surprise invoice.

Per-GB vs percentile: who each model wins for

• Steady, predictable throughput (streaming, replication, CDN origins) → per-GB or a high commit; you know your baseline and flat-rate models make you subsidize other people's spikes.
• Bursty workloads (web apps, APIs, backups, validators) → 95th percentile, almost always. The free top 5% is exactly where your spikes live.
• Mostly idle machines (DR targets, standby databases) → 95th percentile again; low samples mean a low bill.
• Anyone who has ever been burned by an overage invoice → percentile with a hard conversation, never per-GB.

What we include, and why the bill never surprises you

Every xShredo colocation plan rides a redundant 10 Gbps uplink port with a 25 Mbps committed rate — burstable, with no hard cap. That is deliberate: the commit guarantees your floor, and the 95th-percentile model means a traffic spike on your heaviest day does not become an invoice on your worst one.

If your sustained ceiling starts pushing the commit, the right move is a conversation, not a penalty. We would rather upsell you into the capacity you actually need, with a graph on the table, than bill you for exceeding a number you did not know you had.

The takeaway

When you compare colocation quotes, translate everything into the same dialect first: what is committed, what happens above the commit, and how the monthly number is derived. A "cheaper" per-GB quote with bursty traffic is regularly the more expensive one by month three.

Bandwidth should be boring infrastructure, not a monthly puzzle. Ask any provider you talk to — including us — to walk through the math on your real traffic shape. The good ones will enjoy the question.