Power is the line item everybody scrutinises in a colocation contract. Bandwidth is the one that quietly costs more than expected, because two providers can quote what looks like the same price and bill you amounts that differ by a factor of three.
The difference is almost never the rate. It is the model used to turn your traffic into a number.
The Three Models
Flat Port Pricing
You pay a fixed monthly fee for a port of a given speed and use as much of it as you like. A 10 Gbps port costs what it costs whether you push 200 Mbps or 9 Gbps.
This is the easiest model to budget and usually the most expensive per unit of traffic actually moved, because you are paying for capacity rather than use. It makes sense when traffic is genuinely heavy and continuous, or when finance values a predictable number more than a lower average.
Committed Rate
You commit to a minimum monthly volume or sustained rate, pay for it whether you use it or not, and pay an overage rate for anything above. The commit buys a lower unit price.
This is the standard wholesale model and it rewards predictability. The risk is on both ends: commit too low and overage charges erase the saving, commit too high and you pay for capacity you never used. The overage rate matters as much as the headline rate, and it is frequently two to three times higher.
95th Percentile
The provider samples throughput every five minutes for the whole month. At the end they sort the samples, discard the top five percent, and bill on the highest one left.
In a 30-day month there are roughly 8,640 samples. Discarding five percent discards about 432 of them, which is around 36 hours of your busiest traffic. Those hours are, in effect, free.
That design is deliberate. It was built for traffic with short peaks and a lower sustained level, and it is still the dominant model for IP transit.
Why the Shape of Your Traffic Decides the Winner
Consider two tenants who both peak at 4 Gbps and both move a similar total volume.
The first runs inference for a consumer application. Traffic climbs through the working day, peaks sharply for three hours in the evening, and drops to a few hundred megabits overnight. Most five-minute samples sit well below the peak. After the top five percent is discarded, the billable figure might land near 1.2 Gbps.
The second replicates training datasets between regions on a continuous schedule, holding close to 3.5 Gbps nearly all the time. Discarding the top five percent barely moves the number. The billable figure is around 3.4 Gbps.
Same peak, similar volume, nearly three times the bill. The first tenant should be on 95th percentile. The second should be negotiating a committed rate, where the predictability is worth a discount.
The practical test is simple: divide your average throughput by your peak throughput. A ratio below about 0.4 means you are bursty and percentile billing will serve you well. Above 0.7 and you are effectively a steady load, so buy on commit and get the unit price down.
Details That Change the Number
Which Direction Is Measured
Most providers measure inbound and outbound separately and bill the higher of the two. Some bill the sum. On a workload where one direction dominates, which is most AI workloads, that distinction can double the bill. It should be explicit in the contract, not assumed.
Sample Interval
Five minutes is the convention. A shorter interval makes the measurement more sensitive to brief spikes, since each spike occupies a larger share of the samples and is less likely to fall inside the discarded top five percent. If a provider samples at one minute, the same traffic produces a higher bill.
Which Percentile
Ninety-fifth is standard but not universal. Ninety-ninth percentile billing discards only about seven hours a month rather than 36, which is a materially different product. The difference is one character in a contract and a substantial difference in cost.
Per-Port or Aggregated
With redundant ports, whether the percentile is calculated per port or across the aggregate matters. Aggregated is almost always better for the tenant, because traffic that fails over from one port to another does not create a new peak on the second port.
Burstable Capacity
A burstable port is physically provisioned above the level you pay for. A 10 Gbps port billed at 1 Gbps lets you exceed 1 Gbps without hitting a rate limit, and you only pay more if the elevated level is sustained past the discard window. This pairs naturally with percentile billing and is worth asking for explicitly, since the alternative is a hard policer that drops your traffic at the committed rate.
Reading a Quote Properly
When comparing providers, normalise everything to the same question: what would this cost for my actual traffic last month?
Ask for the model, the measurement direction, the sample interval, the percentile, whether ports aggregate, the overage rate if there is a commit, and whether cross-connects and IP space are separate charges. Then run your own traffic data through each. Providers will usually do this exercise with you, and the ones who decline are telling you something.
Watch for cross-connect charges in particular. A low transit rate paired with an expensive monthly cross-connect can lose to a higher transit rate with cheap interconnection, especially at lower volumes. Our guide to interconnection and cross-connects covers how those are priced.
When Transit Is Not the Right Product
Above a certain scale, blended IP transit stops being the economical answer. Direct peering with the networks your traffic actually goes to removes transit cost for that share of traffic entirely, at the cost of the cross-connects and the operational effort of running the sessions.
Carrier-neutral facilities with a well-populated meet-me room make this practical; single-carrier buildings do not. If a meaningful share of your traffic goes to one or two large destinations, ask which networks are present in the building before signing. The trade-offs between buying capacity and buying fibre are covered in dark fiber versus lit services.
What This Looks Like in the Gulf
Regional specifics change the arithmetic. International transit out of the Gulf carries a different cost structure to intra-region traffic because of subsea capacity, so a workload serving users across Europe and South Asia will see a different blended rate to one serving the UAE alone. Providers vary in how much of that they expose in the quote.
For AI inference serving regional users, the combination that usually wins is a burstable port on 95th percentile with local peering for the domestic share of traffic. For training workloads that move large datasets on a schedule, a committed rate negotiated against a realistic volume forecast will almost always beat it.
Rax Data and Energy provisions burstable ports with 95th percentile billing as standard and offers committed-rate pricing where traffic is predictable enough to benefit. If you want us to model your actual traffic against both, send us a month of data and we will run the comparison.
Frequently Asked Questions
What does 95th percentile billing mean?
The provider samples your traffic every five minutes across the billing month, discards the highest five percent of those samples, and bills on the highest sample that remains. In a 30-day month that discards roughly 36 hours of peak traffic, so short bursts are effectively free and sustained load sets the bill.
Is 95th percentile cheaper than a committed rate?
It depends entirely on the shape of your traffic. Spiky traffic with a low sustained baseline is far cheaper on 95th percentile. Steady traffic close to its peak costs about the same under either model, and a committed rate usually wins on unit price once volume is predictable.
Are inbound and outbound billed together?
Normally the provider measures both directions and bills on whichever is higher, rather than adding them. This matters for inference workloads, where outbound responses can be much larger than inbound requests, and for training data ingestion, where the reverse is true.
What is a burstable port?
A burstable port is provisioned at a higher physical speed than the traffic level you pay for. A 10 Gbps port billed at 1 Gbps on 95th percentile lets you exceed 1 Gbps freely for short periods without a rate limit, and only pay more if the elevated level is sustained.
Does 95th percentile billing suit AI inference traffic?
Usually yes. Inference demand is typically peaky, driven by user activity that concentrates into parts of the day, with a much lower average than peak. That profile is exactly what the model was designed for. Continuous large-scale data transfer, such as replicating training datasets, suits a committed rate better.
A Real Energisation Date, Not a Brochure
Rax operates where transmission investment has run ahead of load, and we are direct about staged capacity and real dates. Talk to our team about sizing a deployment against a timeline.
Contact Rax Data & Energy