How to evaluate a data center and its operator
A data center is not judged on a site visit or a brochure. The building almost always holds: what decides your nights is how it is run day to day. Here is the eight-axis method, the exact questions to ask in writing, and the grid to score the answers.
Estimate my cost →You do not evaluate a building, you evaluate an operation #
Most evaluation grids stop at the concrete: Tier level, redundancy, certifications, floor space. Those criteria matter, and they are easy to check once and for all. But they describe a design, not a practice. Two sites identical on paper do not behave the same way when a UPS fails at three in the morning.
What will cost you dearly is almost never a building failure. It is a maintenance window announced three days ahead that lands during your financial close, a cross-connect delivered in five weeks instead of five days, a technician turned away at the door because his badge was not ready, or an on-call number ringing into the void. None of this appears in a Tier level.
The eight axes below are therefore ordered like this: first what happens when all is well and the site performs maintenance, then what happens when things break, and only at the end what you actually buy.
Axis 1. Planned maintenance, the clause that empties the availability figure #
This is the least understood point in the trade, and the most expensive. Virtually every hosting contract excludes planned maintenance windows from the availability calculation. The sentence sits in one line among the annexes, and it changes everything.
Do the arithmetic. A site targeting 99.982% availability, the level usually associated with Tier III, allows itself 96 minutes of downtime a year. If its contract provides for a four-hour maintenance window every month, that is forty-eight hours a year, or 2,880 minutes: thirty times the contractual outage budget, and not one of those minutes counts. Even with a more modest two-hour window each quarter, you reach eight hours a year, five times the budget.
This is neither an anomaly nor dishonest: a site that never performs maintenance is a site that will fail you for good. The point lies elsewhere: the published percentage says nothing about the availability you will experience until you know the volume and the regime of planned windows.
The six questions to ask about maintenance #
Ask them in writing, before signing. The answers take a few lines, and their absence is already an answer.
- How many hours of planned maintenance a year, averaged over the last three years? A serious operator knows that figure because it tracks it.
- What minimum notice period, and in what form? Seven working days is common practice for non-urgent work. Three days stops you from organising anything.
- Who picks the window? Imposed by the site, or negotiable with the customer? Can you refuse a slot and request another, and within what limits?
- Are there frozen periods? Many sites bar any intervention during holidays or their customers' known peaks. If that freeze exists, it is written down.
- What happens in emergency maintenance? What minimum notice, decided by whom, and communicated how? That is the case that reveals the real organisation.
- Is maintenance excluded from the SLA calculation? Almost always yes. Then ask for the availability figure excluding exclusions, that is, the availability a customer actually lived through.
Axis 2. Notifications, the most revealing test #
How an operator warns you says more about it than any certification. A proper maintenance notice contains six things: the exact date and time, the estimated duration, the equipment involved, the expected impact on you (none, automatic failover, one power feed down), the fallback procedure if the work goes wrong, and a contact reachable during the operation.
Half the notices on the market stop at the date and the equipment. They inform you without letting you decide anything.
The test that settles it, and it costs nothing: ask for the last three maintenance notices sent to their customers, anonymised. An operator who supplies them within twenty-four hours has a process, a template and a history. One who promises an example «as soon as I have one to hand» improvises every time, and will improvise on the day of your incident too.
What a notice must contain, and through which channel #
- The channel: a customer portal with a searchable history beats an email to a generic address. Ask whether notices are archived and retrievable six months later.
- The recipients: can you register several contacts, including a group address and an on-call line? A notice sent to an employee who left a year ago is a lost notice.
- Acknowledgement: does the site wait for your confirmation before working, or does it treat silence as consent?
- Closure: do you receive a report afterwards stating what was done and whether it went as planned? That document is your evidence in any later dispute.
- Unplanned incidents: how quickly are you told of an outage in progress, and through which channel? An operator who only tells you after restoration leaves you hunting for the cause on your own side throughout.
Axis 3. The indicators to demand, and over what period #
An indicator without a period and a measurement method is worthless. «We target 99.99%» is an intention; «99.97% measured over the last twelve months at the UPS, excluding two planned maintenance windows» is a fact. Demand the second form, and accept that it will be less flattering: an honest, poorer figure beats a round, unverifiable one.
Here are the indicators that actually decide, and the form in which to ask for them.
| Indicator | How to ask for it | Why it decides |
|---|---|---|
| Actual availability | measured over 12 rolling months, with the metering point and the list of exclusions | the only figure describing what a customer lived through |
| Planned maintenance | count and cumulative duration over 12 months | this volume is excluded from availability: it is its indispensable complement |
| Grid outages | count per year, and the longest generator run | tells you whether generators are a safety net or an operating mode |
| Fuel autonomy | hours at full load, and priority refuelling contract | 24 h of autonomy without a refuelling contract is a guarantee for 24 h only |
| First response time | median and 95th percentile, not the average | a 12-minute average can hide one case in twenty at three hours |
| Restoration time | median across the last 12 months of incidents | this is the restore time as lived, to compare with the one promised |
| Cross-connect delivery | median observed delay, from order to live link | observed delays range from same-day to several weeks, and nobody publishes them |
| Cold-aisle temperature | min, average and max over 12 months | a site that only tracks the average will never see its hot spots |
| PUE | measured over 12 months, not design, with the metering boundary | it multiplies your energy bill whenever energy is rebilled at actual use |
| Site fill rate | in subscribed electrical power, not floor space | a site full on power cannot support your growth, even with empty racks left |
Axis 4. The SLA, and the question nobody asks #
An availability figure is a budget of outage minutes, not a promise of no outage. That distinction, and how to turn an incident into a claimable amount, are covered in detail in our dedicated guide: Restore time, SLA, MTTR: what a commitment is really worth.
One question deserves adding here, because it precedes all the others: what does the contract call an outage? A fibre cut by a digger at two in the morning, repaired before dawn, while your servers kept running on the second feed: was there an outage? The answer does not depend on the facts, it depends on the definition written in the annex. Depending on the scope chosen, the same incident counts as zero minutes or as two hundred and forty.
So ask, in order: the definition of unavailability, the metering point, who declares it, and from which timestamp it runs.
Axis 5. Price, and the only comparison that means anything #
Comparing rack prices means comparing the one line competition has already crushed. The gap is made elsewhere: electricity and how it is billed, cross-connects, the transit port, IP addresses, setup fees, and remote-hands work billed by the hour.
The rule is simple to state and rarely applied: ask every candidate for the monthly budget of an identical configuration, defined by you, line by line. One rack, a given power draw, a number of links, a block of addresses. Anything a candidate does not publish must be named as missing, never estimated on their behalf: an incomplete total presented as complete distorts the decision more surely than a high price.
Two guides detail that calculation: the price of a rack and its hidden costs and comparing offers on an identical basis.
Axis 6. Connectivity, in four numbers and one question #
A site's connectivity is not a marketing claim, it is counted. Four numbers qualify it, and they can be checked in public interconnection registries, independently of what the operator says.
- The number of carriers physically present in the building, not reachable from it. That nuance decides your negotiating power: a site with three carriers gives you three quotes, a site with a hundred gives you a hundred.
- The number of Internet exchange points reachable from the site. They drive latency towards the large content networks and the cost of your transit.
- The cross-connect price and its delivery time. This is the line billed as you go, which can change without reopening the rack contract, and which most often escapes the forecast budget.
- Neutrality: does the site belong to a carrier, or is it independent? A neutral site has no interest in steering you towards one transport provider over another.
Axis 7. Physical access, what you discover on the first visit #
None of this appears on a quotation, and it costs more than a discount on rent. Badge lead time, decisive if you send a technician in an emergency: some sites impose forty-eight working hours, and a technician sent without that notice leaves without getting in. Round-the-clock access without prior notice, or not. Mandatory escort, which lengthens every intervention and often makes night work impossible. The loading dock and goods lift, without which a full rack weighing several hundred kilos is delivered case by case. Parcel reception and how long they are stored. Tooling available on site.
Also ask for the actual usable rack height. A 47U rack holds five more servers than a 42U rack for an often identical rent: on the same budget, you do not house the same amount of hardware.
Axis 8. The soundness of the operator itself #
The last axis is no longer about the site but about the company running it. How many years has it operated this particular building, rather than «data centers» in general? How many people on site, and on what hours? Is there an on-call rota, and a written escalation procedure with names and delays?
Ask for their customer retention rate, or failing that three references you can call, in your sector and of your size. A reference you are refused is information. Finally, check the operator's financial health: a site that is bought out or in difficulty changes its pricing and service quality within months, and you do not move an infrastructure within months.
The scoring grid, to fill in for each candidate #
Score each axis from 0 to 3, then compare totals. The grid is not meant to designate an automatic winner: it is meant to make visible what you agreed to ignore. A candidate at 18 out of 24 with a zero on maintenance is riskier than one at 15 with no zero at all.
| Axis | 0 points | 3 points |
|---|---|---|
| Planned maintenance | annual volume unknown, window imposed | volume quantified over 3 years, negotiable window, written frozen periods |
| Notifications | generic email, minimal content | archived portal, six elements present, closure report |
| Indicators | intentions («we target») | dated measurements over 12 months, with method and exclusions |
| SLA | a percentage with no definition of unavailability | definition, metering point, cap and claim deadline in writing |
| Price | a rack rate alone | full budget line by line, gaps named |
| Connectivity | «we are well connected» | carriers present counted, cross-connect price and lead time published |
| Physical access | discovered on arrival | badge lead time, dock, parcel storage and tooling declared |
| Operator soundness | no reachable reference | references in your sector, on-call rota and escalation in writing |
The three-question test, if you only have ten minutes #
If you could ask only three questions, ask these. They are not about promises, they are about traces, and an operator cannot answer them without genuinely having the matching process.
One. «How many hours of planned maintenance did you carry out last year, and how many unplanned incidents?» The answer exists or it does not; there is no middle ground.
Two. «Send me your last three maintenance notices, anonymised.» You judge on evidence, within twenty-four hours.
Three. «What is your median cross-connect delivery time, observed over the last twelve months?» It is the most ordinary act in operations. A site that cannot measure it probably measures nothing else.
An operator who answers these three precisely has already told you the essential: it measures itself. One who answers with general qualities tells you something else, equally useful to know.
What DataColoc computes, and what it does not know #
On every data center listing we publish what can be verified without the operator: the carriers and exchange points present in the building according to public interconnection registries, the monthly budget of a reference configuration built only from published rates, the trajectory of the cross-connect price where we hold a dated history, and the cost of energy depending on the site's efficiency.
What we do not know, we name rather than estimate. The volume of planned maintenance, badge lead time, actually measured availability, the PUE of the last twelve months: that information belongs to the operator alone, and only it can publish it. That is why every listing carries the list of questions to ask, and why an operator who claims its listing and publishes those answers stands out immediately from those who stay silent.
FAQ #
It can, and that is an answer in itself. The volume of planned maintenance is neither a trade secret nor sensitive data: it is operational information any serious site tracks for its own steering. A refusal signals either the absence of measurement, or a volume the operator would rather not show.
Seven working days is common practice for non-urgent work, fifteen days for heavy operations touching power. Below three days you can no longer move a financial close or warn your own customers. For emergency maintenance the notice is counted in hours, but the contract must state who qualifies it as urgent.
Almost never. It is the most important and least read clause in the contract. Explicitly ask for the figure excluding exclusions, that is, the availability a customer actually lived through over twelve months, and compare it with the commercial percentage.
Neither, as a matter of principle. A large site brings pooling, a rich network ecosystem and written processes; a small site often brings someone who knows your rack and a responsiveness no process replaces. The grid above does not ask for size: it asks for traces. Some small operators answer the three questions better than listed groups.
Three is enough if the requested configuration is identical for all three and you ask each the same written questions. Beyond that, the time spent exceeds the gain, and the comparison degrades because you end up accepting answers in different formats.
The carriers present in a building are declared in public interconnection registries, consultable independently of the operator. It is the only part of the evaluation that rests on no declaration at all: every listing in our catalogue publishes that count and its source.
Written on 14 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers