A provider claims Tier III: how many hours of downtime does that really give you?
It depends on what has been tested, and you can check that yourself on your visit. Six questions are enough.
Estimate my cost →What you are being told, and what it is worth #
Tier is a four-level scale grading an installation by the redundancy of its power and cooling circuits. In concrete terms, redundancy means two separate power feeds into the building, two chains of equipment inside, and right down to two power strips in your rack, fed separately. If one fails, the other holds.
Each level carries an expected downtime over a year:
| Level | Expected downtime per year | What it means for you |
|---|---|---|
| Tier I | 28.8 h | a single path: any maintenance stops your servers |
| Tier II | 22.7 h | duplicated equipment, but still a single path |
| Tier III | 1.6 h | the operator can maintain any component without stopping you |
| Tier IV | 0.4 h | two paths active at once: a failure goes unnoticed |
⚠️ Watch out for two wordings that look alike. "Designed to Tier III criteria" and "Tier III certified" are not the same thing. The first is the operator's own statement about their own design. The second implies that a third party, the Uptime Institute, which created this classification, has examined the site. Ask which of the two you are being sold, and exactly what the document covers: a plan validated on paper, or an installation examined once built.
What those durations assume, and nobody says #
These figures do not describe a perfect installation. They describe an installation of that design maintained as it was conceived.
In other words, they already contain failures: a generator that starts, a transfer that completes, a UPS that holds through that transfer. All of it is inside the figure.
What they do not contain are gaps in maintenance: backup equipment nobody has checked, equipment left to age, and the time an engineer takes to arrive when nobody is on call.
The three practices that widen the gap #
Before reading them, two words about the two devices that keep coming up, because they do different jobs.
The UPS takes over instantly when power fails, but holds only a few minutes on its batteries. The generator takes tens of seconds to start, then can power the site for hours. The UPS exists to cover exactly that starting delay. If either fails to show up, the chain breaks.
Each of the three practices below works through a precise mechanism, not a penalty of principle.
- The generator never tested under load. It is started with nothing connected: that proves the engine runs, and nothing more. A load test has it power the site for real, and it is the only one that proves it holds the power, that it cools, and that the transfer completes without a break. This is the heaviest of the three.
- The UPS beyond seven years. It is no longer just the batteries wearing out. Capacitors, the components that smooth the current, dry out over time, and fans clog up. The device meant to cover the generator's start becomes a cause of outage itself.
- Cooling with no maintenance contract. A room that heats up stops, however perfectly powered. It is the classic omission where all the attention went to power.
What this gives, and what it costs you #
A site claiming Tier III, with a generator never load-tested and a UPS over seven years old.
Downtime announced by the level: 1.6 hours a year. The two maintenance gaps above each add their share, and the expected total rises to 3.31 hours. That is 1.71 hours more, a little over double. This surplus does not replace the 1.6 hours announced, it adds to them.
That surplus turns into euros at your end, and the amount depends entirely on what an hour of downtime costs you:
| If an hour of downtime costs you | What the gap costs you per year |
|---|---|
| €400 | €684 |
| €800 | €1,368 |
| €2,000 | €3,420 |
| €5,000 | €8,550 |
To estimate your own figure: divide your annual revenue by the working hours of the year, then remove the share of activity that would carry on regardless.
That amount is modest next to a rack's rent, but you get it back by asking six questions during a visit you will make anyway.
The six questions to ask on a site visit #
They fit on one page. The third column is the most useful: it says what a reassuring answer looks like, as opposed to one that dodges.
| The question | Why it matters | What a good answer looks like |
|---|---|---|
| When was the generator last load-tested, and may I see the report? | an off-load test proves neither power capability, nor cooling, nor a break-free transfer | a document less than a year old stating the load applied (in kW or per cent), the duration of the test, and who signed it. "We run it every month" with no document is worth nothing |
| How old are the UPS units, and when were the batteries changed? | past seven years, capacitors and fans age too, not only batteries | a precise age, a battery replacement date, and periodic discharge testing. Hesitation about the age is already an answer |
| Is cooling duplicated, and under a maintenance contract? | a room that heats up stops, however perfectly powered | the number of units and how many can fail without consequence, plus a named contract. "There is air conditioning" is not an answer |
| Who responds at three in the morning, and how fast are they on site? | a failure's duration includes arrival time, which depends on organisation, not equipment | staff present around the clock, or on-call cover with a written response time. A phone number with no committed delay is not one |
| Is there any point on the power circuit that, alone, stops everything? | a known single path caps the real level, whatever the design claims | an electrical diagram shown, not merely asserted. Look for the panel or component everything passes through. An operator who produces the drawing unprompted is already a good sign |
| Will my rack have two power strips on two separate circuits? | all the building's redundancy is useless if your rack itself is fed from one side only | two strips, two distinct circuits up to the rack, and confirmation that your dual-supply servers must be plugged into both |
And while you are being shown around, look. Documents say what was written, the room says what is lived. Four signs that prove nothing on their own, but tell you where to press: dust on a generator or clogged ventilation grilles, an alarm light amber or red that nobody comments on, an empty on-call office on a weekday, and cables added as an afterthought hanging outside the trays. Conversely, a site where drawings are posted, last-service dates are written on the machines and aisles are clear has habits that show.
What fixing it costs the operator, and why that concerns you #
These costs are not yours: the operator pays them. They still matter to you, because they say whether he has an excuse not to.
So we restate them per rack, at the scale of the site used as our example:
- The annual load test of the generator recovers 1.32 hours a year for €3,000: that is €13 per rack per month. Next to nothing, in other words. An operator who skips it has no economic argument, and you can tell him so with that figure.
- Replacing an end-of-life UPS recovers 0.39 hours, for €39 per rack per month once the cost is spread over the appliance's nine-year life. Heavier, which explains why an operator may defer it. Your call whether the site you are visiting has deferred it once too often.
A useful order of magnitude for the conversation: the first action costs less than a meal, per rack per month, and it is the one that returns the most hours. Ask about it first, and its answer will tell you the most about the house.
What the contract promises is not what the installation holds #
You will see an availability percentage in the commercial proposal. It does not say the same thing as the Tier level, and the comparison is surprising:
| What is written | What it actually allows per year |
|---|---|
| 99.9 % availability | 8 h 46 min of downtime, with no clause breached |
| 99.99 % | 53 minutes |
| 99.999 % | 5 minutes |
| *for comparison*, a Tier III design | *1.6 hours* |
A 99.9 % commitment is therefore more permissive than the technical promise of level III, while sounding impressive. And the compensation for breaching it almost never relates to what the outage cost you: it is generally computed on your subscription, not on your loss.
Ask the salesperson two things: over what period the percentage is measured, and whether the compensation is a credit note or a payment. The detail is in what a service commitment is really worth.
How we calculate, for those interested in the method #
The hours quoted above come from a calculation, and here it is, so that you can contradict us.
We never count twice. Adding failures on top of the duration a level announces would double-count, since that duration already contains them. From each mechanism we keep only what exceeds the maintenance the design assumes. The check is public: the same site, with a load-tested generator, a recent UPS and on-call cover in place, comes out at 1.6 hours a year, exactly its level's figure, and no correction is proposed.
A spot everything passes through does not downgrade the level, it caps it. A known single power path does not make Tier III unlikely, it makes it false, since the definition of level III is that several paths exist. We then compute no probability: the site cannot be credited beyond level II.
We do not demote a site by a level. Between 1.6 hours and 22.7 hours lies a factor of fourteen: announcing 22.7 hours to a site expecting 3.31 would be a worse error than the one being corrected. We give the duration in hours, and we situate.
And we do not fault a design for being what it is. A Tier I with no generator takes no penalty for that absence: it is its definition, and its 28.8 hours already contain it.
What this guide does not cover #
It does not predict the real availability of the site you are visiting. Expected downtime is an order of magnitude computed from mechanisms, not a measurement. The only measure that counts is a test, which is precisely the first of the six questions.
Nor does it cover the rest of the visit: access security, working conditions in the room, available connectivity. Our guide on assessing a data centre and its operator covers those.
And if you are still weighing keeping your servers in-house against moving them to a provider, the full comparison is in the six omissions that skew the numbers.
FAQ #
No. It is the expected duration for an installation of that design maintained as it was conceived. It is neither a contractual guarantee nor a measurement: it is an order of magnitude.
The first is the operator's own statement about their own design. The second implies examination by a third party, the Uptime Institute, which created this classification. Ask which of the two you are being sold, and what the document covers.
The UPS takes over instantly but holds only a few minutes on its batteries. The generator takes tens of seconds to start, then powers the site for hours. The UPS exists to cover that starting delay.
Because it only proves the engine starts. It tests neither real power capability, nor the generator's cooling, nor a break-free transfer. Yet those three decide the day the outage comes.
A document less than a year old stating the load applied in kilowatts or per cent, the duration of the test, and the signature of whoever ran it. An oral assurance with no document is worth nothing.
Because all the building's redundancy stops at your door if your rack is fed from one side only. Your dual-supply servers must be plugged into two separate circuits for that redundancy to reach you.
It allows 8 h 46 min of downtime a year with no breach of contract, more than five times what a Tier III design promises. Ask over what period it is measured, and whether the compensation is a credit note or a payment.
Ask the six questions on this page, then enter the answers into our assessment tool. It returns the gap in hours and prices it at the downtime cost you declare.
Written on 16 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers