Keeping your server room or moving out: the six omissions that skew the numbers
Your table compares what your server room costs with what a hosting provider would charge. Six items are almost always missing from it, and they all lean the same way. Check whether yours are there.
Estimate my cost →The case we are using, with all its bearings #
Every figure on this page comes from the same example, computed by our tool. Here it is in full, so you can place yourself before reading on.
The company. About sixty people. Six racks in a technical room: a rack is a cabinet roughly 60 cm wide and 2 metres tall, in which servers are stacked. A UPS a few years old, a generator, an air-conditioning contract.
What the room costs in invoices: €113,070 a year, or €1,570 per rack per month. Electricity accounts for 56 % of it, rent for the room 12 %, maintenance contracts 23 %, and the rest is the sum set aside each year to replace equipment at end of life.
The bearing usually missing: at a hosting provider, the same rack costs about €375 a month, all in. Your room therefore costs around four times more per rack. Why, is omission 2.
What an hour of downtime costs: €800 in this example. That is deliberately conservative, and it is the one figure you must replace with your own: it decides everything else. The more an hour of downtime costs you, the wider every gap on this page becomes.
Omission 1: your team's time appears on no invoice #
Escorting a maintenance engineer, walking the room, reading meters, driving in at eleven at night because an alarm went off. That work occupies someone and produces no invoice, so it enters no budget.
On our room it comes to 138 hours a year, close to twenty working days. At a technician's fully loaded rate: €10,350 a year, against €113,070 of invoices. Nine per cent more cost, missing from the table.
Those hours break down as:
- 54 hours escorting maintenance visits. That number follows from your contracts: each triggers two visits a year, and each visit occupies someone for three hours.
- 48 hours of rounds and readings, four hours a month.
- 24 hours of out-of-hours call-outs, across six calls a year. We add no night or weekend premium, so this line is understated.
- 12 hours of patching, meaning network cabling inside the racks, plus adding and removing servers.
And this item does not fall to zero if you leave. Managing a provider takes time, and the patching stays yours since the servers do. €3,263 a year remains, a third of the total. So the real saving is €7,088 a year, not €10,350.
Omission 2: your room is sized for more than it holds #
€1,570 per rack per month against €375 at a provider: the gap has a single cause, and it has nothing to do with the quality of your installation.
A room is built as one piece. The cooling, the room, the UPS and the maintenance contract cost nearly the same whether you put six racks in it or twenty. Divide that by six and each rack carries a very heavy share of fixed costs.
A provider spreads the same equipment across hundreds of racks and bills you only for yours. Sharing what is expensive is the service they sell.
A half-empty room is therefore the worst case, and the most common one, since rooms are always sized for the growth to come.
Omission 3: the robustness claimed is not the one held #
Server rooms and data centres are graded on a four-level scale called Tier. What separates the levels is simple: at level II, maintaining a UPS or a cooling unit forces a service outage, because only one circuit exists. At level III, several exist, so maintenance happens without stopping anything. Hence the difference in expected downtime: 22.7 hours a year at level II, 1.6 hours at level III.
Those durations assume an installation maintained as designed. Two examples of what they do not cover:
- A generator only ever tested off-load. It is started with nothing connected: that proves the engine runs, and nothing more. A load test has it actually power the room, and it is the only one that proves it holds the power, that it cools, and that the transfer completes without a break.
- A UPS past seven years. It is no longer just the batteries wearing out: capacitors and fans age too, and the device meant to protect becomes a cause of failure.
On our level II room with a generator only tested off-load, expected downtime rises from 22.7 to 23.75 hours. One hour more, or €840 a year.
On a small room that gap stays small, and this deserves saying: level II promises little to begin with. It becomes heavy on an installation that promises a lot, which is precisely what a provider sells, and we costed it in what a claimed Tier level is really worth.
*Check on the calculation: the same room with a load-tested generator, a recent UPS and on-call cover in place comes out at 22.7 hours, exactly what its level announces. The gap comes only from maintenance.*
Omission 4: you are comparing a purchase with a subscription #
Staying means buying equipment that lasts fifteen years. Leaving means paying a monthly subscription. Adding both over five years and comparing totals means nothing, since one ties up money at once and the other spreads it.
The usual method turns the purchase into an equivalent rent: what you would pay each year if you rented that equipment instead of buying it, across its whole life. Accountants call it equivalent annual cost. Without that conversion, renting always wins, because it has no large sum up front.
With the three options on that common basis:
| What you do | Cost per year | What it covers |
|---|---|---|
| You change nothing | €111,847 | your current invoices, plus €19,000 of expected losses over the 23.75 hours of downtime |
| You bring the room up to standard | €99,058 | €39,000: partitioning the airflow so cold air no longer mixes with hot, which lightens the electricity bill, plus added capacity |
| You move to a hosting provider | €25,613 | the subscription, the move, and €1,280 of expected losses instead of €19,000 |
A ratio of more than four between staying and leaving looks suspect. It comes from two places, both visible in the table.
The room's fixed costs, explained in omission 2: you pay for a whole installation to hold six racks.
And failure risk, counted on both sides. 23.75 hours at €800 make €19,000 a year of expected losses; 1.6 hours make €1,280. The risk gap alone is close to €18,000, and it depends entirely on the figure you put on an hour of downtime. If your business copes with a day without servers, it collapses and the conclusion changes.
Omission 5: your equipment is still worth something at the end #
If you compare over five years, a UPS bought today still has half its life left at the end of the period. Ignoring that charges the "stay" option for equipment you have not used up.
We measured what this omission produces on its own, by rerunning the calculation without counting what remains: the annual cost of "bring the room up to standard" rises by €1,282, "move out" by €559, and the status quo does not move, since it buys nothing.
The order of the three amounts is logical: the more an option buys, the more the omission penalises it.
Omission 6: you are testing one assumption, not seven #
Your table names a winner, but it does not say whether that winner still holds when electricity rises, when you reason over seven years instead of five, or when borrowing costs more.
Yet that is the question that decides in the room: does the answer change if I got something wrong?
So we rerun the whole calculation across seven sets of assumptions. On our case, the same option wins in all seven. Had it lost in two of them, we would have said so, because that is what separates a solid answer from one that holds by luck.
What this calculation does not do, and when it says change nothing #
It does not know what an hour of downtime costs you, and that is the figure that decides everything. To estimate it, start from your annual revenue divided by the working hours of the year, then remove the share of activity that would carry on regardless. A company of sixty people often lands at several thousand euros an hour, well above the conservative €800 used here.
It does not replace a quote: our amounts are market estimates. Lining up two quotes against each other is a separate exercise, covered in comparing offers on an identical basis.
It does not predict your real availability. Expected downtime is an order of magnitude, not a promise, and only a test measures it.
And sometimes it concludes that nothing should change. A small, well-maintained room whose downtime costs little, with a team already on site for other reasons, often has no financial reason to move.
Take your own table and check one thing first: does it contain the six items on this page?
FAQ #
Because those twenty days exist and could serve another purpose. If the calculation ignores them, it hands the "stay" option free labour the "provider" option does not get. This is not about cutting posts, it is about comparing the same thing on both sides.
They follow from your installation: the number of visits comes from your maintenance contracts, the rounds from a monthly rhythm, the out-of-hours call-outs from a number of calls you can correct. Every line shows its arithmetic.
A metal cabinet roughly 60 cm wide and 2 metres tall, in which servers are stacked. A small firm often occupies two to six.
The number of circuits. At level II there is only one: maintaining a UPS forces a service outage. At level III there are several, so maintenance happens without stopping anything. Hence 22.7 hours of expected downtime a year on one side, 1.6 hours on the other.
Because it only proves the engine starts. A load test has it actually power the room, and it is the only one that proves it holds the power, cools properly, and transfers without a break.
What you would pay each year if you rented a piece of equipment instead of buying it, across its whole life. It is the only way to compare a fifteen-year purchase with a monthly subscription.
Divide your annual revenue by the working hours of the year, then remove the share of activity that would carry on regardless. The €800 on this page is deliberately conservative: at many companies of this size the real figure is several times higher, which widens every gap further.
No. On a well-maintained room whose downtime costs little, staying is often right, and the calculation says so. It also refuses several reliability spends when they cost more than they return.
Written on 16 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers