Research

What a cold data centre costs you.

Most server rooms are kept colder than they need to be. We spent a month measuring what that habit actually buys you.

The short version

Cold rooms cost more than they save.

We ran our servers at four room temperatures across a month, from 18 °C up to 27 °C. The gap in power the machines drew between the coldest and the warmest room was smaller than the industry usually claims. The failure rate barely moved. So the cold room saves very little on the servers themselves. Meanwhile the cooling that keeps it cold runs up a bill of its own, every month, whether the machines need it or not.

How we ran it

One workload, four rooms.

The idea was to change one thing and hold everything else identical. Same machines, same work, same month.

Set four room temperatures

We held a test room at four points between 18 and 27 °C. Nothing else about the servers changed from one run to the next.
Ran the same work for a month

The machines carried a steady, realistic load the whole time. A month is long enough to see past a single good day or a bad one.
Watched power and failures

We measured how much power each machine drew and logged every fault. Then we compared the coldest room against the warmest.

The test at a glance

4 Room temperatures

1 month Continuous run

18 to 27 °C Range under test

Small Power gap, cold to warm

Steady Failure rate across the range

What the month showed

The cold buys less than you think.

The assumption is simple, and nearly everyone shares it. Colder servers draw less power and last longer, so a cold room pays for itself. Our month did not back that up.

Imagery in production

One of four temperature runs, each held steady across the month.

Between the coldest run and the warmest, the power our servers drew hardly changed. There was a gap, but it was small, far smaller than the figures quoted when someone is selling cooling. Stretched over a month, it came to very little.

Reliability told the same story. We logged every fault in every run. The warmer rooms did not produce more of them. The failure rate held flat from one end of the range to the other.

So where does the money go

Not into the servers. It goes into the cooling. Chilling a room to 18 °C and holding it there takes steady energy, and that bill arrives every month whether the servers need the cold or not. That is what a cold data centre really costs, and most of it guards against a problem our test could not find.

Under the hood

The questions behind the result.

Imagery in production

Why warmer is usually fine

Server parts are rated to run far hotter than most rooms ever get.

The parts inside a server are built to tolerate temperatures well above a comfortable office. A room at 27 °C sits nowhere near their limit. Cooling that room down to 18 °C does not make the components any happier. It just makes them cold, and cold is not the same as safe.

Imagery in production

Where the cooling bill comes from

The colder the target, the harder the cooling works, all day, every day.

Cooling is neither free nor occasional. To hold a room several degrees below the air outside, around the clock, the plant runs the whole time. Every extra degree of cold you ask for costs more energy to deliver. The servers pocket a small saving while the cooling spends a much larger one to hand it over.

Imagery in production

What this means for a small server room

You probably do not need the room as cold as it is.

If your server lives in a cupboard or a small back room, the same logic holds at your scale. Letting that room run a few degrees warmer is unlikely to shorten the machine's life or push up its power use, and it eases the load on whatever keeps it cool. We build our servers to be comfortable in an ordinary room for exactly this reason.

See the machines

“We would rather publish a number that costs us a sale than repeat one that sells the room cold. The research desk reports what the test showed, not what the market wants to hear.”

The Calyraen research desk

Questions

Asked and answered.

Does running warmer wear servers out faster?

Not in the range we tested. From 18 to 27 °C the failure rate stayed flat. The parts inside a server are built for far more heat than a warm office ever reaches.

So should I turn the cooling off?

No. Servers still need airflow and a sensible ceiling on temperature. The point is narrower: you can often let the room run a few degrees warmer than habit says, and save on cooling without paying for it in reliability.

Do these results apply to my setup?

The physics travels well, but your room is not our test room. Treat this as a reason to check your own numbers, not a fixed setting to copy. If you run our servers, talk to us and we will help you read your own power and temperature data.

Where can I read more?

This page is the plain version of the work. The research index keeps our write-ups, newest first.

Where to *next*.

Research index Our commitments The machines Talk to us

Keep reading

More from the research desk.

The full index, newest first. We add to it as the tests wrap up.

Back to the index Our commitments

What a cold data centre costs you.

Cold rooms cost more than they save.

Set four room temperatures

Ran the same work for a month

Watched power and failures

The test at a glance

So where does the money go

Why warmer is usually fine

Where the cooling bill comes from

What this means for a small server room

More from the research desk.