What is a data center?

Right now, as you read this sentence, your request flew through a windowless building packed with thousands of servers, past biometric security, along fibre under the ground — and came back in a fraction of a second. That is a data center. You almost never see them, yet without them no website, no app and no bank card works.

Let us break down how they are built — in plain language, without losing the essence.

What a data center is and why you do not keep a server at home

A data center (DC) is a building where companies servers live, plus all the infrastructure that keeps them from switching off: power, cooling, connectivity and security.

Can you put a server in the office under a desk? You can. Once. Until the first power cut, a summer overheat or a voltage spike. A data center is the same server, but with everything duplicated: two power feeds, UPS units, diesel generators, several air conditioners, several internet providers. The idea is simple: a failure of any single part must not take your services down.

A rough analogy: a home server is a boat with one motor. A data center is a plane with two engines that flies calmly on one.

A tour inside: what is behind the biometric door

A DC is not just servers. It is six systems that work to keep the servers from ever stopping.

• Power that never goes dark. The city lights flickered — but the DC did not even notice. For a second or two the load is held by a UPS (battery-backed power), and in that window the diesel generators spin up and carry on for days as long as there is fuel.
• Cooling. A server turns almost all its electricity into heat. Fail to remove it and the hardware cooks. So the halls are kept at ~20–25 °C: cold air is fed into cold aisles and hot air is blown out separately. For powerful AI servers air is no longer enough — they are cooled with liquid.
• Connectivity. A serious DC has several internet carriers coming in at once. An excavator cut one carriers fibre? Traffic instantly moved to another — you noticed nothing.
• Security. In layers: a fence, cameras, guards, and at the entrance a card + PIN + biometrics. There is even a mantrap: a small room with two doors that never open at the same time, so nobody slips in behind your back.
• Fire suppression without water. You cannot pour water on servers. So the fire is knocked down with gas (for example FM-200): it leaves no residue and does not damage the electronics.
• Monitoring. On top of all this sits software (DCIM) that watches temperature, load and power in real time and raises the alarm before anything breaks.

Did you know? Microsoft once sank a data center off the coast of Scotland — a sealed capsule with 855 servers sat on the seabed for about 2 years. Underwater the failure rate came out roughly 8 times lower than an identical setup on land: a stable temperature and the absence of oxygen, which ages hardware, did the work.

A short 50-year history

The idea of splitting one expensive computer among many was born back in 1972 at IBM. In the 1990s servers moved from machine halls into dedicated buildings — and at that same time, in 1993, Kenneth Brill founded the Uptime Institute and a single quality mark for DC reliability (the Tier system). Before that, everyone understood reliable in their own way.

Today history is being rewritten because of AI. A typical large DC eats 100–150 MW. New AI campuses aim for gigawatts — 10–40 times more. The biggest of those under construction approach one fifth of Manhattan by area. The AI race has turned data centers into the build of the century.

Data centers come in different kinds

• Enterprise — a company builds a DC for itself (banks, government).
• Colocation — you rent space for your own servers, while the building, power and security belong to the provider.
• Hyperscale — giants like Google and Meta: thousands of servers, megawatts of power.
• Edge — a small DC closer to people so things are faster (games, streaming).
• Cloud — the same data centers, you just take capacity as a service, without your own hardware.

Tier I–IV: how reliability is measured

Tier is the reliability level of a DC on the Uptime Institute scale, from I (basic) to IV (maximum). The higher it is, the more is duplicated. Here is the short picture:

The main difference is between Tier III and Tier IV (even IT people mix these up), and it is simple: in Tier III you can take any equipment out for planned maintenance without switching the servers off, but a sudden failure can still cause downtime; Tier IV survives even a sudden failure without stopping, because everything runs in two parallel, physically separated systems.

A car analogy: Tier III — you have a spare tyre in the trunk and change a wheel in the car park without trouble. Tier IV — a second car drives alongside, and if the first blows a tyre in motion you simply switch across without stopping. The second car is expensive — so Tier IV is overkill for most businesses, Tier III is enough. More detail: what a Tier III data center is.

An honest note on 99.982% These pretty uptime percentages travel across every website, but few admit it: the Uptime Institute removed them from the standard back in 2009. Today they are just convenient reference points for comparison, not a guarantee. If a provider promises Tier III equals a guaranteed 99.982%, that is marketing. Look at the real SLA.

Three myths worth throwing out:

• There is a Tier 5. No. The scale ends at IV. Tier 5 was invented by one company for advertising.
• Higher Tier is always better. Tier measures only redundancy, not speed, security or price. Higher does not always mean more necessary.
• Anyone can call themselves Tier III. Only the Uptime Institute certifies. Tier III on a website without a certificate is just words.

How certification works

So that not everyone slaps Tier III on themselves, the certificate is issued in three steps, in order:

• 1. Design. The drawings are checked before construction. Only intent on paper.
• 2. Constructed. They come and inspect the actually built DC. This is the key proof: the paperwork can be perfect while the build was done with mistakes. Ask specifically for "Constructed", not just "Design".
• 3. Operational. They assess whether the team can really hold the level (most incidents are human factors). Awards: Bronze / Silver / Gold.

Over the years the Uptime Institute has issued more than 4,000 such awards in 120+ countries — it is expensive, slow and not everyone passes. That is why the certificate is valued. (Do not confuse it with TIA-942 — a different, broader standard with its own Rated 1–4; the terms are not interchangeable.)

A few words in pro language

• PUE. How much electricity actually reaches the servers versus how much the building itself (air conditioning and so on) eats. 1.0 is ideal; the average is ~1.56; top DCs are ~1.1. The closer to 1, the more efficient.
• N+1 and 2N. A spare tyre in the trunk (N+1) versus a second car alongside (2N).
• Rack density. How many kilowatts one rack eats. It used to be ~12 kW; AI racks are 100+ kW. That is exactly why the move to liquid became necessary.
• SLA. A contract with an uptime guarantee and compensation if it is missed. Read the compensation clause, not just the pretty percentage.

Records and fun facts

The largest DC in the world depends on how you count: by area — China Telecom in Inner Mongolia, ~1 million m² (about 140 football pitches); by power — xAI Colossus in Memphis (hundreds of thousands of GPUs, gigawatt-scale power). And because of AI the record changes almost every month.

• Where is growth fastest? The world hub is Northern Virginia (+37% in a year), but almost no free capacity is left there. The main driver everywhere is AI.
• How much energy do they eat? All the worlds data centers in 2024 used ~1.5% of the planets electricity, and by 2030 it will be ~3% (IEA data). AI is half of that growth.
• ~99% of intercontinental internet travels not through satellites but along submarine cables on the ocean floor (about 570 of them).
• In Norway there is a data center inside a mountain below sea level — it is cooled with fjord water and powered by a hydro plant.

What it costs and how to choose

Building your own DC costs tens and hundreds of millions of dollars. So business does not build, it rents: colocation or simply a VPS or server with monthly billing. What to look at when choosing:

• Where it sits — closer to your users equals faster; data inside the country equals calmer with the law.
• Connectivity — several carriers, backup channels.
• Reliability — Tier III/IV, a real certificate (Constructed).
• SLA with compensation — not just a percentage, but a compensation clause.
• Headroom for growth — whether there is enough power per rack.

What about Azerbaijan

Where your server physically sits is about speed and about law. A server in Baku answers local users in single-digit milliseconds (not tens, like a server in Europe), and the data stays in the country.

Epro.io note — our data center. Epro.io equipment sits in Baku, in the Delta Telecom Tier III data center — certified for both the design and the constructed facility (Design + Constructed). Redundant power and cooling, physical security, 99.9% SLA and manat billing. For business in Azerbaijan this is exactly that right balance of Tier III. More: what Tier III is and why Epro.io.

FAQ