What Is a Data Center? Power, Cooling and PUE Explained
A data center is a building that holds many computers, called servers, together with equipment for power, cooling and networking. "The cloud" and AI services actually run inside buildings like these. The building also needs large systems that deliver power without a break and get rid of the heat the servers give off.
What you'll learn
- What sits inside a data center (racks, servers, storage and switches)
- The path electricity takes, and how a UPS covers a power outage
- How heat is removed, and why liquid cooling is spreading for AI
- How to read PUE, the efficiency number, and how much of the world's electricity data centers use
- How spare equipment keeps things running, and the chips that work inside
Prerequisites: none. It helps to know that "the cloud means renting another company's computers" (see "What Is Cloud Computing?").
In everyday life
When you ask an AI chatbot a question, an answer comes back in a few seconds. The phone in your hand is not what writes that answer. Many computers in a distant building do. Video streaming, photo storage and web search work the same way.
A request from your phone travels over the internet to that building (to see how, read "How Is Data Sent Over the Internet?"). This article looks inside the building at the other end.
A data center is "a building for computers"
A computer that sends back data or does calculations when asked is called a "server." A data center is a building made to hold thousands, even tens of thousands, of servers in one place. It is built not for people to work in, but for machines to work well.
Machines mainly need three things: power that never cuts out, cooling to remove heat, and a fast network connection to the outside. Providing these in one place is cheaper and more reliable than looking after each machine separately.
What's inside: racks, servers, storage and switches
Inside, you see row after row of tall metal shelves. Each shelf is called a "rack." Racks follow standard sizes, such as width, so machines from many companies fit in the same rack.
A rack holds many servers shaped like thin boxes, stacked like drawers. Inside, a server is much like a PC. It has a CPU that does the calculations, memory that holds data being worked on, and a device that keeps data stored.
Some machines specialize in holding data. These are called "storage." Inside are many SSDs, which store data electrically, or hard disks, which record data magnetically on spinning platters.
Other machines connect servers to each other and to outside lines, and send arriving data to the right server. This machine is called a "switch," and each rack often has one. For how it differs from a home router, see the table in "Hub vs Switch vs Router."
Outside the server rooms are rooms for equipment that receives and distributes power, batteries and generators for outages, and cooling machines. This "equipment that keeps the servers running" also takes up a large share of the building.
The path of electricity: from the grid to the chip
A data center gets its electricity from the utility's power grid. Grid power runs at a very high voltage so it can travel long distances. So the power first passes through a device that lowers the voltage. This device is called a transformer.
Next comes a device that guards against outages. It is connected to batteries, and the moment the power fails, it keeps supplying power from them. This device is called a "UPS" (uninterruptible power supply). Even a very short break in power can stop a server, so a UPS is essential.
Batteries cannot last through a long outage, so data centers also keep emergency generators that run on fuel. A generator takes a little time to start and deliver steady power. The UPS batteries bridge that gap, and then the generator takes over.
Inside a UPS, parts that change the form of electricity are at work. Power from a wall outlet is "alternating current" (AC), whose direction keeps reversing. A common design first turns it into "direct current" (DC), which flows one way, to connect the batteries. It then turns the DC back into AC and sends it on. Power semiconductors do this conversion. To learn how they work, read "What Is a Power Semiconductor?"
Power leaving the UPS is split up and sent to each rack. The device that does this is called a "PDU" (power distribution unit). Finally, the power supply inside the server turns AC into DC and lowers the voltage (for more, see "AC vs DC Explained").
Chips such as CPUs run on a low voltage of around 1 volt. So the DC voltage is lowered one more step on the circuit board. The circuit that does this is called a DC-DC converter. It sits right next to the chip and feeds it a large current.
Each conversion loses some of the electricity as heat. To cut these losses, Google published a design in 2016 that delivers 48-volt DC inside the rack. Google said it had used this since 2010, with lower losses than delivering power at 12 volts.
Getting rid of heat: cooling with air
After a server uses electricity to do its calculations, almost all of that electricity ends up as heat. If the heat is not removed all the time, the chips get too hot and may stop or break.
Most data centers cool with air. A server draws in cool air at the front and blows warm air out the back. When cool air and hot air mix, the cooling system wastes more energy.
So the rows of racks are lined up in the same pattern. When the fronts of two rows face each other, only cool air flows in the aisle between them. Hot air collects in the aisle between the backs of two rows. This layout is called "hot aisle / cold aisle." The U.S. Department of Energy's design guide presents it as a basic way to manage airflow.
The collected hot air is cooled by cooling units and sent back to the cold aisles. In cold seasons or cold regions, outside air can also be used to cut the electricity spent on cooling.
What AI changed: far more power per rack
A rough guide to the heat a chip gives off under a standard heavy workload is called thermal design power (TDP). Cooling systems are designed with this value in mind. NVIDIA's H100, a GPU for AI, has a maximum of 700 watts in its high-performance version.
NVIDIA's documents say that the "DGX H100," a server with eight of these GPUs, uses up to 10.2 kilowatts. One kilowatt is 1,000 watts. Meanwhile, in Uptime Institute's 2024 survey, the average rack at responding facilities used less than 8 kilowatts. So a single AI server can use more power than an average whole rack.
There are now even products that build a whole rack as one large computer. NVIDIA's GB200 NVL72 puts 72 GPUs and 36 CPUs in a single rack. HPE's product documents say one such rack uses 132 kilowatts. That is more than 16 times the average (under 8 kilowatts).
Still, these are a few cutting-edge examples. Uptime Institute's 2025 survey found that, overall, racks of 10–30 kilowatts were only starting to become more common. Facilities with racks above 30 kilowatts are still few.
To stay within the limits of the power and cooling systems, operators sometimes set a ceiling on server power. This is called power capping. When heat builds up beyond what even that can handle, the answer is liquid cooling, covered next.
Liquid cooling: carrying heat away with water or other liquids
Air is not very good at carrying heat. For the same volume, water can carry far more heat than air. So a method is spreading that presses a metal plate with liquid flowing through it right on top of the chip. This is called "liquid cooling."
The heat picked up by the plate travels with the liquid through pipes out of the rack. There it passes to a separate water loop and is finally released outside the building. HPE's documents say that of the GB200 NVL72's 132 kilowatts, 115 kilowatts are liquid-cooled and the rest is air-cooled.
Another method sinks the whole server in a special liquid that does not conduct electricity. With every method, there is new work: looking after the pipes and the liquid.
PUE: how much of the power reaches the computers?
A data center's electricity goes not only to servers, but also to cooling and to losses in power conversion. Divide the whole building's power by the power used by IT equipment, such as computers. The result is called power usage effectiveness (PUE). If IT equipment uses 1, PUE shows how much the whole facility uses.
For example, a PUE of 1.5 means that for every 1 unit used by IT equipment, an extra 0.5 goes to cooling and other uses. A PUE of 1.0 means zero extra: all the power reaches the IT equipment, which is the ideal. In practice, cooling and conversion losses can never be zero, so PUE is above 1.0.
In Uptime Institute's global surveys, the average for responding facilities was 1.56 in 2024 and 1.54 in 2025. That is not the average of every facility in the world, but it has stayed nearly flat for several years. The company says this is because many buildings are old, and some climates make it hard to cut cooling. Large operators do better: for example, Google reports 1.09 across its large-scale data centers in 2024.
How much of the world's electricity do they use?
Large amounts of electricity are measured in TWh (terawatt-hours). One TWh is 1 billion kilowatt-hours. A 2025 report by the International Energy Agency (IEA) says the world's data centers used about 415 TWh in 2024. That is about 1.5% of the world's electricity.
In an April 2026 report, the IEA estimates that this rose to about 485 TWh in 2025. It expects about 950 TWh in 2030, roughly double the 2025 figure. By then, data centers would use about 3% of the world's electricity. AI servers are a major reason for the growth.
Keep in mind that this is a projection. It depends on whether chips, power plants and power lines can be built in time. The IEA itself says these supply limits will hold back growth for now.
Keeping things running: having spares
Data centers are designed on the idea that every machine will fail someday. So the key is to remove any spot where one failure stops everything. Such a weak spot is called a single point of failure.
A common idea is "N+1." If you need N machines, you install one more. For example, if four cooling units are enough, you install five. If one breaks or is shut down for maintenance, the other four keep the cooling going.
For even more protection, you can provide two complete sets of everything you need. This is called "2N." Even if one whole set stops, the other keeps things running. The cost is higher.
There are also ratings for how well a building is protected. The best known is Uptime Institute's four-level "Tier" system. A tier is a level. Tier I is a basic design with no spares, so everything is shut down for maintenance. Tier II has spare equipment, such as UPS and cooling units. Tier III also doubles the power and cooling paths, so any equipment can be maintained without stopping the service. Tier IV can also survive one machine failing or one path being cut without stopping.
Cloud providers build several such buildings in the same region. A group of buildings whose power and cooling are independent of each other is called an availability zone. This lets work continue in another group even if a whole building stops.
The chips that work inside a data center
A data center is also a huge collection of semiconductor chips. Chips handle not only computing, but also memory, networking and power conversion.
The part that connects a server to the network is called a NIC (network interface card). It combines the port where the cable plugs in with a chip that handles sending and receiving. In AI computing, thousands of GPUs exchange data with each other. So network speed now affects how fast the computing goes.
At the heart of a switch is a special chip that does nothing but route data. A chip built for a single purpose like this is called an ASIC. It does its fixed job much faster, and with less power, than a general-purpose CPU. The table below lists the main chips.
| Chip | Where it works and what it does | Learn more |
|---|---|---|
| CPU | At the center of a server, computes and controls as instructed | What Is a CPU? |
| GPU and AI chips | Do huge numbers of the same kind of calculation at once | Why GPUs for AI? |
| DRAM and HBM | Memory for data being worked on. HBM is stacked and placed right next to the GPU | What Is HBM? |
| Flash memory (SSD) | Stores data so it stays even when the power is off | How SSDs work |
| Networking chips | Handle the server's network port (NIC) and route data inside switches | What Is an ASIC? |
| Power semiconductors | Change the form and voltage of electricity in UPS units and power supplies | What Is a Power Semiconductor? |
Common myths
Myth: Cloud data is somewhere "up in the sky"
In fact, the data is stored on devices in racks, in data centers in real countries. Most of the traffic also travels through optical fiber under the ground or under the sea. The myth comes easily from the name "cloud."
Myth: Data centers already use a large share of the world's electricity
In fact, the IEA estimates about 1.5% of the world's electricity in 2024, and about 3% even in its 2030 projection. But in regions where data centers cluster, the share is much larger and strains the power grid. With so much news about AI's power use, the global share can look larger than it is.
Myth: A data center with a lower PUE uses less electricity
In fact, PUE is a ratio: "total power ÷ IT equipment power." It tells you that little is wasted on cooling and other overhead, but not how much power is used in total. Even with a low PUE, a building full of servers uses a lot of electricity. The myth mixes up "efficiency" and "amount."
By the numbers
| Item | Number | Notes |
|---|---|---|
| Average power per rack | Under 8 kW | Responses to Uptime Institute's survey (2024) |
| Maximum TDP of the H100 | 700 W | NVIDIA specifications (high-performance version) |
| DGX H100 (server with 8 GPUs) | Up to 10.2 kW | NVIDIA documents |
| GB200 NVL72 (one rack) | 132 kW | HPE product documents; 115 kW of it liquid-cooled |
| PUE (survey average) | 1.56 → 1.54 | Uptime Institute global survey (2024 → 2025) |
| Electricity used by the world's data centers | About 415 TWh | About 1.5% of the world's electricity (IEA, 2024) |
| The same, in 2025 and 2030 | About 485 TWh → about 950 TWh | 2030 projected at about 3% of the world's electricity (IEA, April 2026) |
Numbers for AI products grow with each new generation. The figures above are as of each year shown.
Learn more
- Try the math: the availability calculator (how much downtime drops when you use two machines)
- Try the math: the junction temperature calculator (find a chip's temperature from its power and how easily it is cooled)
- See the differences in a table: SSD vs HDD, CPU vs GPU vs NPU
- Terms for the next step: RAID (protecting data with several disks), load balancer (spreading work across servers), thermal resistance (how hard it is for heat to escape)
- Related articles: "Memory vs Storage: What's the Difference?" and "How Do Large Language Models Work?"