🏛️ Issues and Advocacy Training
Courses › Data Centers: Types, Purposes and How They Work

Lesson 1 of 9 · about 14 minutes

What a Data Center Is and What It Does

You'll be able to

  • Describe a data center in plain words: computers, power, cooling, network, security
  • Trace a request from a phone to a server and back, naming each hop
  • Estimate network delay from distance using the speed of light in fiber
  • Put national data center electricity use in household terms with the arithmetic shown

A data center is a building full of computers that never get to go home. Everything else in the building, the transformers, the batteries, the diesel generators, the chillers, the fences and the badge readers, exists so those computers keep running every second of every day. When people say "the cloud," they mean these buildings. The cloud has a street address, a power bill and a property tax assessment.

Every one of them does the same jobs. It holds computers (servers, storage and network gear). It feeds them electricity that does not blink. It removes the heat they make, because nearly every watt that goes into a chip comes back out as heat. It connects them to the rest of the internet through fiber. And it keeps people who should not be there away from them. Size varies wildly, from a single room in a hospital basement to campuses that draw more power than a mid-size city. The jobs stay the same.

One tap, followed all the way

Say a resident of the invented town of Cedar Bluff opens a weather app at 6:40 a.m. Here is roughly what happens in the next quarter second.

Phone Cell towerand carrier Carrierbackbone Exchangepoint Data center load balancer, app server, database solid line: request out dashed line: answer back
A simplified path. Real traffic may cross several carriers and exchange points, and a cached answer may come from a server much closer than the main data center.
  1. Phone to tower. The phone sends a small request by radio to the nearest cell site. The carrier turns that radio signal into data on its own fiber network.
  2. Name lookup. The phone needs the numeric address for the weather company's server, so it asks a domain name system (DNS) server. That answer is usually cached somewhere close and comes back in a few milliseconds.
  3. Carrier to exchange. The carrier carries the request toward a place where it can hand traffic to the network that hosts the weather service. Those handoff buildings are called internet exchange points, and many of them sit inside colocation data centers.
  4. Into the data center. The request enters through a border router, passes a firewall, and lands on a load balancer, which picks one of hundreds of identical app servers that is not too busy.
  5. The work. The app server checks who you are, pulls your saved location from a database, and fetches the latest forecast from storage. If an AI feature writes your morning summary, the request may hop to a separate cluster of GPU servers.
  6. The trip home. The answer retraces the path, often by a different route, and the app draws the screen.

Most of that is invisible. The parts that show up on a county assessor's roll are steps 3, 4 and 5: the buildings where traffic is exchanged, and the buildings where the work gets done.

Why distance matters: the speed of light is slow enough to notice

Light in optical fiber moves at about two thirds of its speed in empty space, close to 200,000 kilometers per second. That works out to about 200 km per millisecond, one way. It sounds fast until you add it up.

Worked example: delay from distance alone

A user is 1,200 km by fiber route from the data center that hosts an app. Fiber routes follow highways and rail lines, so route distance runs longer than a straight line on a map.

  • One way: 1,200 km divided by 200 km per ms = 6 ms
  • Round trip: 6 ms times 2 = 12 ms
  • A web page that needs 10 round trips to load (connection setup, security handshake, several files) spends at least 10 times 12 = 120 ms just on distance.

Move the server to 100 km away and the same page spends 10 times 1 = 10 ms. Nothing about the server changed. Only the distance did.

This is why companies spread data centers around and why content delivery networks exist. A CDN keeps copies of popular files, such as video segments, images and app updates, on servers near large groups of users. Your Friday night movie probably streams from a cache a few hundred kilometers away. The company's main campus may be on the other side of the country. It is also why stock trading firms pay a premium to sit in the same building as an exchange's matching engine, where distance shrinks to the length of a cable.

Some work does not care about delay at all. Training an AI model runs for weeks inside one campus and talks to the outside world very little. That difference, delay-sensitive work near people and delay-tolerant work wherever power is cheap, shapes where buildings go. Lesson 8 returns to it.

How much electricity are we talking about?

The best U.S. estimate comes from the Lawrence Berkeley National Laboratory (LBNL) report released by the Department of Energy in December 2024. It put data center use at 176 terawatt-hours (TWh) in 2023, about 4.4 percent of all U.S. electricity, up from 58 TWh in 2014. Its 2028 scenarios range from 325 to 580 TWh, or 6.7 to 12 percent of U.S. use. The width of that range is the honest part. Nobody knows exactly how many AI chips will be bought, how hard they will be run, or how many announced projects will get built.

Globally, the International Energy Agency estimated about 415 TWh in 2024, roughly 1.5 percent of world electricity, rising in its base case to about 945 TWh by 2030.

Worked example: turning TWh into homes

The Energy Information Administration says the average U.S. residential customer bought about 10,791 kWh in 2022.

  • 176 TWh = 176,000,000,000 kWh
  • 176,000,000,000 divided by 10,791 = about 16.3 million homes' worth of yearly electricity

That comparison is useful and also a little slippery. Homes use most of their power in evening peaks and hot afternoons. A data center runs nearly flat around the clock. Matching total kWh says nothing about whether the grid can serve the load at 5 p.m. in August. Use the comparison for scale and say what it leaves out.

FigureValueSource
U.S. data center electricity, 201458 TWhDOE / LBNL, Dec. 2024
U.S. data center electricity, 2023176 TWh (4.4% of U.S.)DOE / LBNL, Dec. 2024
U.S. scenarios, 2028325 to 580 TWh (6.7% to 12%)DOE / LBNL, Dec. 2024
World data centers, 2024about 415 TWh (about 1.5%)IEA, Energy and AI
World base case, 2030about 945 TWhIEA, Energy and AI

Common mix-ups

  • Claim: The cloud is wireless, so it does not take up space. What the evidence shows: Wireless is only the last few hundred meters from your phone to a tower. Everything after that runs over fiber into buildings with very large electric services.
  • Claim: Every data center is an AI data center now. What the evidence shows: Most data centers still run ordinary business software, websites, email and storage. AI clusters are the fastest growing slice and the largest single buildings, which is why they get most of the attention.
  • Claim: One national number tells you what a local project will do. What the evidence shows: A national share of 4.4 percent can hide a county where one campus doubles the local utility's peak load. Always ask for the project's own megawatt request.

Reporting angle

When a data center is proposed in your coverage area, the first useful number is the electric load the developer asked the utility for, in megawatts. Ask for it in writing from the developer and the utility, and check whether it appears in a utility filing with the state public utilities commission. Ask what kind of work the building will do (cloud services, colocation for many tenants, or AI training), because that tells you how steady the load will be and how many people will work there. Then ask who the end user is. Many projects are filed under a shell company name, and the actual tenant is disclosed late or never.

Research angle

Check yourself

1. Which description of a data center is most accurate?

2. Light in optical fiber covers roughly how far in one millisecond?

3. According to the Lawrence Berkeley National Laboratory report released by DOE in December 2024, about what share of U.S. electricity did data centers use in 2023?

4. Why does a streaming video often load from a server in or near your own region instead of a company's main campus?

5. If a server is 1,500 km away by fiber route, what is the minimum round-trip delay from distance alone?

Terms in this lesson

Data center
A building or campus built to house networked computers and keep them powered, cooled, connected and secure without interruption.
Server
A computer that answers requests from other computers, such as sending a web page or running an AI model.
Latency
The delay between sending a request and getting the first part of the answer, usually measured in milliseconds.
Internet exchange point (IXP)
A place where many networks connect and hand traffic to each other directly.
Content delivery network (CDN)
A system of servers spread across many places that keeps copies of popular content close to users.
Terawatt-hour (TWh)
One billion kilowatt-hours. Used for the yearly electricity use of large sectors or whole countries.

Full glossary

Sources