From what a data center is to how a gigawatt campus gets built.

A data center is a building whose whole job is to keep computers running around the clock: fed with power, kept cool, kept secure and connected to the internet.

Every time you stream a show, send a message, search the web or ask an AI chatbot a question, your request travels to a data center. "The cloud" is not in the sky. It is millions of computers called servers, stacked in metal cabinets called racks, inside buildings like these all over the country.

Why put computers in a special building?

• Power that never stops. Servers need steady electricity. A data center has backup batteries and generators so a storm does not knock your bank offline.
• Cooling. Computers turn almost all the electricity they use into heat. Packed together, they would overheat in minutes without industrial cooling.
• Connectivity. Many fiber lines from different carriers meet in one place, so traffic is fast and has backup routes.
• Security and scale. Guards, cameras and badge access protect the equipment, and running thousands of servers together is far cheaper than each company running its own closet.

Not all data centers are the same. Who owns it, who uses it and how big it is change everything from the design to the headlines.

Data centers are measured by how much power they can deliver to computers, not by floor space. Getting a feel for the units makes every announcement easier to read.

• Kilowatt (kW): 1,000 watts. A home's average draw is about 1.2 kW. A traditional server rack draws 5 to 15 kW.
• Megawatt (MW): 1,000 kW. Running all year, 1 MW uses about as much electricity as 800 U.S. homes.
• Gigawatt (GW): 1,000 MW, roughly the output of one large nuclear reactor. Several U.S. AI campuses now plan for 1 to 5 GW.
• Megawatt-hour (MWh): energy, not power. 1 MW running for one hour is 1 MWh. Your electric bill counts kWh.

From the outside, most data centers look like big, windowless warehouses. Inside, every room supports the data hall where the servers live.

Tap or hover over each part of this cutaway to see what it does.

Read a data center spec sheet and you'll see the building split into "white space" and "gray space". Knowing the difference explains why campuses are so big.

• White space is the data hall floor where racks stand. It's what tenants lease and what capacity in MW ultimately serves.
• Gray space is everything that supports it: electrical rooms, UPS and battery rooms, mechanical plant, generator yards, meet-me rooms, loading docks and offices. In modern facilities gray space is often as large as, or larger than, the white space.
• Critical IT load is the power available to the racks, the number in most announcements. The utility connection is bigger, because cooling and losses come on top.
• Powered shell vs. turnkey: a developer can deliver a "powered shell" (building, power and fiber) for the tenant to fit out, or a turnkey hall with cooling and electrical gear already installed.

Anatomy of a campus

A campus is several buildings on one site sharing a substation, fiber routes, roads, water connections and security. Each building is usually built in phases, so a "1 GW campus" may open its first 100 to 200 MW years before the last building is finished.

Electricity takes a long path from the grid to a chip, and every step is built so that a single failure never turns the servers off.

What happens when the grid goes down

1. 0 seconds: the utility feed drops. The UPS batteries take over instantly, so servers never notice.
2. About 10 seconds: the diesel or gas generators start and reach full speed.
3. Seconds later: the automatic transfer switch moves the load onto the generators, and the batteries begin recharging.
4. When the grid is back: the switch moves the load back and the generators cool down. Sites keep fuel on hand for 24 to 72 hours or more.

The biggest constraint on new data centers in America is not money or land. It is getting enough electricity, fast enough.

How a campus gets power

• Grid interconnection: the developer applies to the utility and regional grid operator (such as PJM in the Mid-Atlantic, ERCOT in Texas or MISO in the Midwest). Studies can take years, and big loads may need new transmission lines and substations.
• Behind-the-meter generation: building gas turbines, fuel cells or solar plus batteries on site to start sooner or avoid grid limits. Several Texas and Wyoming projects take this route.
• Long-term power deals: hyperscalers sign power purchase agreements for new wind, solar, geothermal and nuclear, including restarts of retired reactors and future small modular reactors (SMRs).
• Flexibility: some campuses agree to cut load during grid emergencies in exchange for a faster connection.

Every watt that goes into a server comes out as heat. Cooling is about moving that heat out of the building as efficiently as possible.

From air to liquid

Air works well up to about 20 to 30 kW per rack. Beyond that, fans cannot move enough air, so dense AI racks bring liquid right to the chips. Direct-to-chip cooling pumps coolant through cold plates mounted on GPUs. Immersion dunks whole servers in a non-conductive fluid. The heat then goes to chillers, cooling towers or dry coolers outside.

Two numbers come up in nearly every sustainability claim and local debate. Here is how to read them.

• PUE (power usage effectiveness) = total facility power ÷ IT power. A PUE of 1.0 would mean every watt goes to computing. The industry average is around 1.5 to 1.6; the best hyperscale sites report about 1.1.
• WUE (water usage effectiveness) = liters of water consumed on site ÷ kWh of IT energy. It ranges from near 0 for dry-cooled sites to around 2 L/kWh for evaporative cooling in hot climates.
• CUE (carbon usage effectiveness) = kg of CO₂ ÷ kWh of IT energy, driven mostly by how clean the local grid is.

Data has to move between servers inside the building and out to the internet. Modern data centers use a spine-leaf design so any server can reach any other in a few hops.

• Carriers and the meet-me room: fiber from several providers enters through separate routes so a backhoe cannot cut the building off.
• North-south traffic flows between users and the data center. East-west traffic flows between servers inside it, and it is now the larger share.
• AI clusters add a separate, very fast "back-end" network (400 to 800 Gb/s per GPU) so thousands of GPUs can work as one machine.

Getting data in and out fast

• Interconnection: networks, clouds and enterprises connect to each other inside colocation buildings through cross-connects and internet exchange points. Places like Ashburn, Virginia grew because so many networks already meet there.
• Latency follows distance: light in fiber travels roughly 200 km per millisecond, so every 100 km adds about 1 ms of round-trip time. That is why inference and trading workloads want to be near users and exchanges, while training can go wherever power is cheap.
• Diverse routes: at least two physically separate fiber paths enter a well-built site, so a single cut never isolates it.
• Optics: modern switches connect with pluggable optical transceivers running 400 to 800 Gb/s per port; an AI cluster can need tens of thousands of them.

A rack is a standard steel cabinet about 2 feet wide and 7 feet tall. Its height is measured in rack units (U), each 1.75 inches. Most racks are 42U to 48U.

• Top-of-rack (ToR) switch: connects every server in the rack to the data center network.
• Servers: CPUs, memory and local drives. A common 1U or 2U server draws 300 to 1,000 W.
• GPU servers: 4U to 8U boxes with 8 accelerators each, drawing 6 to 12 kW apiece.
• Rack PDUs: the power strips, usually two (A and B feeds) so either can fail.
• Blanking panels: cover empty slots so hot air cannot loop back to the front.

Three kinds of compute in a hall

Because density jumps so sharply with AI, operators now plan halls by power per rack first and floor area second. A building designed around 10 kW racks can't simply take 130 kW racks without new power distribution, liquid cooling and stronger floors.

Data centers are designed around a simple question: what happens when a part fails, or needs maintenance? The answer is spare capacity, described with N notation.

• N: exactly enough equipment for the load. Any failure means an outage.
• N+1: one spare unit. One failure is fine.
• 2N: two complete, independent systems. An entire path can fail.

Uptime Institute tiers

A data center protects what's inside with layers: every layer someone passes through has its own check, so one mistake never opens the whole site.

Protecting against fire and leaks

• Very early smoke detection (VESDA) samples the air continuously and can spot overheating gear long before visible smoke.
• Pre-action sprinklers keep pipes dry until two separate signals confirm a fire, so a damaged pipe can't flood the hall. Some rooms use clean-agent gas instead of water.
• Leak detection cables run under floors and along liquid-cooling pipes.

The monitoring stack

Thousands of sensors report into software that operators watch around the clock. Each system has a specific job:

• Environmental limits: many operators follow ASHRAE guidance, which recommends server inlet air of about 18 to 27 °C.
• Operational technology security: the networks that run cooling and power are kept separate from customer and office networks, because a cyberattack on them could cause a physical outage.
• Change control: work follows written procedures: methods of procedure (MOPs) for planned work, standard procedures (SOPs) for routine tasks and emergency procedures (EOPs) for failures. Mistakes during maintenance and changes are among the most common causes of outages.
• Testing: generators are run under load, infrared scans look for hot electrical connections, and failover is rehearsed.
• Audits: customers ask for third-party reports such as SOC 2 and ISO 27001 before moving in.

A GPU is a chip with thousands of small cores that do math in parallel, which is exactly what training and running AI models needs. Thousands of them are wired together to act as one machine.

Training vs. inference

• Training builds a model by running huge datasets through tens of thousands of GPUs for weeks. It wants one giant, tightly networked cluster and lots of cheap, steady power, so it can sit far from cities.
• Inference runs the finished model to answer questions. Each request is small, but there are billions of them, and users notice delay, so inference sits closer to people.

Training and running large AI models changed data center design more in three years than the previous fifteen. The reason is the GPU.

• Density: a classic rack draws 8 to 15 kW. A rack-scale AI system holding 72 GPUs draws about 120 to 140 kW, roughly ten times more in the same footprint.
• Liquid cooling becomes mandatory, and buildings need far more piping, pumps and heat rejection.
• One giant computer: training clusters link tens or hundreds of thousands of GPUs with a dedicated network, so they want to sit close together on one campus with a huge, steady power supply.
• Training vs. inference: training builds a model and can sit in remote places with cheap power. Inference answers users' questions and benefits from being closer to population centers.

An announcement is the start, not the finish. Here is what happens between a press release and servers running, and how long each part typically takes.

What developers look for in a site

• Power: hundreds of megawatts available soon, ideally near high-voltage lines.
• Land: flat, outside flood zones, typically 100 to 1,000+ acres for a campus.
• Fiber: at least two diverse long-haul routes.
• Water and climate: water supply for evaporative cooling or a cool climate for free cooling.
• Incentives and permitting: sales tax exemptions on equipment, property tax abatements, and a predictable local approval process.

The money

Building costs commonly run around $10 million or more per megawatt for the shell, power and cooling, before servers. AI hardware can cost several times that again. That is why gigawatt campuses come with headlines quoting tens of billions of dollars.

Data centers bring tax revenue and construction jobs, and they also raise real questions about power bills, water, noise and land. Most of the fights in the news are about balancing these.

Local governments respond with zoning rules, setback and noise limits, special electricity rates, water agreements and, in some places, temporary moratoriums while they write new rules.