The state of data centers

An AI data center is a warehouse full of computers. Building one takes about two years. Getting it plugged in can take longer. This is what goes inside, how much power and water it uses, who pays for it, and why the plug is the hard part.

Choosing where to build a data center

Builders used to ask two questions about a piece of land: is fiber-optic cable nearby, and are customers nearby? Now the first question is how soon they can get power. A 2026 industry outlook ranked "speed to power" first, ahead of whether the town wants them and how far away the users are.

Where a model learns, and where it answers

An AI data center does two jobs: training a model, and answering questions. Training can happen anywhere the power is, like west Texas or North Dakota. Answering has to be nearer to people, but less near than you might think.

The best-known campus being built today is Stargate in Abilene, Texas. It is planned as eight buildings covering about four million square feet.

Does the building that answers me have to be in my city?

No. The International Energy Agency says a chatbot can sit 500 to 5,000 kilometres from its users, because the trip adds only hundredths of a second to an answer that takes seconds to write. What must sit close to a city is older work, like banking and everyday cloud apps, which need a reply within thousandths of a second.

Are data centers being built in cities that use AI a lot, like San Francisco?

They try. Santa Clara, in Silicon Valley, has 57 data centers running or being built. But in 2025 two new ones there stood empty, because the city utility could not yet power them.

How much of the work is training, and how much is answering?

Today they are roughly even. Answering used nearly 60% of the electricity of US AI servers in 2023. Forecasters disagree about what comes next: Berkeley Lab expects training to pull ahead by 2028, and McKinsey expects answering to be the bigger job by 2030.

Is Stargate used for training or for answering?

Both. OpenAI trained its GPT-5.5 model there, and the site answers questions too.

Bring in the power

High-voltage lines pass the site. A substation taps them, and its transformers step the power down to something a building can use.

The Abilene campus is planned to draw 1.2 gigawatts. New York City, in its hottest hour of the year, draws about 11. So one campus is about a ninth of New York at its busiest. If it ran flat out all year, it would use about a fifth of the electricity New York City uses in a year, as much as a million American homes.

All the data centers in the US used about three and a half New York Cities' worth of electricity in 2023. All the data centers in the world now use a little more than France, and by 2030 they are expected to use a little more than Japan does today.

How much of Stargate is running today?

By April 2026 only about 200 to 300 megawatts of the planned 1.2 gigawatts was running. Later in 2026, Oracle said it had handed over six of the eight buildings.

Building and running the data center

The building is a long, low warehouse with no windows. Nothing inside needs daylight.

The empty building alone costs about 11 million dollars for every megawatt it can hold. Fitting it out for AI can more than double that, before a single chip goes in. So a one-gigawatt campus is at least a ten-billion-dollar building.

At the peak of construction about 1,500 people work on one large building. Once it is running, about 50 do.

Everything inside needs two things, all day and all night: electricity and cold.

What do the 50 people do?

About half work for the data center company: managers, engineers, and technicians who repair the computers. The other half are contractors: electricians, pipefitters who look after the cooling pipes, and security guards.

There is always something to fix. When Meta trained one large model on 16,000 chips, hardware faults stopped the work more than 300 times in 54 days.

Keep it cool

Almost all the electricity a computer uses turns into heat. Chilled water carries that heat out to cooling towers, where some of it leaves as steam.

Cooling takes power of its own. Engineers score a building with one number: all the electricity it draws, divided by the part that reaches the computers. A perfect score is 1, where every watt goes to computing. The average data center scores 1.54, so for every 100 watts of computing it spends 54 more on cooling, lights, and power lost along the way. The best come close to perfect. Google's data centers scored 1.09 in 2025: 9 extra watts for every 100.

Water is the other cost. One Gemini question uses about five drops, by Google's count. But Google's data centers together used up about 16,000 Olympic swimming pools of water in 2025, most of it evaporated. And the power plants that feed US data centers use about 12 times more water than the data centers do themselves.

Cooling forces a trade. Air cooling saves water but uses more power. Water cooling saves power but uses more water.

How does that water compare with eating meat?

Raising the beef for one quarter-pound burger uses up about 62 litres of water pumped from rivers and wells, not counting rain. That is the water of about 240,000 AI questions. All the water Google's data centers used in 2025 would make about 640 million burgers. And America's golf courses are watered with about 50 times as much.

So next to food, the total is small. The trouble is local: a data center draws from one town's supply, and some of those towns are dry.

Why can a data center in a desert not save both water and power?

Evaporating water is the cheapest way to shed heat, and it works best in dry air. That makes it tempting in a desert, where water is scarce. Cooling with air and chillers uses almost no water but more electricity: Google says its water-cooled sites use about 10% less energy than its air-cooled ones.

And making that extra electricity takes water too, at the power plant. So in a dry place, saving water at the building usually costs power, and some of that power costs water somewhere else.

Slide in the computers

Racks stand in rows. Each rack is a stack of servers, and an AI server is mostly graphics chips. This one holds eight.

Data centers are not new. The first big computer, ENIAC, filled a room in 1945. Most data center power still goes to ordinary things: email, banking, video, maps, search. AI servers used about a fifth of it in 2025. By 2030 it could be almost half.

What changed is how much is packed into one rack. A typical rack draws about 7.5 kilowatts, about what five space heaters draw. A new AI rack holds 72 graphics chips and draws about 120, nearly what all of ENIAC drew. NVIDIA said in 2025 that a rack planned for 2027 would draw about 600. Air cannot carry that much heat away, so the newest racks are cooled by liquid piped right onto the chips.

How can a newer rack be more efficient and still draw more power?

Each new chip does more work for every watt: about 30 to 40% more each year since 2016, by the International Energy Agency's count. But more efficient does not mean less power. One rack now holds far more chips, and people use far more AI, so the total keeps climbing. Data center electricity use grew 17% in 2025 alone.

Why are so many chips packed into one rack?

So they can work as one. The 72 chips in a new AI rack are wired to each other with short copper cables, which lets the whole rack act as one giant chip.

Add storage

Beside the racks stands a wall of drives. It holds what the chips work on: the text and images a model learns from, and saved copies of the model itself.

How much of a data center is storage?

A small part. In US data centers in 2023, servers used a bit over half of the electricity, cooling and power equipment about 30%, and storage about a tenth.

Do the chips only work on what is stored here?

No. During training, the chips read their lessons from these drives and save their progress back to them. When a model answers you, your question comes in over the network, and so do things like web search results. Only the model itself is loaded from storage into the chips' memory.

Run the fiber

Fiber-optic cable comes in from the road. Inside, more cable links every rack to every other rack, so thousands of chips can work as one machine.

How much cable is inside one rack?

A lot. NVIDIA says one of its newest racks holds 5,000 copper cables, more than two miles of them, linking its chips together.

The data center is built. Let's try it

Now it works. Press "Send a question" on the model and follow it. Your question arrives as light through the fiber. Chips across the racks do the math. The heat goes up the towers, and the answer goes back out a few letters at a time.

One question is small. Google said in 2025 that a typical Gemini text question uses 0.24 watt-hours, less than a second of a microwave. But billions of questions arrive every day, and an AI agent working through a long task can use many times more.

Teaching a model costs more up front. Training Meta's largest Llama 3.1 model took about as much electricity as 2,000 American homes use in a year.

That is the whole machine. There is one catch.

If one question is so small, why does AI use so much power?

Because there are so many questions, and because models have to be trained first. In the US, AI servers went from under 2 terawatt-hours of electricity in 2017 to more than 40 in 2023.

Then it waits years for a grid connection

The building is the quick part. The connection is the slow part. To draw power, a data center has to ask the grid operator for a connection, then wait in line.

In Texas, by March 2026 the big users asking to connect added up to about 410 gigawatts, and 87% of that was data centers. The list has kept growing. The most Texas has ever used at once is 91.1.

Enough electricity, but not in the busiest hour

So is there enough electricity? Most hours, yes. The grid is built for its busiest hour of the year, and the rest of the time it has room to spare. The trouble is that busiest hour. In the largest US power market, the yearly auction for standby power came up about 6.6 gigawatts short of its safety target for 2027, for the first time. That is more than half of New York City at its peak.

The line is slow for everyone. For new power plants across the US, the typical wait from request to switch-on doubled from under 2 years to over 4, and about three quarters of the capacity that asked gave up.

If there is enough electricity most hours, what stops a data center from connecting?

Two things: the hour and the place. A data center runs all day, so it has to fit into the busiest hour of the year too. Every request is studied against that hour, and whatever upgrades the study calls for must be built before the site can connect.

Then the power has to reach the site, and the lines and substations nearby may already be full. A new long-distance power line takes about ten years to plan, permit, and build. A large transformer alone takes about two and a half years to arrive.

Big, fast, firm: pick two

A data center can be big. It can be built fast. It can have a firm connection, meaning grid power that is always there. Today it cannot have all three.

Flip the switches and watch the site change.

Are gas turbines an efficient way to make power?

No. They are picked for speed. A trailer-mounted turbine can be set up in about two weeks, but small turbines turn only about a third of the gas they burn into electricity.

Other choices do better. Fuel cells turn a half to two thirds of the gas into electricity, and big gas engines about half. The best large gas plants reach almost two thirds, but the big turbines they need are booked up for years.

Big and fast: bring your own power

You skip the grid. In Memphis, xAI could get only 8 megawatts from the grid right away, so it parked 35 gas turbines outside, about 420 megawatts. For its second site, Mississippi approved 41 more, about 1.2 gigawatts. Turbines are scarce now too: the biggest maker is booked up for years.

Big and firm: wait in line

You wait for a full connection. In the busiest markets that now takes more than four years. The racks sit dark while the years go by. A large transformer alone takes about two and a half years to arrive.

Fast and firm: stay small

You build only what the power already at the site can feed. A few rows of racks, not a campus.

So they build around the grid

When the grid cannot keep up, the biggest companies stop waiting for it.

They buy whole power plants. Microsoft is paying to restart a reactor at Three Mile Island, planned for 2027. Amazon signed for up to 1.9 gigawatts from a nuclear plant in Pennsylvania. Meta signed nuclear deals for up to 6.6 gigawatts in a single week.

They pay for the grid itself. In Ohio and Georgia, regulators now make big data centers sign long contracts with a minimum bill. In Ohio they pay for at least 85% of the power they reserve, even if they never use it.

They move to where the power is: a five-gigawatt campus in Abu Dhabi, a gigawatt hub in India.

And they change how they cool. Microsoft's newest design cools chips in a closed loop that needs no fresh water. Google has cooled one Georgia site with treated wastewater for more than a decade.

What does cooling with treated wastewater mean?

Water from the town's sewage plant that was headed for the river is cleaned again and run through the cooling towers, so the site does not use drinking water. What does not evaporate is cleaned once more and returned to the river.

Everyone else pays too

A crowded grid costs everyone on it. In PJM, the largest US power market, that standby-power auction went from $2.2 billion to $14.7 billion in one year, then to $16.1 billion, where it hit the price cap. The market's own watchdog said data centers caused about $9.3 billion of the first jump.

These costs land on household bills. Virginia's watchdog estimated a typical home there could pay $14 to $37 more a month by 2040.

It is not the same everywhere. From 2019 to 2024, the US states where electricity use grew fastest saw prices fall after inflation. The pain is sharpest where the grid is shortest, like the mid-Atlantic.

Why does everyone pay, and not only the data centers?

Because of how the bill is split. The grid operator buys enough standby power to cover the busiest hour it expects. It divides that bill among every power company in the region by their share of the peak, and they pass it on to every customer. When data centers push the expected peak up, the price rises for the whole region. In northern Illinois, the utility said the first jump would add about $10.60 a month to a typical home's bill.

New power lines often work the same way. One study counted $4.3 billion of projects approved in 2024 to connect data centers in seven states. In almost all of them, the cost was spread across ordinary customers.

Virginia's watchdog found that data centers there pay their own way today. The extra cost comes from the growth they bring: new power plants and lines for everyone.

And the neighbors feel it

For most people, a data center is a small line on a power bill. For the people next door, it can be much more.

In Memphis, xAI's turbines ran without permits beside a neighborhood that already had some of the worst air in the city. They likely made the site the largest industrial source of smog-forming gas in Memphis. Near Virginia data centers, residents say the fans hum day and night. In Georgia, a couple say their well went bad after a Meta data center went up next door, which Meta disputes. In Oregon, Google used almost a third of one small city's water in 2021.

Towns are pushing back. By early 2025, local opposition had blocked or delayed about $64 billion of projects. By mid-2026, at least 17 Indiana counties had paused new ones.

Which part of a data center makes the smog?

Not the computers, which burn nothing. The gas comes from burning fuel on site: gas turbines like the ones in Memphis, and at ordinary data centers the diesel backup generators, which run for tests and during outages.

Virginia has permitted about 8,000 of those generators. They run rarely, and make under 4% of the smog-forming gas in Northern Virginia.

What it leaves behind

A data center leaves three things behind: heat, carbon, and old chips.

All the electricity ends up as heat. Most of it goes into the sky, but it does not have to. In Denmark, a Meta data center heats thousands of homes through the town's hot-water pipes.

The carbon comes from the power plants. The world's data centers made about 180 million tonnes of CO2 in 2024, about a fifth of what all flying did. Google's and Microsoft's emissions are both up by more than half since before the AI boom, mostly from building and equipping data centers.

The chips wear out. Servers for generative AI could add up to 2.5 million tonnes of electronic waste a year by 2030, if none of it is reused or recycled. That is about as much as all of Brazil or Japan throws away in electronics in a year, and about 3% of all the world's e-waste.

Can old AI servers be reused or recycled?

Partly. The same study found that reusing and recycling the servers could cut that waste by 16 to 86%.

So why do towns say yes?

Money, mostly. A data center pays a lot of property tax and needs few roads, schools, or police. In Loudoun County, Virginia, data centers bring in about 38% of the main county budget, and the tax rate for homeowners has fallen by about 30% since 2016.

States compete with tax breaks to win them, and the breaks are expensive. Virginia gave up about $930 million in one year and got back about half of it in other taxes.

Countries want them for another reason: control. A nation that runs AI on its own soil does not have to depend on someone else's. The US now speeds up federal permits for big sites, the EU is funding AI "gigafactories," and China moves AI work to its emptier west.

Some places have said enough. In Ireland, data centers use 22% of the country's electricity, and new ones may now connect only if they bring their own power.

Do data centers bring many jobs to a town?

Not many, but they pay well. A typical large building has about 50 full-time workers, about an eighth as many as some warehouses of the same size. They earn about $100,000 a year on average.

Who owns it?

Usually not the AI company you have heard of. Abilene shows the chain. Crusoe found the land and built it. Investors led by Blue Owl put up about $15 billion. Oracle leases it for 15 years. OpenAI runs its models inside. NVIDIA made the chips.

There are three kinds of owner. The giants own and run their own: Amazon, Microsoft, Google, and Meta plan to spend more than $700 billion on it in 2026, about twice what they spent in 2025. Then come smaller clouds that rent out chips, like CoreWeave and Nebius, some of them former bitcoin miners. And then landlords and investors who own the buildings.

Each AI lab leans on different partners. OpenAI rents from Microsoft, Oracle, Amazon, and CoreWeave. Anthropic runs mostly on Amazon's and Google's chips. Google's own lab uses Google's. Meta and xAI build their own.

The money often goes in a circle. A chipmaker invests in an AI lab, and the lab spends it on that chipmaker's chips.

How much does an AI lab spend on computing?

Billions, and most of it goes to building the models. Epoch AI estimates that OpenAI spent about $7 billion on computing in 2024: about $5 billion to train and test models, and about $2 billion to answer questions.

One fix: turn down in the busiest hours

The grid is built for its busiest hours. Most other hours it has room to spare.

A Duke University study found that 76 gigawatts of new demand could connect to the grid that exists today, if it agreed to trim back during those peaks. Across a year that means about 85 hours of turning down, and in most of them the site keeps at least half its power. The average cut lasts under two hours, a gap that batteries on site can fill.

Google signed deals like this in 2025, agreeing to slow some AI work when two of its local grids get tight.

What if data centers turned down a little more often?

More would fit. The same study found room for up to 126 gigawatts if new users turned down for about 366 hours a year.

Or put it somewhere else

If land near the grid is the problem, why not build where land is not?

In the desert, land is cheap and the sun is strong, but the heat makes cooling cost more water or more power. In the cold north, outside air cools the racks for free and hydropower is cheap, but the users are far away and the grids are small. In Iceland, the national power company has turned away new bitcoin miners to save power for other users.

Underwater, the sea cools the racks. Microsoft sank 864 servers off Scotland for two years, and they broke down one eighth as often as on land, but it ended the project in 2024. China has since put data centers in the sea, one powered by an offshore wind farm.

In orbit, a solar panel sits in near-constant sunlight and can make up to 8 times more power than on the ground. Investor Guru Chahal ends his whiteboard tour of the data center with this idea, and Google is studying it as Project Suncatcher, with two test satellites planned for early 2027. The catch is heat. In space there is no air or water to carry it away, only radiators. The whole International Space Station can shed about 70 kilowatts, less than one AI rack.

Why did Microsoft end its underwater project?

Microsoft said the test worked and gave no detailed reason for stopping. Outside engineers point to upkeep: servers sealed on the seabed cannot easily be repaired or swapped for newer chips.

How does a radiator in space compare with cooling by water?

Badly. On Earth, the average US data center spends about a third of a litre of water for each kilowatt-hour of computing, so one 120-kilowatt AI rack takes about 40 litres an hour.

A radiator can only glow its heat away as infrared light. The space station needs about 475 square metres of radiator panels for its 70 kilowatts. At that rate, one AI rack would need about 800 square metres, about three tennis courts. A startup planning data centers in orbit calculates that better panels could do it with about 190.