Why Data Centers Build Out, Not Up

This is Random Daily, from All Things Random. One curious question, explained. Today: why data centers build out, not up.

Fly over Northern Virginia, home to the biggest cluster of data centers on Earth, and the landscape looks strange. Mile after mile of enormous gray boxes. Few windows, if any. And most of them low and flat, one or two stories, as if someone took an office tower and laid it on its side.

That's odd, when you think about it. These are some of the most important buildings in the modern world. They store your photos, stream your shows, and answer every question you type into a chatbot. And in a city, important buildings go up. When land gets scarce, we stack. So why do data centers sprawl across the ground like parking lots?

The comparison is closer than it sounds. A parking lot is cheap. A parking garage costs far more for every space, because each upper floor has to carry everything parked on it. So garages only go up where land is too expensive for a lot. Data centers face the same trade-off. Only their "cars" are much heavier, and they never stop giving off heat.

Inside every data center are long rows of metal cabinets called racks. Each is roughly the size of a tall refrigerator, packed with computers called servers. For years, a loaded rack was heavy but manageable. Then artificial intelligence arrived. Fill one of those cabinets with today's AI hardware, and it can weigh about as much as a mid-sized car.

Now picture hundreds of them, on the fourth floor.

Weight is the first reason data centers stay low. Heat is the second. Land is the third. And yet, in a few corners of the world, data centers are climbing anyway. One stands eleven stories tall. Another spent two years at the bottom of the sea.

But first, that cabinet. How does a box of computers end up weighing as much as a car?

Start with how servers are stored. They slide into a rack like trays into a bread cart, one above another. Each slot is measured in a unit called a rack unit, about one and three quarter inches tall. A typical rack holds forty-two of them.

For years, an ordinary server weighed about seventeen pounds for every rack unit it filled. Heavy, but nothing a sturdy floor couldn't handle. Then came the machines built for artificial intelligence. They pack in more chips, bigger metal heat sinks, and plumbing for liquid cooling. An AI server can weigh around thirty-six pounds per rack unit. More than double.

Add it all up, and a fully loaded AI rack may need to hold about five thousand pounds. Even the copper cables inside can add hundreds of pounds on their own. And all of that weight sits on a patch of floor about the size of a bathtub.

On the ground floor, that's manageable. The concrete slab rests directly on the earth, and the earth doesn't mind. In fact, many older data centers raised their racks on a false floor of tiles and posts, with cables and cool air running underneath. With today's heavy racks, strengthening those raised floors gets expensive. So more and more new data halls simply put the racks straight on the slab.

Upstairs is a different story. An upper floor is a slab held up in the air by beams and columns. To carry thousands of pounds per rack, that floor needs far more concrete and steel than an office building, whose floors are designed for desks, chairs and people.

Then there's getting the equipment up there. A single-story building needs a loading dock and a wide hallway. A tall one needs freight elevators strong enough to lift a loaded rack. Engineers designing multi-story data centers say the big operators ask for elevators rated to lift five metric tons. Elevator makers now sell models built just for data centers, with several times the capacity of a passenger elevator.

And in earthquake country, like the West Coast, strict seismic rules make every extra story cost even more.

So weight alone makes a tall data center expensive. But weight is the problem you can see. The bigger one is invisible. Every one of those servers is also a heater.

Here's a basic fact of physics. Nearly all the electricity a computer uses ends up as heat. A server doesn't burn its power away somewhere else. It warms the room.

A household space heater uses about fifteen hundred watts. For a long time, a typical data center rack used somewhere between five and ten thousand watts. So each rack was like a handful of space heaters running nonstop.

AI changed that too. One of Nvidia's best-known AI racks draws around one hundred and twenty to one hundred and forty thousand watts. That's roughly ninety space heaters, packed into a single cabinet. Newer designs aim even higher.

All that heat has to leave the building, or the chips slow down and fail. For decades, the answer was air. Cold air blows in front of the racks, the servers pull it through, and hot air comes out the back. But air can only carry so much. Somewhere past thirty to fifty thousand watts per rack, air struggles to keep up. That's why the newest AI racks are cooled with liquid, piped right to the chips.

Either way, the heat ends up outside. And this is where the shape of the building matters.

Heat rises. In a single-story data center, hot air can travel straight up and out through the roof. And the roof itself becomes a giant work surface for cooling equipment: rows of fans, chillers and cooling towers that throw the heat into the sky. One story means every data hall sits directly under its own roof.

Stack the data halls, and that gets harder. The hot air from the lower floor has to be routed around the floors above it. Meanwhile, the building has more heat to get rid of, but no more roof than before. Two floors of servers share one roof. Four floors share it four ways.

There's water to think about, too. Many data centers cool themselves partly by evaporating water, the same way sweat cools your skin. A mid-sized data center can use up to three hundred thousand gallons of water a day. Moving that water, and those pipes, up through a tall building adds more cost and more risk.

So weight pushes data centers down, and heat pushes them out. Which raises a fair question. If staying low is so much easier, why is anyone building data centers taller at all? The answer is the third reason. Land.

A data center needs a lot of land. Not just for the building, but for everything around it: rows of backup generators, an electrical substation, cooling equipment, fences and a buffer from neighbors. A traditional single-story campus was often planned on twenty acres or more. Today's biggest AI campuses are measured in hundreds or even thousands of acres.

Where land is cheap and plentiful, the math is simple. Spread out. That's why so many new campuses go up in farm country and on the edges of small cities. In October of twenty twenty-six, one major developer announced a campus in Kokomo, Indiana, on nearly five hundred acres, made of three single-story buildings.

But data centers don't only want cheap land. They want land close to fiber optic lines, close to big power connections, and close to the people using them. And in the most popular places, that land runs out.

Northern Virginia is the clearest example. Through twenty thirteen, Loudoun County had twenty-one single-story data centers, and only four with two stories. Then land got scarce and expensive. Of the shared, rental data centers built there over the next four years, only three were single story. Eleven had two stories.

Other places went further. Google built four-story data centers in Iowa and Oklahoma. In Santa Clara, in Silicon Valley, one operator built a four-story data center. In Singapore, where land is famously scarce, Meta built one that stands eleven stories tall. In Hong Kong, data centers fill floors of industrial towers more than twenty stories high. And in lower Manhattan, a thirty-two-story former phone company building was turned into a data center.

So when does building up pay off? Industry engineers who studied it in Northern Virginia found that two and three stories give good value. Past four stories, the gains shrink. Every extra floor needs a stronger frame, bigger elevators, and more work to move heat and power. Eventually, the savings on land are eaten by the cost of climbing.

That's the parking garage again. Nobody builds a garage in a cornfield. But downtown, a garage is the only way to fit the cars. Data centers are heading the same way. Mostly flat where there's room, and taller where there isn't.

And then there are the places where engineers ran out of options on land entirely, and went somewhere much stranger.

In twenty eighteen, Microsoft sealed eight hundred and sixty-four servers inside a steel cylinder about the size of a shipping container. Then it lowered the whole thing into the sea off the Orkney Islands in Scotland, one hundred and seventeen feet down. Inside, there was no oxygen and no people, just dry nitrogen and the cold sea outside to carry the heat away.

Two years later, the team hauled it back up, covered in algae and barnacles. And according to Microsoft, the servers inside had failed at about one eighth the rate of the same servers on land. The company has since said it isn't building more underwater data centers. But it learned a lot about what makes servers last.

Other engineers went underground. In Stockholm, a data center called Pionen sits about thirty meters inside solid rock, in a former Cold War bunker, behind a steel door. Its builders added gardens, waterfalls and a big fish tank, to make it a little more pleasant for the people who work there.

And in Helsinki, there's a data center in an old air raid shelter beneath a cathedral. Its waste heat is put to use, warming homes nearby.

So the next time you send a message or ask a chatbot a question, picture where it goes. Most likely to a long, low building on cheap land, with machines too heavy to stack easily, running too hot to crowd together. Sometimes to a tower in a crowded city. And once, at least, to the bottom of the sea.

That's Random Daily. The full article, with diagrams and sources, is at all things random dot net. Follow the show for tomorrow's question.