What a Hyperscale Data Center Actually Needs: Power, Water, Land

Drive past a hyperscale data center and there is honestly not much to look at. A windowless box, maybe a few hundred thousand square feet, and every so often three or four of them parked on the same campus. The biggest cloud and internet companies are the ones putting them up. But the box is not the story. What the box consumes is. A single site is a monster for electricity, thirsty for water, and greedy for land, and it wants all of that in one place. That is why these keep turning into local brawls, and it is worth taking the three demands one at a time.

Power is the Big Motivator

Nothing happens without electricity. Data Centers run massive computers with chips called GPUs, and they require energy to run tasks; they also give off heat and that means the processors and servers need to be cooled, and that cooling also needs electricity. Ten years ago, data centers weren’t running AI tasks and acting as power users of GPUs and may have only needed a few or ten megawatts off the grid. Today, hyperscale data centers run thousands of GPUs, in thousands of servers, all at once and for 24/7. That means hundreds of megawatts or even gigawatts. The chips behind these models eat power like nothing in a normal server room, and you stack rack after rack of them.

Natural resources: Cooling and water

All that power, used by GPUs, creates heat and if servers get too hot they shut down. To cool them, most data centers use either traditional air conditioning (which doesn’t always scale) or water cooling: Cool water gets pumped in, passes through the server rooms, and takes heat with it on the way out. The larger the data center, the more water it needs to cool things down. If that water is coming from sources used by the local community, questions arise as to whether a data center will leave everyone else—including farmers and ranchers sometimes—dry.

Land and siting

The land requirement is pretty hard to work around. Many acres near water or power, often with room for growth, are a key requirement for the largest data centers.

But having access to power, fiber connections, water, and acreage also means proximity to communities where people have lived for years. Traditionally, that has meant regions like Northern Virginia, central Ohio or swaths of Texas. Flexible or no zoning makes it even more attractive to those siting or building a data center: the fewer the rules, the less the red tape, the faster a data center gets built.

Because the lots are huge and the buildings are basically industrial, the real decisions happen locally, in zoning meetings and land-use hearings. Noise, construction traffic, what it looks like from the neighbor’s porch, how it gets taxed, all of it gets hashed out there. And that hearing, not the press release, is usually where a project lives or dies.

Backup power and on-site generation

The whole premise of a data center is that it does not go down, ever, which means backups behind the backups. Most sites keep rows of diesel generators, plus batteries or flywheels to cover the handful of seconds between the grid dropping out and the generators catching. Neighbors notice the generators. Partly it is what comes out of the stacks, partly it is the noise every time they get test-run. And now that grid hookups have gotten so slow, a few developers have quit thinking of on-site power as a backup at all. They are putting in gas turbines and fuel cells, and a couple are seriously kicking the tires on dedicated nuclear. At that point you have basically built a private power station next to the servers. That changes what the project means for the air and for the regulators, and it is the trend to watch hardest.

Data Centers have been around for years, but AI changed the game

Everyone knows about “the cloud,” which exists as big servers that hold data ranging from your text messages to your health information to real-time stock trades on Wall Street.

AI changed what happens in a data center. Artificial Intelligence applications need to be trained on huge amounts of data most people can’t imagine, like a library of every book ever written. All of this training happens on servers, and because so much data needs to be processed so quickly, they rely on GPUs to handle the load. More data, more GPU processing, more power, more heat, and more cooling are all the byproducts of AI training.

What communities like and what they hate

Data Center operators pay a lot of taxes to local governments, or payments in lieu of taxes. That can reduce a resident’s annual tax bill. But on the other side of the ledger, electric rates can increase, water rates can increase, and the humming noise from hundreds or thousands of servers rings through a community like torture.

The bottom line

In the end it is power, water, and land. Those decide whether one of these can go up somewhere and how bloody the fight gets. Not one of them is unlimited. Every one is local. And all of them have turned touchier as the buildout picks up speed. So when a proposal hits the table, a few blunt questions cut straight through. How many megawatts? How much water, and cooled which way? How many acres? And the one hiding under all the rest: who actually ends up paying?

This one is an evergreen explainer, and it will get updated as the technology and the siting playbook keep shifting.