D espite fears of rogue AI agents and increasing political pushback by local communities and politicians, the U.S. is plowing ahead on a data center construction frenzy that’s unprecedented in its scale.

Depending on who you ask, estimates of the total cost of the buildout range from $2.8 trillion by 2030 to $10.3 trillion by 2032 in the U.S. alone. Those figures could even be conservative, with giants OpenAI and Anthropic revealing more about their spending plans as they prepare to go public in the coming months: A leaked copy of Anthropic’s prospectus revealed the firm expects to spend $518 billion on AI infrastructure over the next decade.

Yet we’re still in the early innings of what could become a new industrial revolution. According to economists at Goldman Sachs , U.S. investment on AI will rise from 1.8% of GDP this year to 2.8% in 2028. That’s roughly the same share of the economy as the federal defense budget.

The railroad boom represented about 2.4% of America’s GDP between 1880 and 1890. If AI spending reaches the 2.8% of GDP predicted by Goldman Sachs, it will outstrip the railroad boom in its economic footprint by next year—and that could be a conservative estimate. A recent study from Columbia University economist Stijn van Nieuwerburgh puts that even higher, at 3.6% of GDP by 2032.

The advent of railroads in the 1800s ushered in the Gilded Age, elevating the era’s industrial barons to unimaginable levels of wealth. When Forbes published its first-ever ranking of the country’s richest people in 1918 , five of the 30 richest made their money in railroads. Another seven, including John D. Rockefeller and Andrew Carnegie, provided the oil and steel to build the rails and keep the trains moving.

Now the AI boom is minting dozens of new billionaires and boosting the fortunes of tech titans to dizzying heights. Elon Musk briefly became a trillionaire after the IPO of SpaceX—which despite its name, spends significantly more on AI than it does on rockets. Today, 11 of the 30 richest Americans have fortunes rooted in companies central to the AI economy, including Nvidia’s Jensen Huang and Google’s Larry Page and Sergey Brin.

The buildout is also reliant on mountains of debt to finance it. According to credit intelligence platform Atrium , data center developers have already raised at least $1.3 trillion in debt. “Who loses money when things die on the vine?” asks Ryan Alfred, Atrium’s CEO and co-founder, adding that he’s seen projections that data center lending could balloon to as much as $10 trillion.

Even small changes could have a knock-on effect if firms find themselves no longer needing—or not having the resources for—the data center capacity they’ve leased or built. “With so much debt, you don’t need a big shock to the values. Ten percent will wipe out the equity and will begin to impair the debt,” says Columbia’s van Nieuwerburgh.

AI firms and developers need to keep raising cash to pay for data centers that are getting increasingly larger, more energy-hungry and more expensive. But even those cost estimates vary significantly. Per construction giant Clayco, which is part of the teams building billions of dollars’ worth of data centers for the likes of Oracle and CoreWeave, a typical 40-megawatt data center—enough to power Burlington, Vermont—costs some $500 million to build.

Half of that cost goes to buying advanced GPU chips and other IT equipment, while power, energy and water infrastructure (including concrete, steel, copper, cooling systems and turbines) add another 35%. On an ongoing basis, the biggest costs are electricity and water, with a new data center of this size using about 60,000 gallons of water per day.

Firms like Clayco have been cashing in on the demand, with data center construction expected to account for 75% of the firm’s nearly $12 billion in revenue this year and powering its founder, Bob Clark, to a $7 billion fortune. He’s far from the only one, as major developers like the Related Companies—which developed New York’s Hudson Yards—also get in on the action with $55 billion worth of data center developments, helping boost the net worths of its billionaire founder Stephen Ross and president Bruce Beal, Jr.

On a larger scale, McKinsey estimates that nearly two-thirds of the cost of data centers comes from GPU servers and data storage, to the tune of $4.4 trillion on a global level through 2030. They estimate $1.3 trillion will be spent on electrical and mechanical equipment, power generation and network infrastructure, with an additional $1 trillion needed for labor, real estate construction and land acquisition.

Trillions of dollars’ worth of land, cement, electricity and water are hard to put into perspective. As of 2024, data centers consumed more power than the entire state of Ohio. By 2030, they’re expected to need as much as what Texas uses today. Over the next four years, they could eat up another 457 terawatt-hours of electricity—enough to power California for nearly two years. The problem facing many data center developers is that capacity just doesn’t exist yet.

Percentage of U.S. electricity used by data centers:

(More than Ohio, about as much as two Louisianas)

(More than Texas, the most power-hungry state in the country) SOURCE: LAWRENCE BERKELEY NATIONAL LABORATORY

Firms that already have access to power—including a host of former bitcoin miners that have transitioned their mining operations to data centers they now lease out to AI firms—have benefited from that shortage. “There is such a shortage of power right now that pretty much any company that has access to prompt power seems like it’s able to get contracts,” adds Greg Lewis, an analyst at financial services firm BTIG, noting the surging stock prices over the past year of outfits like Cipher Digital and TeraWulf.

Private companies are spending tens of billions of dollars on electrical power, all in an effort to keep up with the insatiable demand from data centers. Whether they’ll be able to build it fast enough is still unclear. “A lot of utility companies have committed to doubling their capacity over the next 10 years in order to support deals that are already permitted,” adds Atrium’s Alfred. “And utilities don’t generally double power capacity in 10 years. A lot of them aren’t going to be able to, and a lot of lenders won’t lend unless [the project] has got utility power.”

The firms building the cables that move all that electricity are benefiting, too. Southwire, the privately held cable manufacturer owned by Georgia’s Richards family, posted a record $9.7 billion in revenue last year thanks to rising copper prices and greater demand from data centers, making the Richards family worth $13.1 billion.

Land used by data centers in:

(slightly larger than Manhattan)

(About the size of Des Moines, Iowa; larger than D.C.)

of cement needed by 2028, total (roughly enough to pave the highway between Los Angeles and San Francisco) SOURCES: HINES, AMERICAN CEMENT ASSOCIATION

America is vast, but so is the amount of land and cement needed for the buildout. As of last year, data centers currently took up slightly more land than the island of Manhattan. By 2030, they’re expected to eat up more than the entire District of Columbia. That may seem small given the country’s size, but developers are snapping up as much land as they can get so they can be prepared several years down the road.

“There is no power available in 2026, 2027, 2028. It’s all been taken because this is a multi-year development cycle business,” says Raul Martynek, CEO of Dallas-based data center builder DataBank. “We’re planning for 2030 or beyond, literally acquiring land now where we think the ready for service date is 2030, 2031, 2032.”

Even if the land is available, securing the labor to build the facilities is another major bottleneck. “Labor is just a real pinch point because you can’t just crank out more skilled electricians,” adds Martynek. That’s also an issue when it comes to key electrical components such as generators, where delivery times have slipped from 12 weeks a few years ago to nearly 2 years today.

Massive amounts of cement are also necessary to mix into the concrete to build the hulking data center structures. One beneficiary has been Tennessee-based SRM Concrete, which is pouring concrete for 13 data centers for hyperscalers including Meta, Oracle and Google, helping the firm grow and vaulting its founder, Mike Hollingshead, to a $7.2 billion fortune.

Water consumption by data centers globally in:

(About 58 days of NYC daily water consumption; 88,800 Olympic-size swimming pools)

(About 5.5 months of NYC daily water consumption; 257,600 Olympic-size swimming pools)

*Water-efficient scenario

(About five months of NYC daily water consumption; 217,200 Olympic-size swimming pools)

*Aggressive water-saving scenario (About three months of NYC daily water consumption; 155,200 Olympic-size swimming pools) SOURCE: RYSTAD ENERGY

A key component of the pushback to data centers spreading across the U.S. is the strain on water resources. A recent Gallup poll found that 70% of Americans opposed building AI data centers in their communities, with 50% of those citing the effect on resources including excess water usage.

According to consultancy Rystad Energy , data centers used 59 billion gallons of water last year globally—enough for 58 days of New York City’s water consumption. Under even the most aggressive water-saving scenarios, that could grow to 102 billion gallons by 2030, and far higher if developers don’t take steps to mitigate water use.

Some investors have been securing water capacity to profit from that demand. In West Texas, a growing data center hub, billionaire David Capobianco has amassed companies that handle more than 6 million barrels of wastewater that could be pumped to cool data centers, in addition to 320,000 acres of land where he can build them himself.

Firms are now installing more advanced cooling systems that use less water, but they consume more electricity. That adds to the already enormous pressure on the electrical grid from the data centers. At the scale of a 1-gigawatt data center—the size of two Meta campuses in Indiana and Texas that aren’t even among its largest projects—those will use up about as much electricity as Seattle does, plus nearly 3% of the city’s annual water use.

That amount of demand, the equivalent of plugging the city of Seattle into the grid at once, is almost unheard of. “We have very seldom in America ever seen a gigawatt load of demand pop down in a single place in the grid,” says Kiran Bhatraju, CEO of digital utility firm Arcadia.

Some are also skeptical that the full scale of the buildout will ever happen. In an August research note, analysts at equity research firm Bernstein counted a total data center pipeline in the U.S. of more than 400 gigawatts, which would be more than enough to power the entire country of Japan. Yet they estimated that only 35% of that will actually see the light of day.

The upside of this investment, if all these data centers get built, could be enormous. Bain & Company estimates AI could drive more than $1.5 trillion in growth and new value, ranging from physical AI, robotics and autonomous vehicles to search and advertising and new drug discovery and scientific breakthroughs.

Still, it’s unclear when—or if—that growth will actually happen. “We need to generate $3.7 trillion in annual revenue by 2032 to earn a 10% return on all that investment,” says van Nieuwerburgh. “And that’s just an astronomical number.”