Tech giants want to double A.I. electricity consumption in 5 years by enough to power more than 30 million homes. America can do it.
Between now and 2030, the giants of A.I. like OpenAI, Google, Microsoft, Amazon and Meta aim to more than double the computing power dedicated to growing and operating their non-human minds. They currently use about 40 gigawatts of power, enough for 30 million homes.
The cost of this ambition will be astronomical — about $50 billion per gigawatt of computing power built for a total of $2.5 trillion over the next five years alone. Roughly 80% of that will go to buy GPUs made by the likes of Nvidia and AMD; the rest — some $500 billion — will provide the energy via new power plants and transmission lines.
At the trajectory these hyperscalers are on, Goldman Sachs figures that by 2030 American datacenters will consume 500 terawatt hours per year — more than 10% of total domestic electricity. “I think we should already be raising the alarm on the potential for facilities to complete construction but be without power in 2028 and 2029,” says Zach Krause, an analyst at East Daley, a Denver energy consultancy. “I hope they don’t march into a wall.”
Some already have. In Oregon, Amazon Data Services has filed a complaint against Berkshire Hathaway subsidiary Pacificorp, which has refused to provide power to energize some of Amazon’s $30 billion in data center investments there. In Santa Clara, Calif. two 50 megawatt centers developed by Digital Realty and Stack Infrastructure are ready to go but can’t get electricity until Silicon Valley Power completes $450 million in grid upgrades – not expected until 2028 or later. Faced with new demand for 30 gigawatts of power, utility AES in Ohio told developers they had to enter into long term contracts to buy 85% of the power they want. (The queue dropped by more than half to just 13 GW).
Amid the doomsday scenarios are optimists. Joseph Majkut, director of energy security at CSIS, writes in a new report that this is a nice problem for the United States to face, and overcome: “Rapid demand growth should be good news. Despite trade conflicts and macroeconomic uncertainty, the United States is positioned to power more economic growth and strategic industries than we’ve seen in decades.”
“A lot of people are saying power is the constraint. Increasingly we believe that is not necessarily true,” says Carson Kearl, power markets analyst at Enverus in Calgary. He thinks building enough, fast enough shouldn’t be a problem. “The initial reaction is one of disbelief,” he says. “But there’s a lot of excess capacity in the market,” if you locate your project in the right place.
Alex Tang of VC firm 50 Years agrees. “If a hyperscaler has committed to it, it’s going to happen,” says Tang, who has invested in nuclear, batteries and solar startups. “We are one of the most efficient capital markets. We have shifted on a dime and can stand up a massive effort.” According to federal data the U.S. built 40 GW of new power plants in 2023, and is on track to erect 63 GW this year — half of which will be solar panels.
Plenty of datacenter developers are taking the situation into their own hands, building their own power generation on site rather than relying on utility companies to hook them up. These “behind-the-meter” generators are especially prevalent in Texas, which has its own power grid that isn’t subject to federal regulatory oversight enabling easier permitting. In Abilene, Texas, the Stargate project being developed by OpenAI, SoftBank, Oracle and investment firm MGX is building 10 gas turbines to serve as backup power.
Unexpected entrants into this space include the big oil companies, which aim to arbitrage the price difference of their low-value gas against high-value electricity. Chevron plans to build 5 GW of gas turbines for datacenters by 2027 in the Permian Basin of Texas where gas is so plentiful that prices at the Waha pipeline hub have gone negative this year. Giant oil companies, which already operate dozens of power plants at their refineries, will be able to lean on vendors for utility-scale turbines.
Others face 4-year wait times for big turbines from the likes of G.E. Vernova, Siemens and Hitachi. So they have turned to other options. Private equity giant Brookfield inked a $5 billion deal with Bloom Energy for their fuel cells that run on gas. Meanwhile, Elon Musk’s xAI, for its Memphis, Tennessee data center, has deployed dozens of smaller gas turbines (~30 megawatts) procured from Caterpillar subsidiary Solar Turbines. Kearl at Enverus thinks developers will be able to source an impressive 25 GW a year of smaller-size gas generators.
Natural gas will power about 60% of all this new datacenter demand, says Goldman Sachs. Such a domestic natural gas building boom is not unprecedented. According to the Energy Information Administration, in 2002 developers added 57 GW of gas turbines to the grid, led by independent power producer Calpine, which subsequently went bankrupt in 2005 due to surging natural gas prices amid tight supplies. (The company is now being acquired by Constellation Energy).
Demand from A.I. could even give coal a second life. Usage has ticked up in the past year as Trump’s EPA has proposed to repeal Biden-era anti-coal rules. Officials in Pueblo County, Colorado recently asked Xcel Energy to delay the closingCK of two coal plants until replacements have been found.
Longer term, a renaissance in nuclear power will also ensure ample electricity. Meta, Microsoft and Amazon have all contracted for years of power from decades-old nuclear reactors, and Constellation Energy even landed a federal loan guarantee to restart a mothballed reactor at Three Mile Island.
New nukes are in the works as well. Westinghouse and Brookfield landed federal backing to build $80 billion of new AP1000 reactors, while a dozen startups are working on new small modular reactors. Trump’s energy secretary Chris Wright wants to see nuclear reactors and datacenters built on federally owned lands, particularly military installations that could benefit from redundant power supplies in exchange for easy permitting.
International energy consultancy WoodMackenzie, forecasts that A.I. will help find even more energy than it uses. The firm took decades of data on 2,500 of the world’s biggest oil and gas fields, then fed it into a proprietary A.I. Using “integrative modeling,” the A.I. showed how the industry could boost global oil reserves by 500 billion barrels by applying best techniques to newer fields.
Likewise, Rocky Mountain Institute says plenty more power can be unlocked (50 GW or more) by increasing grid efficiencies, upgrading high-voltage transmission lines, and creating so-called “demand response” programs, where big customers agree to curtail electricity use during times of peak demand. Researchers at Duke University figure that if datacenter operators agreed to dial back power use during just 1% of their expected uptime, it would have the effect of giving the power grid “curtailment-enabled headroom” to the tune of 125 GW.
It’s easy to be a naysayer. In generations past, bubbles of overbuilding in railroads and fiberoptic networks and even gas turbines took years to absorb, and bankrupted plenty of operators. But this isn’t subprime lending — the data center building boom is backed by the richest companies and most powerful government in the world. If they need more electricity to remain dominant in A.I. (and protect their market capitalizations), they will find a way.
