Nvidia’s next-gen chips will guzzle electricity. Will the US have enough power?

Big tech companies, which are all betting heavily on AI, will undoubtedly buy those chips, even if the price skyrockets. But there’s growing evidence that there won’t be enough electricity to power all of their AI dreams.

The virtual world is starting to hit a physical limit.

It all starts with gas turbines, the spinning machinery that transforms natural gas—the largest source of America’s electricity—into electric power. Power company executives say they’re building more big turbines but are unwilling to vastly boost capacity in a way that would satisfy every potential AI company, out of fear of overbuilding. The wait for new turbines now stretches more than three years. And the costs to buy them are jumping faster than Taylor Swift concert tickets.

The power crunch could come soon. Morgan Stanley estimated last year that data centers will need 57 gigawatts worth of new electricity capacity by 2028, or as much as 70 large-scale power plants produce.

There are three dominant turbine makers—GE Vernova, Mitsubishi Power and Siemens Energy. The boom in power demand associated with AI has been incredibly rewarding to their investors. GE Vernova stock is up 160% in the past year; Mitsubishi Heavy Industries, the parent of Mitsubishi Power, is up 60%; and Siemens Energy has rocketed 300%.

The industry has gone from a deep lull to a dead sprint in a very short period. In 2022, only one large gas turbine was sold by all manufacturers in the U.S., said Rich Voorberg, president of Siemens Energy in North America. In just the last six months, Siemens Energy has signed deals for dozens of large turbines in North America.

There’s much more growth ahead. As of the end of 2024, companies and utilities were planning 363 new natural gas plants in the U.S., with planned output jumping 70% in just two years, according to the Sierra Club, which tracks the industry.

The turbine companies are scaling up somewhat to address the demand. Mitsubishi is on track to expand production capacity by 30%, and Siemens Energy is doubling the output of one turbine equipment plant in Tampa. GE Vernova is moving at a faster pace, planning to produce 70 to 80 heavy-duty turbines a year by the second half of 2026, up from 48 in 2024.

But the ramp-up also comes with considerable risks for those same companies. Expanding turbine production aggressively now would expose them to major losses in the event that power demand growth slows. Past busts have led to mass industry layoffs.

Executives at Mitsubishi Power and Siemens Energy said in interviews in recent days that they’re being cautious about adding capacity. GE Vernova has made comments recently about growing at reasonable rates. The companies are churning out more turbines, but almost certainly not fast enough to meet all the potential demand from those AI chips. They want to be prepared for a bust in the future, even if they don’t see it on the horizon just yet.

“We’ve been through a lot of these upcycles and downcycles,” Rich Voorberg, president of Siemens Energy in North America, said in an interview. “A lot of us have been burned over the years.”

On paper there are hundreds of billions of dollars worth of data centers coming—up to $500 billion alone for the Stargate project involving OpenAI and SoftBank. In reality, they may not all pan out. “The challenge we’ve got is trying to figure out who’s real and who isn’t real today,” Voorberg said.

Bill Newsom, CEO of Mitsubishi Power Americas, said in an interview that while he is enjoying the boom, he’s already got an eye out for a bust. “My biggest question is how long is this going to happen,” he said. “Every day I wake up trying to answer that question. Is this a 3, 5, 10, or 15-year run?”

AI progress has proven unpredictable. Stocks of the turbine-makers tumbled in January after the debut of Chinese AI company DeepSeek, which promised top-tier AI performance with much lower energy use.

Already, some power plant projects are falling by the wayside for lack of turbines. Power company Engie canceled two major Texas power plant projects last month—one of which was on a shortlist for state support—because of equipment shortages. Among the equipment that the company couldn’t procure in time was gas turbines.

Voorberg said that it was “very accurate” to assume that other power plant projects will suffer a similar fate. Voorberg said that tech CEOs often have unrealistic assumptions about how the power market works, seeming to think it can scale up the way software does. Some of them want to buy turbines directly so that they can create off-grid power plants that hook directly into data centers without having to travel through the larger grid. Companies like Chevron and Exxon Mobil are partnering with tech companies to make that happen.

Voorberg said Siemens Energy tells tech firms to come up with comprehensive plans to partner with utilities or other power companies, and have a clear path to regulatory approval. He also needs to ensure he has enough turbines for the company’s tried-and-true customers—utilities. “The ones that are the flash in the pans, we will definitely sell to them,” he said. “But we’re also not going to forget our long-term customers.”

The leading manufacturers are now sold out of big turbines until 2028 or later—and it can take up to two years after turbine delivery for a company to actually generate power from a natural gas power plant. There’s lots of other equipment that needs to be hooked up, and permits to get, after the turbine arrives on site.

“To get your hands on a gas turbine and to actually get it built across the market, you’re really looking at 2030, or later,” said John Ketchum, the CEO of NextEra Energy, at an energy conference in Houston last week. Ketchum’s company owns Florida’s biggest regulated utility, which depends heavily on natural gas, and is one of the country’s largest renewable power developers too.

That delay clearly poses a problem. Patience is not a virtue that anyone associates with the AI race today. Advances in the technology are happening lightning-fast, and companies are not keen to wait until 2030 to secure the power they need.

The shortage is a threat to the growth of AI—and more power plants will invariably be canceled—but some companies are coming up with workarounds too.

Some tech companies are buying a large number of smaller turbines and combining their power to equal the generation of a big one. It’s an inefficient solution but one that’s gaining momentum. The big players make small turbines, but there are other suppliers as well, including Baker Hughes and Caterpillar. They should add market share in turbines in the years ahead. Bloom Energy, which uses fuel cells to turn natural gas into electricity, is also benefiting.

In addition, tech companies may have to invest even more in renewables, which can come online faster than natural gas and already make up the largest share of new electricity generation. When coupled with batteries, they can provide power for much of the day. Ketchum of NextEra says he thinks renewables will have to fill the gap for the next few years, at least.

And companies may simply have to make do with the power that’s available, by becoming more efficient. Nvidia and Schneider Electric announced a partnership on Tuesday to find ways to cut power use through efficiency gains. Duke University researchers published a paper last month that found existing electricity resources could accommodate dozens of massive new data centers as long as the tech companies were willing to curtail power demand temporarily at moments of high stress on the grid.

AI seems unstoppable. The looming power crunch could be its biggest test.

Write to Avi Salzman at [email protected]