AI Is Already Wreaking Havoc on Global Power Systems

These are only estimates and there remains a high degree of uncertainty about how the current AI frenzy will play out. There’s also a difference between the projections for how much electricity data center developers want and how much generation actually gets built.

While tech companies are quick to point out that data centers account for less than 2% of global energy use even with all the expansion, an April report from Goldman Sachs estimates that figure could rise to 4% by the end of the decade. Any percentage point increase is monumental, given overall electricity demand has remained almost flat for years — if not declining in some regions.

In the US, power demand is expected to grow by 40% over the next two decades, compared with just 9% growth over the past 20 years, according to John Ketchum, chief executive officer at NextEra Energy Inc., the world’s biggest builder of wind and solar energy that isn’t backed by a government. Data centers are the biggest reason for that demand boom, Ketchum said, citing electrification and manufacturing as other factors. Asked why data centers were suddenly sucking up so much power, his answer was blunt: “It’s AI,” he said, citing the energy needs for training models and also the inference process by which AI draws conclusions from data it hasn’t seen before. “It’s 10 to 15 times the amount of electricity.”

The biggest cloud services providers, Amazon.com Inc., Microsoft Corp. and Alphabet Inc.’s Google, have announced goals to run their data centers entirely on green energy — Amazon by next year, Google and Microsoft by 2030. All three say they are working on technological methods to use less power or balance the demand on the grid more efficiently. That can include wringing more efficiency from chips and servers, laying out equipment in ways that require less cooling and shifting loads to different areas based on where energy — particularly green energy — is available.

But some tech leaders like OpenAI CEO Sam Altman have said an energy breakthrough — likely from nuclear power — is needed to adapt to this new picture. Microsoft and Amazon are also betting on nuclear energy, with Amazon recently buying a nuclear-powered data center in Pennsylvania and indicating it’s open to more. For now, the path forward remains unclear. Microsoft recently admitted that its AI push is jeopardizing its long-held goal to be carbon negative by 2030.

“We need terawatts and terawatts more of traditional green energy, whether it’s wind or solar, and that’s across the globe,” said Amanda Peterson Corio, Google’s global head of data center energy, speaking broadly rather than solely about AI demands. For context, a single terawatt is equivalent to the output of about 1,000 nuclear power plants. “Of course we want to decarbonize ourselves, but if we’re just decarbonizing ourselves and not the whole grid, what’s the point?”

In today’s data centers, you might find thousands of Nvidia Corp.’s coveted H100 chips — the engine of the generative AI boom — each of which draws as much as 700 watts, or nearly eight times the power used by a typical 60-inch flat screen TV. Data centers built for training AI models require even more. Microsoft, for example, strung together tens of thousands of Nvidia processors inside a facility used to develop OpenAI’s technology. These are fed by networking and other types of chips which, combined with machinery in data centers used to prevent the gear from overheating, saps up even more power.

And the conventional wisdom in Silicon Valley is that the amount of energy needed will only go up. Nvidia’s newest chip, the B100, can consume nearly twice as much power as the H100. Nvidia contends companies will be able to do more with fewer chips, but Ian Buck, the company’s head of accelerated computing, admits it’s likely AI deployments will increase. “People like to fill their data centers,” he said.

AI development is evolving fast, too, with a feverish push toward developing ever larger artificial intelligence models. The Microsoft supercomputer built in 2020 that trained OpenAI’s GPT-3 system used 10,000 of what was then the latest AI chip. A November 2023 Microsoft supercomputer relied on 14,400 of the top of the line Nvidia H100 chips and the next one, which is already being designed, will be 30 times more powerful, according to a May slide presentation by Microsoft Azure Chief Technology Officer Mark Russinovich.

Meanwhile, Nvidia CEO Jensen Huang said many nations will push to build their own “sovereign” AI systems to stay competitive and process data locally. The global battle for AI supremacy may well depend on which countries have enough data centers and power to support the technology. If so, Loudoun County is a vision of what’s to come for the rest of the world — and the challenges other countries will face keeping up.

The data center capitals of the world

Over the past five years, Dominion Energy Inc., the power company that services Loudoun County, also known as “data center alley,” has connected 94 data centers that consume about four gigawatts of electricity, combined. Now it’s fielding requests for data centers campuses that would consume multiple gigawatts — enough to power hundreds of thousands of homes — two or three of which could use as much electricity combined as all the facilities hooked up since 2019.

The surge in demand is causing a backlog. Data center developers now have to wait longer to hook their projects up to the electric grid. “It could be as quick as two years, it could be four years depending on what needs to be built,” Dominion Energy Virginia president Edward Baine said in an interview.

Dominion is trying to build out the infrastructure to support it. New power lines hang from massive metal towers and run along roads and over creeks to feed electricity to these towering, windowless data centers. The company is building a large new wind farm off the coast and a lot of solar farms, but coal and gas powered plants could also stay online longer.

In late 2022, Dominion filed a previously unreported letter to its regulators asking for permission to build new substations and power lines to serve “unprecedented” load growth. In the letter, Dominion said it experienced 18 load relief warnings in the spring of that year. These warnings occur when the grid operator tells the company that it might need to shed load, the technical term for the controlled interruption of power to customers, which could include rotating outages.

“This is far outside of the normal, safe operating protocol,” Dominion told regulators.

Virginia is not alone in struggling to keep pace with demand. In West London, historically considered a hub for data centers, new facilities have to wait until 2030 to connect to the grid, according to David Bloom, chairman of UK-based data center operator Kao Data. “We are now being pushed into new locations,” he said. Likewise, in the south of Sweden, a strong market for renewables, there is so much demand to get connected that businesses may have to wait years. “We have one queue, and you need to get your ticket,” said Peter Hjalmar, German utility E.ON SE’s regional manager for southern Sweden. And across the US, many new AI data centers are expected to consume 100 megawatts each, according to a recent Morgan Stanley analysis, prompting the analysts to wonder “how all of the proposed data center projects will be powered in a timely manner.” Demand is so high that large tech companies are having bidding wars over data center sites with ready access to power, according to NextEra.

Data center growth, as it’s being forecast, may also run up against the limits of how much power can be carried through transmission lines, said Ali Farhadi, CEO of the Allen Institute for AI. “I don’t think we can move that much electricity around the globe, forget about generating it,” he said.

Data centers will get more efficient over time, energy analysts say, but they’re also getting significantly bigger. The average size of data center facilities worldwide is now 412,000 square feet, an almost fivefold increase from what it was in 2010, according to data from DC Byte, a market intelligence firm.

The surge in data center demand, combined with heavy investments from power companies like Dominion on new substations, transmission lines and other infrastructure to support it, are also increasing the likelihood customers will see their energy prices go up, experts say. The cost of some upgrades are typically allocated among electricity customers in an entire region, showing up as a line item on everyone’s monthly utility bill.

Goldman Sachs estimates that US utility companies will have to invest roughly $50 billion in new power generation capacity to support data centers. “That’s going to raise energy prices for both wholesale energy and retail rates,” said power market analyst Patrick Finn of energy consultancy Wood Mackenzie.

Costs including grid improvements are divided among each customer class, from residential to industrial, based on how much it actually costs to serve each, according to a Dominion representative. As a result, residential customers have seen their share of transmission costs drop in recent years while data centers have seen their portion rise, the representative said.

In Ireland, another heavily saturated market, there are some early signs of rate increases. Blessed with a moderate climate, Ireland has attracted so many data centers from Microsoft, Amazon and others that these facilities are forecast to consume a third of the country’s energy by 2026, up from 18% in 2022.

Wholesale power prices in Ireland have been a third higher on average this year than the rest of Europe. Other factors, including the country’s geography, play a role, but Sarah Nolan, senior modeler at Cornwall Insight, said growing data center demand can contribute — and that’s in a country that took strong steps to limit buildout just before the AI craze kicked off.

To manage energy constraints, Ireland’s state-owned electricity operator imposed a moratorium in Dublin in early 2022 and set out conditions to connect new data centers to the grid, including a preference for those generating their own electricity. The operator has “comprehensive” plans to build out the grid, said Donal Travers, the head of technology, consumer and business services for IDA Ireland, the state agency tasked with attracting foreign direct investment. But he said restrictions on large new connections are expected to continue “probably until 2028 or thereabouts.”

In the meantime, other regions are all too eager to open their doors.

The next Virginia

When Rangu Salgame looks at Malaysia, he sees the next Virginia “in the making.” Johor, the southernmost state in peninsular Malaysia, has a policy to speed up clearances for data centers. Crucially, it’s also a short drive to Singapore, a longtime data center hub that imposed a moratorium for several years on new facilities to manage energy growth on the tiny island.

Once a sleepy fishing village, the suburbs of the city of Johor Bahru are now marked by vast construction sites. Microsoft and Amazon are investing in the region, as is Salgame’s company, Princeton Digital Group (PDG). At Sedenak Tech Park, a sprawling complex about 40 miles south of Johor Bahru’s city center, towering cranes dot the sky. PDG’s new 150 megawatt data center occupies one corner of the park, across from similar facilities from other providers.

“We have gone from shovel to production in 12 months,” said Salgame, who expects to have 300 megawatts of capacity in Johor within two years. “Not all parts of the world can execute at this speed and scale.”

But even markets eager to streamline data center buildout face constraints. What’s missing in Johor, especially for an industry like tech that is known for its climate pledges, is renewable energy. The power supply at Sedenak comes from Tenaga Nasional Berhad, which uses coal or gas-fired plants. While Malaysia has ambitious goals to bolster renewables, including plans to build a 500-megawatt solar farm in Johor, today it relies on coal for more than a third of its generation. Most of Malaysia’s data center capacity is not in use yet, but factoring in everything under construction, the amount of electricity used just by data centers would exceed the country’s total renewable output in 2022, the latest year for which data is available, a Bloomberg analysis found.

Like Malaysia, Texas has emerged as one of the fastest growing data center markets in the US, thanks in part to the promise of shorter wait times on its independent and deregulated grid. Texas sites can get on the grid in the one to two years it takes to build data centers, said Bobby Hollis, the vice president of energy at Microsoft, which is the largest player in Texas by megawatt, according to DC Byte.

Texas offers plentiful solar power and, in the state’s panhandle, some of the best access to wind power in the world, Hollis said. But a hotter climate and strained water supplies are pushing Microsoft and Google to try weaning their data center cooling gear off water. Alternate approaches, however, require more energy — about 5% more on average, according to Microsoft.

While power in Texas looks plentiful, there are limits. Solar panels and other gear needed for clean power are starting to see some supply constraints, Hollis said. The Electric Reliability Council of Texas also recently cautioned that it has underestimated demand in the San Antonio area, where Microsoft’s big data center campus is located, potentially causing cascading outages statewide in the future.

Back in Virginia, opposition to data centers is growing. At a March supervisors meeting in Prince William County, frustrated residents spoke out against a zoning change that would allow data centers on a specific plot of land to be built about twice as tall. One woman told officials that data centers were turning her quiet neighborhood into a “dystopian nightmare.”

A 48-year-old homemaker named Rachel Ellis spoke remotely to say the change would mean more demand on a grid that’s already strained. “It’s reckless governing to continue to approve data centers without knowing the full impacts of where this electricity will come from,” she said.

After hearing a dozen people speak against the zoning change, the supervisors voted. The bigger data centers were approved.