AI Could Upend Electricity Demand
AI’s energy needs are enormous, and could have a once-in-a-generation impact on the electricity market.
Sign up for smart news, insights, and analysis on the biggest financial stories of the day.
Generative AI is the future. It’s also an energy glutton. That’s potentially bad for our future. Training and running AI systems takes a heck of a lot of compute power, and compute power needs energy — mostly in the form of electricity. The International Energy Agency predicts electricity consumption from data centers could double from 2022 to 2026 to 1,000 terawatt-hours, roughly the same amount of electricity consumed yearly by Japan.
It’s another big infrastructure challenge for Big Tech. The energy it needs to realize its AI dreams simply does not exist. The industry has overcome its infrastructure challenges before — Google, Meta, Amazon, and Microsoft have spent years laying subsea internet cables as data highways. But setting up new energy sources takes time, which is frustrating for an industry that’s used to moving fast and breaking things. Still, Big Tech is pitching in and trying to grab as much forward capacity as it can, and as electricity becomes scarcer and scarcer, companies will be competing not just with one another but with everyone else who wants to turn on a lightbulb.
Hungry, Hungry Servers
Although the IEA has had a go at quantifying how much more electricity the AI sector will eat up in the next few years, it’s extremely hard to know what the upper limit will really be. Deborah Perry Piscione, a former Silicon Valley VC, told The Daily Upside that it’ll be a few years until the full scale of AI’s energy demand really comes into focus. She added that in the US, the electrical grid is not equipped to deal with any massive extra load. “Generating the power is one thing, and then the distribution is something totally different,” she said.
Evan Caron, co-founder and chief information officer at climate-tech and energy transition incubator Montauk Climate, agreed that it’s very hard to know exactly how high the energy demand from AI might climb, adding that it partly depends on whether the fervor surrounding the technology persists. This week showed that investor enthusiasm for AI is not bottomless, as Nvidia lost over $550 billion in market cap over a three-day period.
But even if investors cool on AI, the tech companies that want to be players in the AI space are already thinking years ahead to shore up the energy they’ll need for data centers, pouring billions into deals with energy companies and utilities:
- Microsoft reached a deal with Brookfield Asset Management last month to inject $10 billion into renewable electricity projects.
- At the end of May, Amazon, Google, and Microsoft all signed deals with Duke Energy proposing new energy tariffs to develop clean energy sources in the southern US.
- Earlier this month, Google agreed to partner up with Nevada utility company NV Energy and Warren Buffett-backed geothermal energy startup Fervo Energy.
The Nuclear Option: Microsoft and its partner Sam Altman, the tech CEO du jour who heads OpenAI, have been particularly bullish on nuclear energy, investing in companies that promise to make small modular reactors: a new kind of nuclear power plant that’s theoretically smaller, faster, and cheaper to build. SMRs have yet to prove themselves as a viable part of the energy supply chain, however.
None of the Big Tech companies have managed to get much of a head start on capturing forward capacity, Caron told The Daily Upside. “They’re all scrambling for megawatts,” he said. He predicted competition for energy hook-ups will only get hotter, adding, “They are going to be fiercely competitive on acquiring electrons and molecules of natural gas. And I personally think they’re going to try to get it from any source they possibly can.”
There’s already evidence to suggest that AI energy demand is giving the fossil fuel sector, which is supposed to reach peak demand by the end of the decade, a little boost. Microsoft’s emissions climbed almost 30% from 2020 to 2023, and US coal concerns are stretching out timelines for ramping down their operations in response to the demand. Both Caron and Piscione said they believe the AI boom will hamper global targets to wean ourselves off of climate-damaging fossil fuels.
Less Than Friendly Competition
As Big Tech signs more and more deals with energy firms, electricity will become harder to come by, Caron said. “Resource constraint is the name of the game,” he told The Daily Upside. There’s even competition for AI compute power within tech companies. Bloomberg reported earlier this month that teams at Google’s DeepMind unit focused purely on research have a tighter compute supply than teams focused on more commercial applications.
But Big Tech firms are still competing with anyone else who wants to build something — e.g., houses, schools, and hospitals. The waiting lists are pretty long too, as getting a new power station online and feeding into the grid is a pretty lengthy process. “There’s a physical limitation on how much can be built and when,” Caron said, adding: “Despite the increase in demand, the supply chains have not changed materially over the last couple of years.”
Even demand for data centers pre-AI boom was driving this kind of competition; the Financial Times reported in 2022 that developers in west London were facing a ban on building new housing until 2035 because a blossoming of new data centers in the area had maxed out the grid’s capacity for the next decade.
With AI in the picture, residential competition could get more dire, Piscione said. “It could really lead to the supply-demand imbalances, especially in regions with limited energy infrastructure, or where AI development is concentrated,” she said. “It absolutely is going to impact and hamper economic development in some regions, because data centers end up snapping up most of the spare capacity in areas with constrained transmission or generation and it could make it harder or more expensive for new homes or businesses to secure the power connections that they need.”
When it comes to signing deals with energy utilities, there’s no prioritization process: If you have the money, then it’s first come, first served. “Everything’s electrifying. It’s all competing,” Caron said. “Right now a good political question is: If a hospital and an elder care facility ask for an interconnection at the same time as Sam Altman’s OpenAI data center, who wins? Right now, it’s Sam Altman’s OpenAI data center, because they have more money.”
Regulation Stations
What happens when a crucial resource’s demand outstrips its supply is a big, dark question-mark over the future of electricity. “We were in a place globally where energy was getting cheaper and cheaper and cheaper because of renewables,” Caron said. “I fear that that won’t be the case for the next generation.”
Policymakers have been eager to talk about how to regulate AI’s output — the disinformation, the spam, the copyright issues — but there has not been much discussion about its energy-guzzling. Piscione said if the US wants to maintain its pole position as a leader in AI, policymakers need to get involved in preparing the nation’s infrastructure. “I do think a lot of the onus is going to fall on chipmakers. Working and figuring out something with [Big Tech] companies with a subsidy from the government,” she said, warning that “it is going to be a massive battle working with policymakers.”
Caron thinks multi-stakeholder co-operation and fast, top-down policymaking will be key to managing any kind of orderly transition. He believes it’s possible that energy market regulation could include remedies such as moratoriums on new interconnections for AI or laws that give prioritization to critical infrastructure over data centers.
“I think the genie’s out of the bottle on AI,” Caron said. “It feels like it’s moving towards dystopian versus utopian, I think, but it’s on a knife’s edge. It could be very utopian if they figure out how to do it.”