FRESH TAKES ON BIG IDEAS

Can the energy sector deliver the power needed to fuel the AI revolution? That’s a great question. The answer is—well, it’s complicated. 

The need for computational load has been growing steadily since the invention of the internet, and until recently, a host of efficiencies has done a pretty good job of keeping up. But now, at the dawn of generative AI, which demands an unprecedented amount of computational power, innovation is occurring faster than new efficiencies are being created. In some locations outside of the United States, a lack of power is already curbing data center expansion. Amsterdam, Dublin, and Singapore, for example, have placed moratoriums on many new data centers recently, primarily because they lack the power infrastructure to support them.

Power needs for data centers in the United States (the fastest-growing market for data centers) are expected to grow from between 3 and 4 percent of the country’s power demand today to between 11 and 12 percent in 2030. To meet demand, the United States would have to generate 10 percent more electricity by 2030 compared with today. That may not sound like much to ask, but bear in mind that the US power sector has grown little in the past 16 years.

Data centers also need a lot of additional capacity in one place. Historically, five, ten, or even 20 megawatts was a pretty common size for data centers. The average household typically consumes power at a rate of two kilowatts. So in the past, a data center would turn to a local utility and basically propose adding the equivalent of 2,500 to 10,000 additional houses to the grid. But today, hyperscalers—the industry term for builders of lots of data centers—would love to build facilities that draw 200 to 500 or even 1,000 megawatts (also called a gigawatt). That’s akin to asking the local utility to add 100,000 to half a million new houses to the grid all at once. Needless to say, there aren’t too many places left in the country with that much excess capacity.

This situation creates a challenge for the industry, for the nation’s power supply, and for the world’s aspirations for AI. But it also creates an opportunity. The unique thing about finding creative ways to power data centers is that these customers, relative to other big power users, are not very price sensitive. Electricity is only about 15 percent of a data center’s annual operating cost. Owners care more about accessing power fast so they can get facilities up and running. Investors and companies that provide power solutions, therefore, can look at this opportunity in three ways: providing near-term, medium-term, and long-term power solutions. 

“The US data centers’ power needs are expected to grow from between 3 and 4 percent of the country’s power demand today to between 11 and 12 percent in 2030.”

In the near term, investors can funnel investments into utility companies to support the build-out of infrastructure in key markets. There could also be an opportunity for “bridge” solutions as data centers work on longer-term solutions. Some data center owners might pay a relatively large premium for a fuel cell if they can get it quickly and have it placed right next to their facility. Generator sets, small gas turbines, distributed batteries, even mobile-power generation—these are all under consideration to provide power immediately to new data centers.

In the medium term, many hyperscalers have clean-energy commitments and are signing power-purchasing agreements. They are the biggest buyers of clean energy in the United States. There is a challenging situation now where there aren’t enough clean-energy projects coming online. But because there aren’t enough projects, prices are rising—which could be a market signal that says more projects will emerge. While many technologies, including offshore wind, fission, fusion, geothermal, gas carbon capture and storage, and clean fuels, may be able to supply energy in the medium to long term, the bulk of new clean-energy generation is expected to come from solar and onshore wind.

Concurrently, more natural gas power plants are likely needed. They’ve been the main factor in the decarbonization of the US power sector for the past 15 years. Relative to solar and wind, natural gas may not sound so “clean energy,” but it is compared with coal-powered plants. Gas plants are a tricky market, but those who overcome obstacles may be richly rewarded. 

In the long term, significant transmission build-out and investments into fusion companies, small modular nuclear reactors, geothermal, and next-gen battery storage look different in light of power customers who are not very price sensitive. The conversation around nuclear has changed in the past three years. It was once viewed as too expensive to compete economically. Now, hyperscalers are saying they will pay the costs because they need power and they want it to be clean. 

Within the overarching goal of bringing more power online are numerous related opportunities. The data center value chain needs more fiber and power infrastructure in secondary locations, new and emerging technologies that would be used by new power grids, and a scaled talent pipeline to work on all of these projects. 

Like I said, it’s complicated. As with any worthy goal, this one requires lots of creativity, investment, and labor. But to power the AI and digital future, it’s work that must be done.

ABOUT THIS AUTHOR

Jesse Noffsinger is a partner in McKinsey’s Seattle office.