A Data-Center Crunch is Coming

AI needs lots of data centers and electricity. The world soon won’t have enough of either.

Although there are many constraints on AI’s growth—a shortage of memory chips, an uncertain return on investment, rising interest rates—the dearth of data centers looks like the more significant and immediate one.

A recent Morgan Stanley analysis estimated global demand for data-center capacity will be about 56 gigawatts next year, with just 51 gigawatts of new facilities being energized. It gets worse in 2028, when estimated demand climbs to 86 gigawatts, but only 68 gigawatts of computing power come online.

Those are huge amounts of new power to expect electric operators to provide, and data-center developers to accommodate. In the next two years, the combined electricity demand is enough to power around 118 million homes. There are about 150 million homes in the U.S.

Of course, the companies most affected by the coming shortfall—tech giants, AI labs, chip-makers and data-center developers—have levers they can pull to prevent it from slowing their expansion.

Innovation is one. Advances in how AI calculations are made could reduce the amount of energy required. More advanced chips should also be able to do more number-crunching with less energy.

Data-center developers could also move more quickly toward power sources like fuel cells or natural-gas turbines that provide electricity directly to facilities. That avoids some of the time and regulatory hassle associated with linking up to the grid.

Indeed, one big consequence of the data-center shortfall will be more money going into so-called behind-the-meter energy, Morgan Stanley’s analysts said. But a lot would have to go right to keep the AI boom humming.

In his own recent dive into the energy situation, UBS semiconductor analyst Tim Arcuri concluded that enough data centers could theoretically be built to house all the AI chips companies like Nvidia want to sell through 2030.

Yet getting there would require most of the roughly 110 gigawatts of possible behind-the-meter power to come online. Most of the roughly 150 gigawatts of grid-connected power plants that have yet to receive regulatory approval would also have to get built.

In the nearer term, Arcuri estimated, covering the energy shortfall would require just around half of early-stage power projects to be converted into on-grid supply. This scenario, he said, was “not impossible but also high enough to require an accommodative legislation environment and limited supply chain delays.”

Is that panning out? Not really. Political opposition to data centers is mounting. And congressional efforts to oversee their growth more closely could intensify if Democrats take power after the midterm elections.

Regulatory and supply-chain delays are also biting big projects. Oracle is trying to delay making full lease payments on a huge data-center facility in New Mexico amid power and permitting issues.

One possible source of comfort is that chip-makers aren’t yet blaring their sirens about a capacity squeeze. Annual revenue outlooks for Nvidia and Broadcom, Morgan Stanley notes, take into account those companies’ assessments of data-center availability.

There is reason to wonder about that confidence, though. Chip-makers can’t reliably predict supply-chain delays or political problems. And the AI industry can’t even agree on how much data center capacity is being built: a Bernstein Research analysis found forecasts of U.S. data-center capacity ranged from 60 gigawatts to 180 gigawatts, “with little consensus even among credible industry observers.”

That’s a lot of uncertainty for a build-out the AI boom’s future depends on.

Hackers using a Chinese AI tool attacked South Korea’s biggest banks and stole the personal information of 68,000 people, showcasing the danger AI agent incursions could pose to the global financial system. The recent cyberattack follows numerous other recent agent-driven hacks, starting with an OpenAI agent hack of Hugging Face in July. Those attacks have raised security concerns around the globe and led leading AI labs to suggest a pause in model development.

Percentage of businesses that were able to accurately forecast AI spending in a recent study.

Concerns about returns on big investments in AI haven’t slowed down the tech giants and AI labs plowing trillions of dollars into the technology.

But here’s a sobering estimate: to justify all the cash being invested in AI, American businesses and consumers would have to spend about as much on AI in 2032 as they do on food.

Is that possible? There is a case that it could be, if AI joins labor and capital as a third factor of production—if, in other words, it completely transforms the economy.

But the history of computing suggests another likely outcome: that people and companies will eventually exhaust the obvious productive uses for the technology, after which additional dollars spent on AI will yield fewer gains.

U.S. soldiers are testing out smartphones and an array of other new battlefield technology to adapt to rapid changes in how wars are fought. The Army program aims to put cutting-edge AI, mapping and communication technology in soldiers’ hands at a time when cheap drones and precision missiles force troops to be more spread out and nimble.

WSJ’s Tech: California event returns to Napa Valley Nov. 3-4. We’ll explore the latest tech industry news with Sequoia Capital partner Pat Grady, Zoox CEO Aicha Evans, Take-Two Interactive CEO Strauss Zelnick and more. If you’re interested in attending, request an invitation.

WSJ AI & Business is a weekly look at AI’s transformation of the business world. This newsletter was curated and edited by Asa Fitch. Reach him at asa.fitch@wsj.com (if you’re reading this in your inbox, you can just hit reply). Got a tip for us? Here’s how to submit.