Todd G. Buchholz, a former White House director of economic policy and managing director of the hedge fund Tiger Management, is author of “The Price of Prosperity: Why Rich Nations Fail and How to Renew Them.”
In 1946, ENIAC — an ancestor of today’s computers — weighed 30 tons, could fill a house and gulped 140 kilowatts just to add numbers faster than a bookkeeper. Today, the phone in your pocket can translate ancient Greek, recognize your face, fly a drone and remind you that Rocky Marciano defeated Joe Louis in the eighth round in 1951. NASA would have been delighted with the first three.
Keep that history in mind as technology companies race to build data centers — the windowless warehouses stuffed with chips, cooling systems and enough electrical gear to shock Nikola Tesla. As politicians from both parties fight over whether to oppose these developments, investors and engineers should wrestle with a different question: What will a data center look like in 10 or 20 years, and what exactly is worth owning?
Gordon Moore supplied an answer 60 years ago. The Intel co-founder observed that the number of components that fit on a chip rises exponentially. Experts have pronounced Moore’s Law dead almost as often as humanities professors have pronounced capitalism dead. Both keep outliving their obituaries. The precise engineering has changed — chips are stacked, cooling improves, software gets leaner — but the direction remains stubborn: We extract more computing from less.
Computing has migrated from a room to a desktop to a laptop and then to a phone. Now it lives in wristwatches, earbuds, doorbells and even devices not suitable for mention in a family newspaper. Why assume the data center has reached its final physical form?
In Ohio, a SoftBank-backed campus is rising, with commitments from OpenAI and Nvidia, hundreds of billions of dollars in long-term contracts and plans for 10 gigawatts of power. (The Washington Post has a content partnership with OpenAI.) Investors and communities locking in 20- and 30-year commitments should ask a sharp question: How much of that development’s land, electricity and equipment will be required to produce a unit of intelligence?
That number will probably keep falling as computing migrates into smaller devices. Your car should not need to consult a 500,000-square-foot warehouse in Kansas before swerving around a pothole.
There is a catch, which a 19th-century English economist spotted long before anyone had heard of Bluetooth. William Stanley Jevons noticed that more-efficient steam engines did not cut Britain’s coal consumption. Because cheaper steam power created more uses for it, Britain needed more coal, not less.
Artificial intelligence could spark a powerful Jevons effect. If intelligence is 10 times cheaper, entrepreneurs may devise 20 times as many uses for it. Hospitals, tractors and household appliances will grow smarter. At some late-night hour, your refrigerator may develop strong opinions about your cholesterol and refuse to surrender the Häagen-Dazs.
Two apparently contradictory things can therefore happen at once: The world can consume vastly more computing while requiring less infrastructure for each computing unit.
That’s why communities should welcome new investment but insist that developers pay the costs of power, water and grid upgrades. They should be especially skeptical of subsidies that assume today’s physical footprint will last for decades.
Done correctly, these investments could end up benefiting nearby residents. A June study by the Electric Power Research Institute estimated that data center growth between 2015 and 2024 modestly reduced U.S. residential electricity prices because large, steady customers spread fixed grid costs across more kilowatt-hours. Of course, if utilities choose to build for a giant new customer and then stick households with the bill, families will pay more. That is not a law of computing; it is a bad contract.
Investors face a different question: What survives when the technology inside the building improves and the building itself begins to resemble a giant box stuffed with fast-aging silicon? Some data centers might start to resemble dreary indoor shopping malls in the internet age — expensive shells hoping for another use, like an indoor skating rink or pickleball courts.
But remember: Although railroads made horses less valuable, rights of way held their value. A site with contracted power, transmission access and fiber optic cables has something far harder to replace than chips: a place at the front of the line to connect to the electrical grid. Securing that access can take years, but data centers might one day serve as a shortcut. The building is just a box; the more valuable asset may be the right to plug in.
America should build data centers along with more energy capacity. AI may prove as transformative as its evangelists claim, and starving it of infrastructure would be foolish, especially as China plows ahead. But it would be equally foolish to draw a ruler from today’s acres, megawatts and cooling towers straight through to 2050.
In 1946, a computer took up the size of a house. Now, it fits in the palm of one’s hand. Don’t bet that the house has won.
The post The data center boom will shrink. That’s good news. appeared first on Washington Post.
