At this point in the artificial-intelligence boom, it seems like a new data center is announced every morning. Many will never see the light of day, according to one Columbia Business School professor.

A draft paper from Stijn Van Nieuwerburgh, who researches and teaches real estate at Columbia, attempts to model just how real the data-center buildout is. His answer: Even if it is the largest capital-expenditure boom in U.S. history, trillions of dollars worth of planned projects may never materialize.

July data from Cleanview, a power-project repository, showed roughly 509 gigawatts of new data-center capacity planned in the U.S.-almost nine times the amount in operation today. Of that 509 gigawatts, Van Nieuwerburgh's central scenario estimates that 45% will never be built and nearly a quarter won't come online until 2033.

Each 200-megawatt AI campus costs about $8.2 billion, Van Nieuwerburgh estimates. That means the unbuilt facilities in his model could amount to $9.3 trillion in investment that is never realized, according to Barron's calculations. After accounting for rising construction costs, the number would be even larger.

The estimate is a reminder to investors that many planned data centers could face community backlash, construction delays, or abandonment. The setbacks don't mean the AI boom is overstated, but they could put individual projects and companies at risk.

"I think we're in the middle of an arms race, where everyone's building capacity," Van Nieuwerburgh tells Barron's. Some companies won't survive.

Van Nieuwerburgh's central scenario sees about 300 gigawatts of currently planned projects getting built, triggering $10.3 trillion in investment through 2032. That would average about 3.6% of U.S. GDP a year-larger than the railroad, highway, and telecom buildouts of the past.

These numbers aren't an official forecast. They are one scenario, based on assumptions about which projects are most viable. Data centers scheduled to come online in 2026 have a 90% chance of being completed in Van Nieuwerburgh's model. The probability drops five percentage points each year, with another discount if the developer hasn't disclosed the project's size.

The reasoning is simple: The longer the timeline, the more can go wrong, especially in the rapidly-evolving world of AI. Construction costs rise, computation technologies could change, and the companies involved in a given project may implode, Van Nieuwerburgh explains.

"If Oracle is planning to build something in 2032, well, who knows if Oracle will be around in 2032?" he asks.

The Oracle example was timely: Shares fell last Thursday after reports that the cloud provider was seeking to shield itself against potential delays at a planned data center in New Mexico. An Oracle spokesperson later said in a statement that the project remained on schedule.

Van Nieuwerburgh's paper, presented at a Brookings Institution conference last week, comes at a pivotal moment in AI. Anthropic is expected to go public in the coming weeks, raising tens of billions of dollars to help support its enormous data-center rental costs. OpenAI may follow in 2027.

Recent moves by Anthropic and OpenAI to slow aspects of AI development over safety concerns could weigh on the need for capacity, or at least make the growth in demand choppier.

Meanwhile, some cloud providers, such as Alphabet and Amazon.com, are turning cash-flow negative because of their data-center investments. They are increasingly financing new projects with project debt, joint ventures, private credit, and other vehicles.

That shift makes data centers "a major institutional real estate and infrastructure asset class," Van Nieuwerburgh writes. But it also spreads the underlying risks of a project more and makes them more difficult to track.