Columbia model: $9.3 trillion in planned US data centers may never be built

The latest:

Nearly half of the data-center capacity announced across the United States may never be built, according to a draft paper by Columbia Business School real-estate professor Stijn Van Nieuwerburgh. His central scenario puts the unbuilt share at 45% of 509 gigawatts of planned projects — the equivalent of about $9.3 trillion in investment that never materializes, by Barron’s calculations.

Details:

• The pipeline: Data from Cleanview, a power-project repository, showed roughly 509 gigawatts of new US data-center capacity planned as of July. That is almost nine times the capacity operating today, a gap that sits at the center of Van Nieuwerburgh’s attempt to model how much of the announced buildout is real.
• The central estimate: Van Nieuwerburgh’s central scenario projects that 45% of the 509 gigawatts will never be built. Separately, nearly a quarter of the planned capacity would not come online until 2033 or later, meaning the announced pipeline is both smaller and slower than the headline figure implies.
• The math: The professor estimates each 200-megawatt AI campus costs about $8.2 billion to build. Applying that cost to the capacity his model leaves unbuilt produces roughly $9.3 trillion in unrealized investment, a figure calculated by Barron’s rather than stated in the paper itself.
• The upside case: Even after writing off nearly half the pipeline, the model still projects $10.3 trillion of data-center investment through 2032. That averages about 3.6% of US gross domestic product a year, a scale the paper frames as the largest capital-expenditure boom in American history.
• Cost inflation: Van Nieuwerburgh notes the $9.3 trillion shortfall is a conservative reading. Once rising construction costs are factored in, he estimates the value of the investment that never happens would be larger still, because each campus built later in the cycle carries a higher price tag.
• Status of the work: The findings come from a draft paper, not a finished peer-reviewed study, and Van Nieuwerburgh researches and teaches real estate at Columbia rather than semiconductors or cloud computing. The paper does not name which specific announced projects it expects to fall away.
• Companies in frame: Barron’s flagged Oracle, Alphabet and Amazon alongside the report, three of the largest announced builders of AI computing capacity. No projected cancellation figures were attached to any individual company.
• The announcement pace: New data-center projects are being announced at close to a daily rate during the current AI boom, a pace that drives the gap between the 509 gigawatts on paper and the far smaller base in operation.

Between the lines:

The model’s two headline numbers point in opposite directions and are meant to. A 45% failure rate implies announcement inflation on a historic scale, while the surviving $10.3 trillion still equals about 3.6% of annual US output. The paper’s reading is that the boom can be both massively oversold and genuinely enormous at once.

What’s next

Watch whether the draft paper is published in final form, whether updated Cleanview capacity data shifts the 509-gigawatt baseline, and whether Oracle, Alphabet or Amazon revise announced data-center plans in upcoming capital-spending disclosures.