A Columbia Business School analysis by professor Stijn Van Nieuwerburgh estimates the buildout could involve $10 trillion in investment and approach 4% of U.S. gross domestic product. That would dwarf the roughly $173 billion, in current dollars, spent on railroad construction from 1870 to 1890, which amounted to more than 2% of annual economic output over two decades, according to the analysis.

The scale of the spending is reshaping construction investment. U.S. data center construction value more than doubled over two years to $85 billion annually, while the value of other construction fell 8%, according to an analysis of Census Bureau data through August. Since 2024, data center construction value rose by $48 billion as the value of new houses and apartments declined by $29 billion.

The shift has prompted concern that a narrow, capital-intensive boom could displace investment in projects that employ more people. “Because data centers employ so few workers, the vast spending on AI currently taking place will also significantly reduce the demand for labor,” economist Paul Krugman wrote in an Oct. 11 analysis. Data centers do create construction jobs, but their ongoing operations require comparatively few workers.

Josh Bivens, chief economist at the Economic Policy Institute, described the economy’s reliance on AI spending as “a narrow base of growth, basically one sector, and a pretty fragile one.” If expectations for future AI profits weaken, he said, investment could be cut sharply. The International Monetary Fund has also warned that disappointment over AI’s prospects could trigger a stock-market correction.

For communities hosting the facilities, electricity and water demands are immediate concerns. Van Nieuwerburgh estimates the buildout could double electricity consumption. In Oregon, utility affordability was the most frequent concern raised in public comments reviewed for a state report, and the state continued a pause on approvals for data center projects.

Some states are reconsidering incentives they offered before the AI-driven construction surge. Nevada Gov. Joe Lombardo’s September executive order trimmed tax abatements and said data center operators would be asked to cover their energy and water needs. Virginia, meanwhile, began applying an electricity-consumption surcharge to data centers in July while retaining major tax breaks.

Industrial Info Resources projects that data center costs over the next three years will be concentrated in Texas, at $638 billion, followed by Virginia at $249 billion, Ohio at $169 billion and Georgia at $160 billion. The projections underscore how investment and its demands are concentrated in particular states, even as the consequences for power systems and public budgets can extend beyond the facilities themselves.

The boom’s profitability remains uncertain. “It is not a pretty picture right now,” said David Pickering, senior vice president of research at Industrial Info Resources, adding that profitability may improve as the industry settles and regulations governing AI develop. He said major companies will likely find ways to make the investments profitable.

The comparison with railroads is not only about scale. Historian and author Mark Higgins noted that railroad construction attracted large amounts of foreign capital before financial panics left distressed rail lines to be bought by investors. The current buildout, Van Nieuwerburgh’s analysis suggests, is much larger relative to the economy than earlier waves of infrastructure construction, including canals, highways and telecommunications.

The central question for states and local communities is whether the promised economic returns will justify tax incentives and the strain on energy and water systems — particularly if AI revenues fail to match the cost of building and powering the facilities.