America has built canals, railroads, power grids, interstates and fiber networks. According to a new paper for the Brookings Institution, none of them consumed as large a slice of the economy as the data centers now rising across the country. Columbia University economist Stijn Van Nieuwerburgh estimates that US investment in AI infrastructure will total $10.3 trillion between 2025 and 2032, an average of 3.63% of GDP every year. No single-industry buildout in US history has taken a bigger share.
The paper, Financing the AI Buildout, was released on September 23 as a conference draft for the fall 2026 Brookings Papers on Economic Activity. It is scheduled to be presented on September 25. Its message goes beyond scale. The money paying for the boom is moving out of Big Tech’s audited balance sheets and into a web of leases, joint ventures, private credit and special-purpose vehicles that regulators and investors struggle to see into.
Bigger Than the Railroads
The comparison with history is what made the headlines. From roughly 1870 to 1890, the railroad buildout absorbed about 2.2% of annual GDP, which was the previous record. According to Bisnow’s reading of the paper, the AI buildout is running at more than three times the GDP share that went into the interstate highway system from the 1950s. It is also about six times the share spent on electrification at the turn of the 20th century. IBTimes reported that the telecom expansion of the internet era took about 1% of GDP a year.
“The projected buildout would be larger relative to the economy than the major U.S. canal, railroad, electrification, highway, and telecommunications investment booms,” Van Nieuwerburgh writes.
The $10.3 trillion covers more than server halls. It includes the buildings, power systems, networking gear and specialized chips needed to add about 183 gigawatts of data-center capacity by 2032, according to IBTimes. Roughly 57 gigawatts are installed today. By some measures the model is conservative. Bisnow reported that project-level data puts the full pipeline at 509 GW. Van Nieuwerburgh assumes 227 GW of that will never be built and another 117 GW will arrive after 2032. His working cost figure is about $8.2 billion for every 200 megawatts of compute.
Following the Money Off the Balance Sheet
The biggest spenders are already stretched. Bisnow reported that capex at Oracle, Amazon, Alphabet, Microsoft and Meta grew from $97 billion in 2020 to more than $400 billion in 2025. It is projected to top $800 billion in 2026, which is more than the companies’ combined operating cash flow. Cash-rich hyperscalers could pay for the early phase themselves. The paper argues the next phase depends more and more on outside capital arranged through joint ventures, private credit, securitization, special-purpose vehicles, lease commitments and loan guarantees.
That web is hard to follow. “This is freaking complicated,” Van Nieuwerburgh told reporters, according to IBTimes. He also made a pointed comparison: “This opacity of all these special purpose vehicles is somewhat reminiscent of what happened in the subprime mortgage crisis.”
Those off-balance-sheet structures, the paper warns, depend on cash flows and collateral values exposed to “uncertain AI demand, rapid technological change, timely access to power and hardware, and the continued credit quality of a small number of tenants.” Bisnow reported that at least $1.3 trillion in debt has already been committed to the data-center boom. That is less than half of the roughly $3 trillion tied up in subprime mortgages before the Great Recession. “It’s systemic in the sense that every bank is now involved in this and is actually having a lot of concentrated exposure to AI, but in terms of its magnitude, it’s still relatively modest,” Van Nieuwerburgh said.
Why It Matters
The key figure is the revenue required to justify the spending. IBTimes reported that Van Nieuwerburgh estimates the AI industry would need about $3.7 trillion in annual revenue by 2032 to earn the expected return on the infrastructure. OpenAI and Anthropic together currently bring in around $100 billion a year. Closing that gap would take roughly 80% annual growth. “Silicon Valley wants all of us to believe that this is a miracle technology, it’s going to generate trillions of dollars of revenues, and it has to generate trillions of dollars of revenues to be financeable,” he told Bisnow. “I’m sure there is a state of the world where that happens. I’m just not sure how likely it is.”
The paper does not predict a crash. Van Nieuwerburgh writes that “it would be premature to conclude that AI infrastructure already poses systemic risk comparable to earlier credit booms.” His concern is visibility. Off-balance-sheet structures “may make correlated exposures hard to observe before a downturn.” The main risks are a handful of single-tenant megaprojects, hardware that can become obsolete within a few product cycles, and financing whose structure is still taking shape. If demand disappoints, losses would travel through lenders, insurers, pension funds and private-credit funds, not only through tech stocks. Speaking to Bisnow, Van Nieuwerburgh relied on real-estate history. He expects about five to eight years of strong growth, followed by oversupply, “just like we do in every real estate cycle, and then the prices will collapse. I don’t see why this one time is different.”
His policy recommendation is modest and practical. “The most important policy contribution at this stage may therefore be to improve measurement and transparency while the capital structure of the industry is still evolving,” he writes. He added to Bisnow that “a little bit of gradualism would actually serve us very well.”
What to Watch
Watch the reaction when the paper is presented at the Brookings Papers on Economic Activity conference on September 25, where discussants can challenge its capacity and cost assumptions. Beyond that, the indicators are the ones the paper highlights. The first is whether hyperscaler capex keeps outrunning operating cash flow in coming earnings reports. The second is how much new data-center debt moves through SPVs, securitizations and private credit rather than corporate bonds. The third is whether bank regulators start asking for the exposure data Van Nieuwerburgh says is missing. The most important signal is AI revenue. If the gap between today’s roughly $100 billion and the $3.7 trillion the buildout requires does not start closing, the largest capital bet in US history will come under growing pressure.
“This opacity of all these special purpose vehicles is somewhat reminiscent of what happened in the subprime mortgage crisis.”— Stijn Van Nieuwerburgh, Professor of Real Estate, Columbia University