AI Data Center Build-Out to Cost $10.3 Trillion, Brookings Study Warns of Market Correction

The infrastructure push will consume 3.6% of U.S. GDP annually through 2032, exceeding railway and highway build-outs combined, with at least $1.3 trillion in debt already committed.

By the Family Office Real Estate Daily Desk·Thursday, September 24, 2026·2 min read

Editorial summary of reporting byBisnow Our editorial standards →

The answer · checked against Bisnow

How much will the AI data center build-out cost and what is the risk of a market correction?

A new paper by Columbia University economist Stijn Van Nieuwerburgh and the Brookings Institution estimates the AI infrastructure build-out will cost $10.3 trillion between 2025 and 2032, equal to roughly 3.6% of U.S. GDP per year. At least $1.3 trillion in debt has already been committed. Van Nieuwerburgh warns that historical precedent points toward eventual oversupply and a market correction.

Key facts

The Brookings Institution paper, co-authored by Columbia University economist Stijn Van Nieuwerburgh, estimates the AI infrastructure build-out will require $10.3 trillion in investment capital between 2025 and 2032.
The $10.3 trillion price tag represents roughly 3.6% of U.S. gross domestic product per year, according to the Brookings Institution analysis.
The projected cost of the AI build-out is more than three times what was spent to build America's highway system and six times more than was spent on electrification, according to the Brookings Institution paper.
The major hyperscalers — Oracle, Amazon, Alphabet, Microsoft and Meta — have grown their combined capex from $97 billion in 2020 to more than $400 billion in 2025, and are projected to clear $800 billion in 2026, according to the Brookings Institution analysis.
At least $1.3 trillion in debt has already been committed to underwriting the data center boom, with capital sources including private credit, insurers and pension funds, according to the Brookings Institution paper.
Stijn Van Nieuwerburgh said the $10.3 trillion build-out assumes a cost of roughly $8.2 billion for every 200 megawatts of compute power built, with 183 gigawatts of additional computing power expected by 2032.

AI Data Center Build-Out to Cost $10.3 Trillion, Brookings Study Warns of Market Correction — Image: editorial illustration · Story sourced from Bisnow

The artificial intelligence infrastructure build-out will require roughly $10.3 trillion in investment capital between 2025 and 2032, consuming 3.6% of U.S. gross domestic product per year, according to a new analysis from Columbia University economist Stijn Van Nieuwerburgh and the Brookings Institution. The projection, presented at the Brookings Institution's semiannual academic conference, estimates the capital will fund an additional 183 gigawatts of computing power by 2032.

The scale of the AI build-out is unprecedented in American history. The next-largest capital expenditure boom came from 1870 through 1890, when the U.S. directed an average of 2.2% of gross domestic product toward the railway network, the analysis found. The projected AI cost is more than three times what was spent to build out America's highway system and six times more than was spent on electrification at the turn of the 20th century.

At least $1.3 trillion in debt has already been committed to underwriting the data center boom. The debt total is less than half the $3 trillion that was tied up in the subprime mortgage crisis that led to the Great Recession, Van Nieuwerburgh said. The sources of capital are increasingly diverse as private credit pushes in along with insurers and pension funds, while banks hitting lending ceilings are using syndication and other financing vehicles to keep deploying capital into the space.

The major hyperscalers — Oracle, Amazon, Alphabet, Microsoft and Meta — have grown their capex from $97 billion in 2020 to more than $400 billion in 2025. They are projected to clear $800 billion in 2026, more than the firms' combined operating cash flow. The paper assumes it costs roughly $8.2 billion for every 200 megawatts of compute power built.

"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," Van Nieuwerburgh said. "I'm sure there is a state of the world where that happens. I'm just not sure how likely it is."

Debt that looks manageable in the trade tape almost always looks thinner once a forced seller appears, family office advisor Jaf Glazer has cautioned.

The increasingly complex network of financing structures that developers, tech firms and capital sources are leveraging to finance new construction is making it harder to track exposure in the marketplace, with private credit and off-balance-sheet structures adding a layer of opacity that conceals where risk lies. The estimates for project completions are in some ways conservative, with project-level data suggesting the pipeline totals 509 gigawatts, including compute power that will come online after 2032. In the analysis, Van Nieuwerburgh assumed 227 gigawatts of proposed capacity will never be built, while another 117 gigawatts will come online after 2032.

"If history is a guide, credit constraints will loosen, more speculative development will take place, more marginal compute will be built, and sooner or later we're going to have oversupply, just like we do in every real estate cycle, and then the prices will collapse," Van Nieuwerburgh said. "I don't see why this one time is different." Demand is outstripping supply for computing power, but historical precedent makes it likely that dynamic will one day flip, even if that is several years away. Van Nieuwerburgh estimated roughly five to eight years of strong growth before oversupply becomes a concern.

The Deployment Angle

Family Office Real Estate Daily Desk · our analysis, not the source's

The arithmetic family offices must run starts with the revenue requirement. If hyperscalers are deploying $800 billion in capex in 2026 against operating cash flow they have already exceeded, the implied revenue ramp to service that debt load is steep. At a 10% cost of capital, $10.3 trillion requires roughly $1 trillion in annual debt service by 2032. That is revenue Big Tech must extract from AI applications that do not yet exist at commercial scale. Family offices underwriting co-GP positions alongside data center sponsors should model what happens if that revenue does not materialise on the timeline the debt assumes.

The platform capital route — taking minority stakes in operating portfolios or development platforms — faces a different risk. The large number of massive single-tenant facilities presents concentrated credit exposure if one hyperscaler faces a liquidity crunch. A family office backing a platform with three hyperscaler tenants across five facilities is not truly diversified if all five leases reprice simultaneously in a downturn. The analysis suggests five to eight years of strong growth before oversupply becomes a concern, but the opacity of private credit structures and off-balance-sheet vehicles makes it harder to see where the marginal debt sits. Underwrite tenant credit as if it were a single counterparty, not five separate bets.

Direct ownership through separate accounts or programmatic joint ventures offers more control but requires longer hold periods to survive a correction. If the cycle follows historical precedent and oversupply arrives within eight years, a family office buying stabilised data center assets today at a 5% cap rate must assume those assets could trade at 7% or higher in a repricing event. The gap between $8.2 billion per 200 megawatts of new supply and the replacement cost of existing facilities narrows quickly once speculative development floods the market. Family offices should avoid levered bets on continued cap rate compression and instead structure acquisitions with enough equity to hold through a full cycle without a refinancing event.

The comparison to the subprime mortgage crisis is instructive. The $1.3 trillion in debt already committed is less than half the $3 trillion that was tied up in subprime, but the distribution of risk matters more than the headline figure. If every bank now has concentrated exposure to AI infrastructure and private credit has layered in additional leverage that does not appear on balance sheets, the trigger for a correction could be smaller than the total debt stock suggests. Family offices should price in the possibility that exit liquidity disappears before the eight-year window closes and structure positions accordingly.

Questions this story answers

01How does the AI data center build-out compare to past U.S. infrastructure booms?

The Brookings Institution paper found the next-largest capital expenditure boom was the U.S. railway build-out from 1870 through 1890, which averaged 2.2% of GDP. The AI build-out's projected cost is more than three times what was spent on America's highway system and six times more than was spent on electrification at the turn of the 20th century.

02How much debt has already been committed to AI data center development?

At least $1.3 trillion in debt has already been committed to underwriting the data center boom, according to the Brookings Institution paper. Stijn Van Nieuwerburgh noted this is still less than half the $3 trillion tied up in the subprime mortgage crisis that led to the Great Recession, though he said every bank now has concentrated exposure to AI.

03What are the main risks that could cause a correction in AI data center investment?

Stijn Van Nieuwerburgh identified several risks: the large number of massive single-tenant facilities creates credit risk if a hyperscaler faces a liquidity crunch; new data center hardware develops so quickly that new builds can become obsolete; and complex financing structures involving private credit and off-balance-sheet vehicles are making it harder to track where risk lies in the marketplace.

04When does the Brookings study expect AI data center oversupply to become a problem?

Stijn Van Nieuwerburgh estimated roughly five to eight years of strong growth before oversupply becomes a concern. He said historical precedent suggests credit constraints will loosen, more speculative development will occur, and oversupply will eventually follow, causing prices to collapse, as happens in every real estate cycle.

05How large is the total pipeline of AI data center projects beyond the 2032 forecast?

Project-level data cited in the Brookings Institution paper suggests the total pipeline totals 509 gigawatts. Stijn Van Nieuwerburgh assumed 227 gigawatts of proposed capacity will never be built and another 117 gigawatts will come online after 2032, leaving 183 gigawatts as the basis for the $10.3 trillion cost estimate through 2032.

Browse every question the desk has answered →

Original reporting

Bisnow

Read the original at Bisnow →

data-centersartificial-intelligencedebt-marketsmarket-correctionhyperscalers

Peer Network · By Invitation

The Thesis Exchange

Share an investment thesis in confidence. We pair you anonymously with up to two other family offices running adjacent strategies. Reviewed by Gallium's editorial team. No vendor pitch.