Brookings prices the AI build-out and cuts 227 gigawatts from the pipeline

A paper from Columbia University economist Stijn Van Nieuwerburgh and the Brookings Institution, as Bisnow reported, puts the artificial intelligence build-out—the largest capital expenditure program in American history—at roughly $10.3 trillion in investment capital between 2025 and 2032, about 3.6 percent of U.S. gross domestic product a year, and concludes the program will likely end in a market correction.

The size comparisons are the paper's own, and they are not small: the railway network built between 1870 and 1890, the next-largest capital expenditure boom in American history, consumed an average of 2.2 percent of gross domestic product, while the projected AI spend runs more than three times the cost of the interstate highway system and six times what electrification cost at the turn of the 20th century. Presented at Brookings' semiannual academic conference, the work carries institutional weight without obliging anyone to mark a portfolio against it, but for real estate capital the useful part is the gigawatts, not the headline number.

Project-level data cited in the paper puts proposed data-center capacity at 509 gigawatts, including compute that would come online after 2032. Van Nieuwerburgh's analysis assumes 227 gigawatts of that proposed capacity never gets built and another 117 gigawatts arrives beyond the end of the decade, leaving the $10.3 trillion to fund an additional 183 gigawatts of computing power by 2032. He prices the build at roughly $8.2 billion for every 200 megawatts delivered, the one figure in the paper a sponsor can hold directly against its own pro forma.

Oracle, Amazon, Alphabet, Microsoft and Meta grew capital expenditure from $97 billion in 2020 to more than $400 billion in 2025, and the five are projected to clear $800 billion in 2026—above their combined operating cash flow. That pace of financing will require the firms to generate enormous revenue growth in the coming years simply to service the debt, the quiet reason a paper about computing power reads as a paper about credit. Capital spending above operating cash flow leaves a funding gap for some other balance sheet to carry, and the paper is explicit that the mechanisms filling it are getting more complex, more opaque, and likely to accelerate as the cycle runs, though it does not say which vehicles those are.

Van Nieuwerburgh's own framing is blunter than the spreadsheets: “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 said. “I'm sure there is a state of the world where that happens. I'm just not sure how likely it is.”

None of that is a call on the current quarter. The paper concedes that demand is outstripping supply for computing power today and predicts a flip only on the strength of historical precedent, and only some years out, which makes private capital's problem one of schedule: money committed now sits against a demand curve the paper expects to invert eventually, so the capacity financed earliest in the queue carries residual risk the announced pipeline does not price.

Read as a real estate story rather than a macro one, the warning lands where information is thinnest. Hyperscaler capex guidance is public, but project-level returns and the financing terms behind a given shell are not, and the coverage of the paper does not supply them. That gap is why a pipeline count of 509 gigawatts circulates at all, and why a study that deletes 227 of those gigawatts is worth more at a credit committee than another demand forecast.

The data-center pipeline rewards sponsors who secure substations rather than steel, and the binding constraint on the roughly $73 billion of construction already underway is grid access; Brookings extends that argument in a direction the market will not enjoy. Institutional capital's share of industrial outdoor storage investment climbed to 45 percent from 30 percent over four years, pulled in by construction staging, demand with an expiration date. Van Nieuwerburgh now sizes the ending: 227 gigawatts of proposed capacity the paper assumes never gets built, and the staging demand that leaves with them. We have also reported on a $400 million water system attached to an Amazon campus in Shreveport, which reads as operators underwriting permanence at a moment when the capital behind them is underwriting options.

The narrower trade runs through power and offtake rather than the correction headline: capacity that clears an interconnection queue with a tenant signed behind it holds value in a credit-led downturn because the tenant is the cash flow, while capacity announced on a site with grid position and no lease is an option on a demand curve the paper's own base case truncates. A sponsor still sizing a development program to 509 gigawatts is underwriting a number this paper has already cut by 227.

Five firms projecting $800 billion of annual capex can absorb an earnings hit; the land basis, the shells, and the unnamed financing behind them cannot. Interconnection throughput is the number to hold over the next two years: capacity that clears a grid queue with a tenant attached gets built even in the down case, and the rest is a substation with nothing behind it.