A paper presented at the Brookings Papers on Economic Activity conference in September 2026 offers one of the clearest pictures yet of how the AI data center boom is being financed, and where the risk ends up. Its author, Stijn Van Nieuwerburgh of Columbia Business School, looks at the United States. The lessons travel well to Europe.
A BOOM MEASURED IN PERCENT OF GDP
The paper estimates that a 1 GW AI campus costs around $41 billion. For a 200 MW campus, about one third goes into the building and power infrastructure and two thirds into IT equipment. Under the author's central scenario, US investment would average about 3.6% of GDP a year over 2025–2032, a larger share of the economy than the railroad, electrification, highway or fibre booms at their peak.
WHEN CASH FLOW IS NO LONGER ENOUGH
According to the paper, 2026 is the first year in which the capital expenditure of the five largest cloud and tech groups exceeds their combined operating cash flow. The financing is therefore shifting toward leases, joint ventures, project debt, private credit and special-purpose vehicles. Users of compute are being separated from the owners of the assets, and increasingly from the investors who carry the risk.
WHAT A MEGA-FINANCING LOOKS LIKE
The paper takes Meta's Hyperion campus as its case study. A project vehicle raised about $27 billion of investment-grade debt against an asset of about $30 billion, a debt-to-asset ratio of roughly 90%, with a debt service coverage ratio of only 1.12. Meta's flexibility comes from short renewable leases backed by a residual value guarantee, commitments that do not appear on its balance sheet until they take effect. The author also cites estimates of several hundred billion dollars of future hyperscaler lease commitments not yet recognised on balance sheets.
Asset value: about $30 billion
Project debt: about $27 billion, rated A+
Debt-to-asset: about 90%
Debt service coverage: 1.12
WHERE THE RISK SITS
The paper identifies three main vulnerabilities: concentration on a few tenants, technological obsolescence of equipment that ages faster than the debt financing it, and execution. The last one covers power access, transmission, permits, local opposition and hardware supply. The author's point is precise: the risk is not only that costs rise, but that revenue starts later than assumed at underwriting, or never.
This is the layer that standard technical due diligence often covers least. With debt at around 90% of asset value and thin coverage ratios, a year's delay on a grid connection or a suspended permit weighs directly on lenders.
ANNOUNCED IS NOT BUILT
The paper's pipeline analysis tells the same story from another angle. Of a little over 500 GW of US data center projects announced, its central scenario sees about 183 GW operational by 2032, about 117 GW completed later, and about 227 GW never built. The author presents this as a scenario, not a forecast. See our analysis Compute ordered is not compute installed.
WHAT IT MEANS FOR EUROPE
The same financing tools are arriving in Europe: infrastructure funds, private credit, joint ventures and project debt. The execution risks are present too. In France, RTE had reserved close to 18 GW for around 80 data center projects by May 2026, part of which will not be built. Grid commitments are due at reservation, building permits can be challenged, and local opposition is starting to organise nationally. Lenders who underwrite tenant credit without underwriting the delivery schedule are taking a risk they are not measuring.
Is high-voltage equipment ordered, with delivery before that date?
Are all authorisations final and clear of challenge?
How exposed is the project to organised local opposition?
How much delay can the debt service coverage absorb?
A NOTE OF CAUTION
The paper is a conference draft, and its figures may change. Its estimates are scenarios, not forecasts, and they concern the United States. The author does not conclude that AI infrastructure poses imminent systemic risk; his main concern is visibility, knowing who actually bears which risk. That is exactly the question a project-level due diligence should answer. See our due diligence for lenders and funds.
Source: Stijn Van Nieuwerburgh (Columbia Business School), Financing the AI Buildout, Brookings Papers on Economic Activity, conference draft, September 2026.
FREQUENTLY ASKED QUESTIONS
How much does a 1 GW AI data center campus cost?
According to a September 2026 Brookings conference paper, about $41 billion, with roughly one third for the building and power infrastructure and two thirds for IT equipment.
How was Meta's Hyperion data center financed?
Through a joint venture that raised about $27 billion of investment-grade project debt against an asset of about $30 billion, a debt-to-asset ratio of roughly 90%, supported by Meta leases and a residual value guarantee.
What is execution risk in data center financing?
The risk that a project is delayed or never completed because of power access, grid connection, permits, local opposition or equipment supply, so that revenue starts later than assumed or never.
Does the Brookings paper say AI infrastructure is a bubble?
No. The author does not conclude that systemic risk is imminent. He highlights uncertain demand, rapid obsolescence, execution bottlenecks and high asset-level leverage, and stresses the lack of visibility on who bears the risk.