AI may look like software, but building the capacity behind it looks more like a giant infrastructure boom. In this Brookings conference draft, economist Stijn Van Nieuwerburgh estimates that a 200-megawatt AI campus costs about $8.2 billion: roughly two-thirds for computing equipment and one-third for the building and power infrastructure.

His central scenario assumes about 183 gigawatts of additional US data-center capacity comes online by 2032. Including spending on projects completed later, that implies $10.3 trillion invested over 2025–32, averaging 3.63% of GDP a year. The paper compares that share with railroads, highways and the telecom boom — but emphasizes that this is a scenario, not a forecast. Much of the announced project pipeline has no completion date, and the model assumes substantial capacity is never built.

The financing is changing along with the scale:

  • The paper projects 2026 capital spending by Oracle, Microsoft, Amazon, Meta and Alphabet will exceed their combined operating cash flow. More of the buildout must draw on outside capital.
  • Developers and investors own facilities and lend against future lease payments, while hyperscalers rent the capacity. That can keep debt off a technology company’s balance sheet without making its economic commitments disappear.
  • Meta’s Hyperion project is the example: the paper describes roughly $27 billion in debt against a $30 billion facility, backed by Meta leases and a promise to cover certain shortfalls if it walks away. The high leverage sits at the project level.

The sharpest test is whether the new capacity can earn its keep. Under the paper’s assumptions — a 10% required return, a 50% operating cash-flow margin and a six-year economic life for IT equipment — the capacity completed by 2032 would need about $3.7 trillion in mature annual revenue. That is a hurdle implied by the model, not a prediction of 2032 AI sales. If the hardware lasts only three years, the required revenue rises to about $6 trillion.

The takeaway is less “an AI crash is coming” than “follow the obligations.” Power delays, weak demand, falling GPU prices or faster obsolescence would not land neatly on a single company’s books. Leases, guarantees, project debt and a small set of shared tenants can spread the loss through investors and lenders while making the total exposure harder to see.