Brookings model sets a $3.725T annual AI revenue hurdle for 2032

Stijn Van Nieuwerburgh’s Brookings model estimates $10.3 trillion in U.S. AI infrastructure investment during 2025–2032. Its central case separately calculates that the resulting capacity would need to generate $3.725 trillion in mature annual revenue by 2032 under specified assumptions. Brookings published its summary on September 23, 2026, and said the paper was presented on September 25.

What the Brookings model estimates

The $10.3 trillion is a modeled estimate of investment made during 2025–2032, not a tally of completed spending. It covers data-center buildings, power systems, networking, specialized chips and other equipment. It also counts some spending on projects that the scenario places into service after 2032.

Across those eight years, the investment averages 3.63% of GDP. In the model, annual investment reaches $2.058 trillion in 2032, or 5.1% of that year’s modeled GDP. These are investment measures: they describe money going toward building infrastructure.

A representative 200-megawatt AI training campus in the paper has an estimated initial capital cost of $8.216 billion: $2.2 billion for the facility, $400 million for incremental power infrastructure and $5.616 billion for installed IT equipment. That example makes the spending categories tangible, but it is a modeled campus configuration, not a price tag for every data center.

Investment and the revenue hurdle are different measures

The central case calculates a requirement of $3.725 trillion in mature annual revenue by 2032, equivalent to 9.2% of the paper’s modeled 2032 GDP. That is revenue the modeled infrastructure would need to generate under the case’s assumptions—not a forecast that AI companies will earn that amount or a measure of current consumer spending.

The 3.63% figure is an average investment-to-GDP ratio across 2025–2032. The 9.2% figure compares one year of required mature revenue with modeled GDP in 2032. Different economic quantities, periods and denominators; they answer different questions.

The assumptions behind the central case

The calculation assumes a 10% unlevered return—a return measured before borrowing costs—and a 50% operating cash-flow margin, meaning half of revenue is available as operating cash flow. It also assigns a six-year economic life to IT equipment, which makes up 68% of the modeled assets, and 20 years to the remaining 32%.

Change the return or margin assumptions and the revenue requirement changes, too. Across the combinations shown in the paper, the modeled values range from $2.291 trillion to $6.724 trillion in mature annual revenue. Those are alternative scenario results, not a forecast range.

How the buildout compares with earlier infrastructure investment

The paper compares average annual capital expenditure as a share of GDP across selected U.S. infrastructure episodes:

The model places the AI infrastructure scenario above these historical benchmarks on this measure. The paper also cautions that the historical sources use different scopes and accounting conventions. In particular, gross investment in short-lived GPUs is not directly equivalent to net capital formation in longer-lived infrastructure.

Financing structures and risks

The financing discussion covers corporate capital spending and private mechanisms including leases, joint ventures, project debt, private credit, securitization and special-purpose vehicles. The $10.3 trillion figure is an infrastructure-investment estimate, not a taxpayer-spending estimate.

Van Nieuwerburgh identifies uncertain demand, technological change, access to power and hardware, and tenant credit quality as risks. He says it would be premature to conclude that AI infrastructure already poses systemic risk comparable to earlier credit booms.