The AI industry is building like it expects to become a pillar of the economy. A new Brookings paper has worked out what that would actually take.
The estimate: AI must generate about $3.5 trillion a year by 2032 to justify the current data center expansion. The paper’s detailed math lands even higher, at approximately $3.725 trillion annually.
The math behind the buildout
The paper comes from Stijn Van Nieuwerburgh of Columbia Business School. It was published through the Brookings Papers on Economic Activity, better known as BPEA.
Van Nieuwerburgh estimates US AI infrastructure investment will total $10.3 trillion between 2025 and 2032. That works out to an average of 3.63% of GDP every year.
The paper finds this spending notably surpasses earlier US infrastructure booms, including the railroads and the interstate highways.
The paper uses a 10% unlevered return as its benchmark. To hit that bar, the AI sector would need roughly $3.725 trillion in annual revenue by 2032. That would equal approximately 9% of projected US GDP in that year.
From $100 billion to trillions
OpenAI and Anthropic, the two best-known AI labs, currently bring in about $100 billion in combined revenue, according to the research.
Getting from there to the required level would mean sustained growth of about 80% per year through 2032.
The paper projects the US will add roughly 183 GW of computing capacity by 2032. The research describes that as doubling current residential electricity consumption.
Van Nieuwerburgh presented the findings at the Brookings BPEA fall conference, held September 23 to 25, 2026. A follow-up podcast discussion is scheduled for October 8, 2026.
Why the comparison to railroads matters
The research also flags how this buildout is being funded. Much of it relies on off-balance-sheet financing structures, which keep obligations away from the main corporate ledger. That arrangement makes it harder for outsiders to see who is carrying the risk.
What this means for investors and the AI trade
For investors, the central question becomes revenue, not capacity. The paper argues the real test is whether customers pay roughly $3.725 trillion a year for what that spending produces.
Revenue growth at leading labs like OpenAI and Anthropic is the most direct gauge. If combined revenue grows anywhere near 80% annually, the buildout looks defensible.
The financing angle deserves equal attention. The research notes that opacity in current arrangements creates potential vulnerabilities tied to specific tech giants.
Power is the third pressure point. A projected 183 GW of new capacity means utilities, grid operators, and regulators all become stakeholders.
Spending equal to 3.63% of GDP annually is large enough to influence overall economic growth figures.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.