Models
AI data centers: $3.7 trillion a year in revenue by 2032 to pay off, according to a paper published by Brookings
A Columbia economist puts US AI investment at $10.3 trillion from 2025 to 2032, and the revenue needed to pay it off at $3.725 trillion a year.
How much does AI need to earn to pay off the data centers under construction in the United States? A paper presented on September 25, 2026 at the fall Brookings Papers on Economic Activity (BPEA) conference gives an order of magnitude: about $3.725 trillion in annual revenue by 2032, or 9.2% of projected US GDP for that year (paper, conference draft).
To be precise. This is not an institutional Brookings study: it is a research paper by Stijn Van Nieuwerburgh, a real estate professor at Columbia Business School, published by Brookings as part of the BPEA (conference draft dated September 4, 2026). Each publication there represents the sole views of its authors (Brookings).
The key figures
- Investment: $10.3 trillion from 2025 to 2032 in buildings, power, networking, chips and other equipment, an average of 3.63% of GDP a year. That is more than the canal, railroad (2.24%), electrification, highway and telecom (1.10%) booms.
- Capacity: 182.8 GW of additional capacity coming online between 2025 and 2032, out of an announced pipeline of about 509 GW; the central scenario assumes 117.2 GW arrives after 2032 and 226.9 GW is never built.
- Unit cost: about $8.2 billion for a 200 MW campus, two-thirds for IT (GPUs, networking, storage) and one-third for the building and power.
- Required revenue: $3.725 trillion a year once all this capacity is running, against a projected 2032 GDP of $40.481 trillion.
The method: a break-even threshold, not a forecast
The $3.725 trillion figure answers a specific question: what revenue must the capacity completed between 2025 and 2032 generate to give investors a 10% unlevered return, with a 50% operating cash-flow margin, a six-year economic life for IT and twenty years for everything else? The author says it himself in a piece published by Columbia: it is a required-revenue calculation, not a prediction of what customers will spend (Columbia Business School).
The result is highly sensitive to the assumptions (paper, table 7):
- from $2.29 trillion (5% return, 67% margin) to $6.72 trillion (15% return, 33% margin);
- if GPUs stay competitive for only three years instead of six, the requirement climbs to about $6.0 trillion, or 14.8% of 2032 GDP.
For scale: with combined OpenAI and Anthropic revenue estimated at about $100 billion a year, reaching that threshold would take growth of roughly 80% a year until 2032, the author calculates.
Two versions in circulation. In his October 6 piece for Columbia, the author cites a revised version of the paper: nearly $9 trillion of investment, about 188 GW, and $3.55 trillion of required revenue (8.8% of GDP). The figures quoted by the press, and here, are those of the conference draft published by Brookings.
Other comparable estimates
- Bain & Company (September 2025) estimated that about $2 trillion in annual revenue would be needed worldwide by 2030 to fund the required computing power, and that $800 billion would still be missing even counting AI-related savings (Bain). The scope (global, 2030 horizon) and method differ: the two figures do not add up.
- The paper notes that capital spending by Oracle, Microsoft, Amazon, Meta and Alphabet rose from about $97 billion in 2020 to more than $400 billion in 2025, and is projected to exceed $800 billion in 2026, above their combined operating cash flow for the first time. Hence the growing use of leases, joint ventures, private credit and off-balance-sheet vehicles, which the author considers harder to monitor (Bisnow, Brookings podcast).
What it implies for the price of coding tools and APIs
The paper translates its threshold into GPU rental prices: about $5.5 per installed GPU-hour at full utilization, $6.9 at 80% utilization, within the range of current frontier compute prices. For the author, the real question is therefore whether providers will keep that pricing power once some 180 GW has been added. Two readings, which are not mutually exclusive:
- As long as capacity is short, prices hold. Price lists already show it: at Google, Gemini 4 Argon goes from $2/$10 to $4/$20 per million tokens on December 1, and Gemini 3.8 Flash from $0.75/$3.75 to $1.50/$7.50 on January 1, 2027, when introductory prices end (Google). A coding subscription or API that is "cheap" today can go back up.
- If supply outruns demand, they fall. "Sooner or later we're going to have oversupply, just like we do in every real estate cycle, and then the prices will collapse," the author told Bisnow; he expects that risk to appear only after five to eight years of strong growth.
In practice, for a team budgeting its coding agents: compare the price per task rather than the list price, check the end date of introductory prices, and keep a cheaper or self-hosted fallback. The cost calculator and the model comparison let you redo the math with your own volumes; see also our guide to choosing an agent by budget.
Sources
- Stijn Van Nieuwerburgh: Financing the AI Buildout (BPEA, conference draft) (Sep 4, 2026), verified 2026-10-21
- Brookings: Financing the AI buildout (summary) (Sep 23, 2026), verified 2026-10-21
- Brookings: The $10 trillion question (podcast and transcript) (Oct 8, 2026), verified 2026-10-21
- Columbia Business School: The AI Buildout: What Real Estate Investors Need to Underwrite (Oct 6, 2026), verified 2026-10-21
- Bisnow: Brookings Study Shows AI's $10T Build-Out Cost Will Likely End In Market Correction (Sep 23, 2026), verified 2026-10-21
- Bain & Company: $2 trillion in new revenue needed to fund AI's scaling trend (Sep 23, 2025), verified 2026-10-21
- Google: Gemini API pricing, verified 2026-10-21
Key takeaway: $3.725 trillion a year is a break-even threshold computed under assumptions (10% return, 50% margin, GPUs written off over six years), not a revenue forecast.
Spotted an error in this article? Report it · Our corrections policy