• A new analysis estimates the U.S. AI infrastructure buildout could require $10.3 trillion in investment from 2025–2032, demanding roughly $3.7 trillion in annual AI revenue by 2032—about 9.2% of projected GDP.
  • The revenue hurdle implies ~80% annual growth from a combined $100 billion base for OpenAI and Anthropic, raising concerns that falling compute prices and rapid obsolescence could leave investors with weak returns.
  • Financing is shifting from cash-rich balance sheets to private credit, securitizations, and special-purpose vehicles, potentially obscuring correlated risks, according to Columbia professor Stijn Van Nieuwerburgh.

The $10 Trillion Question

America’s race to build artificial intelligence infrastructure may require a scale of spending that dwarfs historic U.S. investment waves, according to a new scenario analysis from Columbia Business School professor Stijn Van Nieuwerburgh. The study, presented in late September, estimates that the planned buildout could demand about $10.3 trillion in capital from 2025 through 2032—roughly 3.6% of GDP annually. To justify that outlay, AI services would need to generate about $3.7 trillion in annual revenue by 2032, or 9.2% of projected GDP.

The headline figure has grabbed attention, but Van Nieuwerburgh is not predicting the U.S. will literally spend 9% of GDP on AI. Rather, the analysis frames the revenue requirement as a high hurdle that must be cleared for current investment to earn a 10% unlevered return, assuming a 50% operating cash-flow margin. The $10.3 trillion covers chips, data-center buildings, networking, and power infrastructure, and assumes roughly 183 gigawatts of new AI-campus capacity by 2032, up from about 57 GW today.

Revenue Growth Must Be Extraordinary

The commercial challenge is stark. The study’s $3.7 trillion annual revenue requirement implies an approximate 80% annual growth rate from the roughly $100 billion combined annual revenue base attributed to OpenAI and Anthropic in the analysis. That pace would need to be sustained through 2032—a tall order even for the fast-moving AI sector.

Large technology platforms are already spending heavily. J.P. Morgan (JPM) estimates capital expenditure by five leading U.S. hyperscalers could reach $697 billion in 2026, while S&P Global (SPGI) projects spending by Alphabet (GOOG), Amazon (AMZN), Microsoft (MSFT), Meta (META), Oracle (ORCL), and SpaceX (SPCX) could exceed $1.3 trillion in 2027. These separate estimates illustrate the speed and scale of the current buildout, but they also underscore the gap between today’s spending and the revenue needed to support it.

The Financing Shift

As the required investment grows beyond what even the largest firms can fund internally, capital is increasingly coming from real-estate developers, infrastructure funds, private-credit providers, joint ventures, leases, securitizations, and special-purpose vehicles. That broadens access to funding but can make leverage and risk concentration harder to assess.

Van Nieuwerburgh compared the opacity concern—not necessarily the scale or inevitability of an outcome—to dynamics seen in pre-2008 structured mortgage finance. The highest-risk combination, he warned, would be a demand slowdown occurring after debt and off-balance-sheet financing have become widespread.

Prices, Power, and Productivity

AI compute is a capital-intensive service: providers spend heavily upfront on land, electricity, buildings, chips, cooling, and networking, then try to recover costs by selling training and inference capacity. If more firms add capacity faster than revenue-generating workloads grow, compute prices can fall—a boon for users but a potential margin killer for investors.

“Economic surplus and investor returns can diverge,” the analysis notes, drawing a parallel to the telecom and fiber buildout of the late 1990s. Communications capacity ultimately became economically valuable, but excessive investment and competition damaged many early investors.

Electricity is the immediate physical constraint. The International Energy Agency projects global data-center electricity use will roughly double to about 945 TWh by 2030, with the U.S. and China accounting for nearly 80% of the increase. In the U.S., data-center consumption could rise by about 240 TWh from 2024 to 2030, a 130% jump, and account for nearly half of electricity-demand growth.

That dynamic is already driving strong demand for generation, transmission, transformers, turbines, gas capacity, nuclear development, renewables, storage, construction labor, and semiconductor equipment—while raising risks of local power-cost disputes, land and water competition, and inflationary pressure in infrastructure-heavy regions.

Policy Tailwinds and Geopolitical Stakes

Federal policy is explicitly oriented toward accelerating AI infrastructure. Executive Order 14318, issued in July 2025, directs agencies to facilitate permitting for qualifying AI data-center and related energy projects exceeding 100 MW of incremental load or $500 million in capital expenditure. It contemplates support such as loans, loan guarantees, grants, tax incentives, and offtake arrangements.

Supporters argue faster permitting is needed to preserve U.S. technological leadership. Critics worry about reduced environmental review, water use, local grid affordability, and subsidies that transfer risks from technology firms to taxpayers or ratepayers.

The buildout also sits inside U.S.–China strategic competition. Export controls on advanced chips and certain servers remain in place, making AI infrastructure part of a broader contest over compute scale, energy capacity, supply chains, and technical talent.

What to Watch

The analysis does not say a bearish outcome is inevitable. It says the revenue hurdle is extraordinarily high, while demand, compute prices, technology changes, power availability, and financing structures remain uncertain. The most useful indicators to watch are AI-service revenue growth rather than announcements alone, data-center utilization and lease terms, hyperscaler capital-expenditure guidance, electricity interconnection timelines, AI-compute price trends, private-credit and SPV exposure, and whether enterprises report measurable returns from AI deployment.

As the public debate shifts from “Will AI be transformative?” to “Who funds the buildout, who bears the downside risk, and who receives the gains?”, Van Nieuwerburgh’s scenario offers a sobering reminder: infrastructure can be socially valuable even when the timing, financing, or distribution of returns is poor for early capital providers.