Economist Estimates AI Buildout Would Need $3.55 Trillion in Annual Revenue

In a working paper, Columbia Business School finance professor Stijn Van Nieuwerburgh puts U.S. AI data center construction on a cash flow statement: how much money must these facilities earn once mature to cover capital recovery, equipment replacement, and investor returns? The paper's core scenario is about 188 GW of new U.S. data center capacity coming online between 2025 and 2032, with nearly $9 trillion of investment over the same period.

Under this scenario, the paper estimates the relevant facilities need about $1.77 trillion in mature-state annual operating cash flow; at a 50% operating cash flow margin, that corresponds to about $3.55 trillion in annual revenue. By the paper's framing, this is equivalent to the relevant companies needing to take in nearly $10 billion per day.

The nature of this number must be clear first. The paper's author emphasizes that these figures are modeled revenue requirements under conditional scenarios, not forecasts of AI companies' actual revenue or AI's economic value. They describe the revenue level completed facilities would need to reach in a mature operating state, not a sales target set for any company.

The revenue estimate rests on three key assumptions: a 10% annual return on investment (pre-borrowing), a 50% operating cash flow margin, and a six-year economic life for compute equipment. Change any one of the three, and the revenue threshold moves noticeably.

The paper runs estimates across different combinations of required return and operating cash flow margin, producing an annual revenue requirement range of roughly $2.13 trillion to $6.55 trillion. In other words, $3.55 trillion is one point under the core scenario, not the only answer.

The cost side also has an explicit framing. The paper estimates a representative 1 GW AI campus costs about $41 billion initially, and a 200 MW campus about $8.2 billion under the same assumptions; about two-thirds goes to compute equipment and related systems, with the rest for buildings, cooling, backup power, and electrical infrastructure. The paper also assumes project costs rise 4% per year and allocates spending by construction and completion year.

The paper assumes compute equipment and related IT systems account for 68% of investment and have a six-year economic life, while other assets are calculated on a 20-year life. Buildings and power infrastructure can last decades, but GPUs face faster technological iteration: older equipment can still run, yet may no longer be favored by customers because of performance or cost.

If the assumed economic life of compute equipment is shortened from six years to three, with other assumptions unchanged, the required mature-state annual revenue rises to about $5.7 trillion. A faster replacement pace means facilities must generate more cash to sustain the same investment return.

The paper also notes this does not necessarily mean GPUs cannot be commercially used after three years. The figure illustrates the project economics' high sensitivity to the equipment-life assumption, while actual life depends on future hardware, customer demand, and the competitiveness of older systems.

The paper converts the revenue requirement into compute units: under the paper's assumed hardware configuration, the core revenue requirement is about $5.10 per installed GPU hour at full utilization; about $6.40 per billed GPU hour at 80% utilization; and about $7.30 at 70%.

The paper compares these estimates with then-available frontier GPU rental prices and finds the core requirement falls within their approximate range. But this comparison is based on prices at the time of analysis and is not a forecast of future rents.

Utilization itself is also a variable. Facilities bear investment costs regardless of whether compute is fully utilized; whether there will be enough customers to buy compute services at the modeled prices after large-scale expansion remains uncertain. The paper also notes competition could push prices down, and more efficient models or nearby compute could change demand for centralized data centers, but it does not determine which scenario will dominate.

The paper cites Cleanview project-level data showing the U.S. already has about 63 GW of operating data center capacity, with a development pipeline of about 494 GW (including under construction and planned projects). But the paper does not assume all 494 GW of pipeline will be built; instead it sets completion and cancellation assumptions based on project timing and documentation completeness.

The core scenario's 188 GW of new capacity includes projects already operating in 2025 or 2026, plus capacity completed later, while a substantial portion of the pipeline will never materialize. The nearly $9 trillion investment estimate includes spending from 2025 to 2032, but some of those projects will open after 2032.

The boundaries of the estimate matter equally: the revenue requirement covers only capacity completed and brought online between 2025 and 2032, excluding the revenue requirement of projects coming online after that. This limitation means the paper's figure corresponds to an asset set with a defined time window, not an open-ended industry total.

The paper discusses how leasing, joint ventures, and external borrowing spread financial risk exposure beyond tech companies to developers, investment funds, and lenders. A tech company can obtain compute capacity through a lease while another company owns the facility and borrows to build it.

This arrangement reduces the tech company's immediate funding pressure, but the resulting payment and guarantee obligations make the related economic consequences hard to judge from corporate balance sheets alone. The paper therefore brings financing structure into the analysis, rather than looking only at the company using the facility.

This also explains why the revenue threshold deserves separate discussion: when obligations are spread among multiple parties, judging whether the entire buildout cycle is sustainable requires returning to the facility level to see how much it must earn, rather than looking only at one company's statements.