Insider Brief
- An economist estimates that the U.S. AI data center buildout would need about $3.55 trillion in mature annual revenue to support investment under the study’s central assumptions.
- Shortening computing equipment’s assumed useful life from six years to three raises the annual revenue requirement to about $5.7 trillion.
- Leases, joint ventures and outside borrowing distribute financial exposure across technology companies and investors, while the study’s estimates remain conditional scenarios rather than forecasts.
The U.S. data center expansion could require about $3.55 trillion in annual revenue to support its investment costs, according to a new working paper that examines whether demand for artificial intelligence (AI) can sustain the infrastructure being built around it.
In plain terms, this means the companies selling computing capacity from those data centers would collectively need to bring in nearly $10 billion a day to make the investment pay off.
The calculation puts a financial test behind the construction boom. Developers must secure electricity, install costly computing equipment and attract customers willing to pay enough for capacity to cover operating expenses, equipment replacement and returns to investors.
In the study, Columbia Business School finance professor Stijn Van Nieuwerburgh estimates that a scenario involving about 188 gigawatts of additional U.S. data center capacity becoming operational during 2025–2032 would generate nearly $9 trillion in investment over that period. A gigawatt is a measure of power demand, used here to describe the scale of data center capacity.
The revenue requirement depends on several assumptions, including a 10% annual investment return before borrowing, a 50% operating cash-flow margin and a six-year economic life for computing equipment. It describes what the completed facilities would need to earn once operating at mature levels, rather than predicting what AI businesses will actually collect.
The paper also looks at how leases, joint ventures and outside borrowing distribute financial exposure beyond the technology companies using the facilities. These arrangements can expand the money available for construction while creating obligations whose economic consequences can be difficult to assess from corporate balance sheets alone.
What the Buildout Must Earn
Van Nieuwerburgh estimates an initial cost of about $41 billion for a representative 1-gigawatt AI campus, including computing systems, the facility and associated power infrastructure. A smaller 200-megawatt campus would cost about $8.2 billion under the same assumptions.
Roughly two-thirds of that spending goes toward computing equipment and related systems. The remaining share funds buildings, cooling, backup power and electricity infrastructure.
To estimate the national buildout, the researcher used Cleanview project-level data, a database that showed about 63 gigawatts of operating U.S. data center capacity and a development pipeline of about 494 gigawatts, including projects under construction and those still planned.
Of course, not all of this will be built out and the study does not assume that the entire pipeline will be on line. It applies completion and cancellation assumptions that vary with project timing and documentation. Its central scenario produces about 188 gigawatts of additional capacity during 2025–2032, including projects already operating in 2025 or 2026, with further capacity completed later and a substantial portion never realized.
The nearly $9 trillion investment estimate includes spending during 2025–2032 on some projects that would open after 2032. It assumes project costs rise 4% annually and distributes expenditures across construction and completion years.
For the revenue calculation, the paper considers only capacity completed during 2025–2032. It calculates the annual cash flow needed to recover capital, compensate investors and account for the different useful lives of computing equipment and other assets.
Based on the key assumptions of the paper, those facilities would need about $1.77 trillion in mature annual operating cash flow. At a 50% operating cash-flow margin, meaning half of revenue remains as operating cash flow, that translates into roughly $3.55 trillion in annual revenue.
That figure should not be looked at as some sort of sales target for OpenAI and Anthropic. It should also not serve as an estimate of the economic value AI might create for users. It is a modeled revenue requirement for the infrastructure included in the calculation.
The outcome changes substantially with financing expectations and profitability. Across the paper’s combinations of required returns and operating cash-flow margins, the annual revenue requirement ranges from about $2.13 trillion to $6.55 trillion.
The Hardware Replacement Problem
The useful life of computing equipment is a critical — and often overlooked — assumptions in the study.
Buildings and power infrastructure can remain useful for decades. Graphics processing units, or GPUs, which perform much of the computing required to train and run AI models, face a faster cycle of technological change. Older equipment can remain operational while becoming less attractive to customers seeking better performance or lower costs.
The paper assumes that computing equipment and related IT systems account for 68% of investment and have a six-year economic life, while assigning a 20-year life to the remaining assets.
If the computing equipment’s economic life falls to three years, with other assumptions unchanged, the required mature annual revenue rises to about $5.7 trillion. Faster replacement increases the cash that facilities must generate to sustain the investment.
This does not necessarily mean that GPUs will become commercially unusable after three years. The figure illustrates how sensitive project economics are to equipment longevity, an issue that depends on future hardware, customer needs and the competitiveness of older systems.
Use presents another test as a facility faces investment costs whether its computing capacity is fully occupied or partly idle.
Using the paper’s assumed hardware configuration, the central revenue requirement translates into about $5.10 per installed GPU-hour at full utilization. At 80% utilization, the required price rises to about $6.40 per billed GPU-hour. At 70%, it reaches about $7.30.
These are modeled revenue requirements, not just operating-cost estimates or predictions of future rental prices. The study compares them with frontier-GPU rental prices available at the time of its analysis and finds that the central requirement is within their general range.
One of the questions might be whether enough customers will purchase computing services at those prices after a large expansion in supply. Competition could reduce prices, while more efficient models or greater use of computing near end users could change demand for centralized facilities.
Strong growth in AI applications could support high utilization and the paper identifies both possibilities without resolving which will dominate.
Following the Financial Obligations
The financing structure is considered because infrastructure users increasingly share ownership and funding responsibilities with developers, investment funds and lenders.
A technology company can secure capacity through a lease while a separate business owns the facility and borrows to build it. The arrangement reduces the technology company’s immediate funding burden, but payments and guarantees can leave it exposed to the project’s performance.
Van Nieuwerburgh uses Meta’s Hyperion development as a case study. The paper covers the roughly $30 billion transaction involving about 2 gigawatts of capacity, with computing equipment financed separately by Meta.
According to the study, Meta sold an 80% equity stake to Blue Owl, and the resulting joint venture raised about $27 billion in external debt. That amounted to borrowing equivalent to roughly 90% of the project’s asset value.
The financing reconciles long-term borrowing with a series of shorter leases. According to the paper, the five four-year lease periods begin in 2029 and extending through the debt’s 2049 maturity.
Meta can terminate leases at renewal dates for some or all of the campus. But that flexibility carries a financial obligation. If affected assets are sold for less than a contractually specified minimum value, Meta must cover the shortfall through a residual-value guarantee.
The guarantee illustrates why outside ownership does not necessarily remove exposure from the technology company. A tenant may retain obligations even when the project’s debt sits elsewhere.
According to the paper, that reported corporate leverage can therefore give an incomplete picture of economic commitments. It also notes that the Hyperion borrowing carried a higher interest rate than Meta likely would have paid on comparable direct corporate borrowing. In the researcher’s assessment, preserving corporate financing flexibility can justify accepting that additional cost.
For lenders, the structure creates dependence on contractual payments, guarantees and asset values. High project-level borrowing leaves less room to absorb weaker revenue or lower sale proceeds.
Future Demand?
The paper combines a project-based investment scenario, financial calculations and a transaction case study. It does not directly measure future AI demand or establish the likelihood of widespread defaults.
Its capacity estimate is particularly sensitive to large projects, multiphase developments and projects without reported completion dates. Applying a benchmark AI-campus cost across the scenario also makes the investment total dependent on assumptions about equipment, facility design and future construction costs.
The estimate excludes later additions to the project pipeline and subsequent replacement of computing equipment. Those exclusions could raise eventual spending, while cancellations, downsizing or lower costs could reduce it. The revenue calculation separately accounts for assumed equipment lives when estimating the annual cash needed to support completed capacity.
It should also be noted that historical investment — like the ones covered in the paper — measures differ in coverage, and assets such as railroads and fiber networks generally have longer useful lives than advanced computing equipment. A larger amount of gross spending does not necessarily create a proportionately larger stock of durable capital.
As for financial exposure, the researcher adds that many bank loans and private-credit arrangements provide less public information than rated debt transactions, limiting the ability to trace obligations across the sector.
Further research would need to establish where borrowing and guarantees are concentrated, how lenders value aging equipment and whether trouble at individual projects could spread across firms and financial institutions.
For a deeper, more technical dive, please review the National Bureau of Economic Research working paper. It’s important to note that arXiv is a pre-print server, which allows researchers to receive quick feedback on their work. However, it is not — nor is this article, itself — official peer-review publications. Peer-review is an important step in the scientific process to verify results.
