A couple of weeks ago, I began seeing a remarkable number in online discussions. Building data centers, the chatter said, would require more than 3 percent of U.S. GDP.
That estimate made me raise my eyebrows. The US is experiencing a boom in data center building, no doubt, but 3 percent would require a massive redirection of material and labor. Initially, I chalked it up to the excessive optimists in the technology community. Somebody probably made a heroic projection, I guessed, and others repeated it on social media. Repetition gives it authority. It soon became conventional wisdom.
Such numbers deserve sober scrutiny. Grounded analysis should guide policy. We could end up making major policy decisions based on conventional wisdom that turn out to be wildly inflated or, at least, a high-end scenario rather than a likely outcome.
Here is my bottom line: Data center construction will require massive investment, but nearly 3 percent of GDP is unlikely.
A reasonable scenario
Start with a new NBER working paper by Stijn Van Nieuwerburgh, which estimates that a 188-gigawatt U.S. data center buildout completed by 2032 would require nearly $9 trillion in investment, averaging 3.2 percent of GDP between 2025 and 2032. An earlier version of the paper, presented at Brookings in September, put the cumulative figure above $10 trillion. The headline number has already shifted between drafts, which is another reason not to treat any single version as a forecast.
Most of the paper examines how an investment boom of that size might be financed. That is a worthwhile question. Even if firms build only half of the 3.2 percent (which I consider plausible), the answer helps public and private planning.
The author wants to understand the financial implications of a huge building boom but needs to base that on a buildout scenario. Deep in the appendix, the paper describes a realized-capacity scenario, using information about thousands of planned and actual projects, sourced from a reputable consulting firm that tracks the industry.
Scenario is the right label. Scenarios and forecasts are not the same. A scenario asks what would happen under a certain set of assumptions. A forecast asks what is likely to happen after markets, prices, bottlenecks, politics, and strategic responses have done their work, and assigns probabilities to several potential outcomes.
Forecasts guide action. Public officials use them when considering transmission lines, new generation, utility rate cases, tax abatements, water allocations, zoning, and permits. Private actors use them to underwrite loans, order turbines and transformers, expand semiconductor capacity, acquire land, and value unfinished projects.
3.2 percent is not likely
The estimate in the appendix for the cost of an individual data center is in the right ballpark, by today’s standards. A 200-megawatt AI campus costs about $2.2 billion for the facility, $400 million for power infrastructure, and $5.6 billion for computing equipment and associated systems. The total is about $8.2 billion. The paper then estimates $41 billion for a 1 gigawatt campus.
Doubts arise when moving from one to many. The paper uses industry lists of planned and actual projects and assigns each pipeline project its own probability of completion. The list contains more than a thousand likely projects, with most projects slated for completion before 2030.
This assumes each project’s odds don’t depend on other projects. Economists call this the fallacy of composition. The whole is not simply the sum of many isolated projects.
Why? Because buyers compete and prices change. Supplies run short. Competitors react. Builders apply to the same regulators for permits, and they do not receive approval for all of them. Supply chains slow as builders compete for the same turbines, electricians, chips, and permits. Some buyers retreat after getting bad news in a competitive race with other builders.
To be fair to the paper, it offers sensible reasons for discounting the project pipeline, including grid access, equipment availability, skilled labor, financing, local moratoria, and regulatory scrutiny. It also assigns some zeroes for projects unlikely to ever happen. Those forthright, sensible qualifications did not make it into the 3.2 percent figure on social media.
That is abstract, so consider the specific constraints that emerge when everybody acts simultaneously. Each one points toward a more cautious forecast. Here are reasons why the United States is unlikely to build 188 gigawatts of data center capacity on the paper’s schedule.
Not enough power is available
The United States currently lacks enough readily available power for generating electricity to support the proposed buildout at full scale. The paper notes that, at high utilization, the new capacity would consume roughly as much electricity as the entire U.S. residential sector consumes today.
Where would all that electricity come from on the proposed schedule? It would have to come from something new. The US has no plans for major hydroelectric dams. New natural gas is a realistic possibility, but every builder has already considered that, so equipment is in short supply. Buyers are being told that backorders will delay deliveries for years. Some firms are exploring a small-scale rebirth of nuclear power, but that will take time, too.
It does not make headlines, but suppliers see it on the ground. Turbines, transformers, substations, and transmission equipment cannot be summoned on demand.
Insufficient skilled labor
Consider constructing a 150 Megawatt data center, which is much smaller than 1 gigawatt, but was the frontier last year. As a rule of thumb, building such a data center on a one-year timeline required about 1,000 licensed, trained electricians, depending on the building design, and almost as many pipe fitters. If project managers sign up only half that many, construction takes much longer.
There are only so many electricians, and it takes several years to become one. Once again, it is not the sort of topic that makes headlines, but there are not enough electricians in the US to accelerate construction.
Not enough semiconductors
Memory, GPUs, networking gear, and other electronic components are in short supply. Supplying the projected buildout would require large additions to manufacturing capacity, or considerable substitution away from Nvidia systems, or both.
Some of that could happen. New capacity is being built. Google has shown a willingness to substitute at scale, as it uses its own Tensor Processing Units (TPUs). Other firms are starting to design their own too.
Adding capacity takes years and many billions of dollars. It is sober to say, but additional supply will not arrive anytime soon.
Government permitting is slow
Data centers are unpopular. This matters when builders apply for permits. Permitting poses another constraint.
Local opposition has already delayed or altered major projects. Data centers require land, transmission, water or alternative cooling systems, and agreements about who will pay for new electrical capacity. No sensible baseline can presume that resistance, regulatory review, and cost disputes will disappear everywhere.
Higher prices do not close the gap
These constraints limit how many gigawatts get built. They do not automatically limit how many dollars get spent. When turbines, electricians, and chips are scarce, their prices rise, and a smaller buildout at higher prices could in principle cost as much as the scenario assumes.
Two things make that unlikely. The scenario already prices each campus at today’s elevated costs, so matching its dollar total on, say, half the capacity would require roughly doubling the cost per gigawatt. And higher input prices feed back into the decision to build. When a gigawatt costs more, fewer projects clear a buyer’s return hurdle, and buyers delay, shrink, or cancel. Scarcity rations through both price and quantity, and when the buyers are a handful of firms watching their own returns, quantity gives way first.
The data so far do not support 3.2 percent
The scenario attracts attention because many people believe expenditures will accelerate in the future. Still, comparing the estimate to the previous year’s actuals can be instructive.
The paper included an estimate for 2025, which has concluded and for which we have data. That lets us gauge how accurate the estimates are likely to be.
The paper assigns $790 billion, or 2.6 percent of GDP, of data center investment to 2025. Estimates from US government sources suggest that, under generous assumptions, about $300 billion went to data centers in 2025. That’s just under 1% of GDP.
That estimate starts with data from two data series. First, the US Census estimates that data center structures accounted for around $50 billion, excluding land. Second, the Bureau of Economic Analysis reports that private investment in computers and peripheral equipment was about $280 billion in 2025, up more than 50 percent from about $180 billion in 2024, which is astonishing.
The second series isn’t the right number because the data doesn’t tell us who bought the equipment, including laptop purchases and spending on older data centers. Let’s estimate how much of that $280 billion went to new data centers. First, we’ll generously attribute all the annual growth to data centers and most of the remainder, so we end up with about $250 billion. Adding $50 billion for structures gives us $300 billion, far less than the $790 billion the paper estimates for 2025. Adding land costs and other miscellaneous expenses will not make up the difference.
Could this number rise? Certainly. By the same estimates, data center construction in 2026 is growing quickly. Could the total reach 3.2 percent of GDP year after year? It would require breaking nearly every observable constraint or paying far more per gigawatt than anyone now budgets.
Any negative revenue shock could delay investment
The proposed buildout depends on a small group of firms continuing to anticipate enormous revenue from AI-based services.
Like most economists who have examined the topic, I think the latest frontier of artificial intelligence will yield productivity gains, and firms will pay for those. I am less optimistic than many about when productivity gains will arrive. It took the commercial internet more than a decade to create some of its most productive applications, and I expect the same here. More to the point, seller revenue differs greatly when those gains arrive tomorrow versus a decade from now.
If even one major buyer falls substantially short of its revenue projections in the next year or two, it will delay its data center projects, renegotiate commitments, or cut and run. I am not predicting that every major buyer will do this. I am saying it is entirely plausible that at least one will. Indeed, given the uncertainty surrounding AI applications and revenues, it would be astonishing if none did.
Who pays if the forecast is wrong
Scenarios can expose risks, but do not substitute for private forecasting. A lender should not underwrite a project as though every neighboring project will obtain power, equipment, tenants, and financing on schedule. They cannot replace judgments about probability.
The estimates matter for public policy because losses do not necessarily remain private. Utilities may build generation and transmission and turn to ratepayers to recover those costs. State and local governments may grant tax abatements, reserve water, widen roads, or issue debt. Regulators may approve investments expecting enormous loads to materialize. If the forecast proves wrong, the data center developer may walk away while households and ordinary businesses continue paying for the infrastructure.
Public planning must consider a high-demand scenario. It must also consider a low-demand case, staged investments, cancellation clauses, financial guarantees, and rules assigning the cost of unused capacity to the firms that requested it. That is good planning.
The sensible policy response to uncertainty is not to select the largest published number as a forecast. Instead, consider the distribution of possible outcomes, preserve options, and protect the public from forecast error. The paper examined a neglected scenario involving 3.2% of GDP, and I am glad it did so. That scenario is now part of general thinking.
Summary
Will data centers account for 3.2 percent of GDP? Maybe, but probably not. Plans collide. Constraints bind. Prices rise. Projects disappear when they do not meet revenue targets.
The AI buildout is real. It is large. It may transform the economy. Public agencies and private firms should plan for its growth. Still, public policy should not bet on only one scenario.
Shane Greenstein is the MBA Class of 1957 Professor of Business Administration at Harvard Business School and co-chair of the Harvard Business School Digital Initiative. He teaches in the Technology, Operations and Management Unit. Greenstein is also co-director of the program on the economics of digitization at The National Bureau of Economic Research. Encompassing a wide array of questions about computing, communication, and Internet markets, Greenstein’s research extends from economic measurement and analysis to broader issues. His most recent book focuses on the development of the commercial Internet in the United States. He also publishes commentary on his blog, Digitopoly, and his work has been covered by media outlets ranging from The New York Times and The Wall Street Journal to Fast Company and PC World. Greenstein previously taught at the Kellogg School of Management, Northwestern University, and at the University of Illinois, Urbana/Champaign. He received his B.A. from University of California at Berkeley and Ph.D. from Stanford University, both in economics.
