If the US stopped building data centers, where would it show up?
In August, American companies were building data centers at an annual rate of $85 billion, according to the Census Bureau's construction survey. A year earlier the figure was $49 billion, and in August 2023 it was $20 billion. At $85 billion, the building work alone is roughly a quarter of one percent of US GDP.
Over the same twelve months, all other private nonresidential construction taken together fell by about 6%. Outside data centers, American commercial building is shrinking.
A slowdown is no longer a purely hypothetical case. In July, New York's governor signed an executive order pausing permits for large new data centers across the state, the first statewide pause of its kind. In many regions, connecting a large new load to the grid now takes years. So a stop does not need the companies paying for these buildings to change their minds. Local permits or a lack of grid capacity can produce the same result.
The usual answer this year has been that a stop would hit the whole economy hard, because AI investment is holding up American growth. The national accounts support a narrower version of that.
Why the GDP number is smaller than it looks
Gross domestic product measures what is produced inside the country. When an American company buys a server assembled abroad, the purchase counts as investment, but the import is subtracted again, so the net effect on GDP is close to zero. Only the domestic share of the spending adds to growth.
Imports matter here because most AI-related spending goes on equipment rather than buildings. Epoch AI, counting only AI-related data centers and only the first quarter of this year, puts their construction at about 0.15% of US GDP, below the Census figure above, which covers all data centers. It puts AI computing hardware at about 0.56% of GDP and networking at about 0.12%. Together that is about 0.8% of GDP, and roughly four fifths of it is equipment rather than buildings. Much of that equipment is manufactured or assembled outside the United States.
Economists who net out imports get much smaller numbers than those who do not. An August analysis from the Flossbach von Storch Research Institute found that hardware imports have risen from 0.5% to 1.6% of GDP since 2024, and that once they are subtracted, the contribution of hardware and software investment to growth falls from 0.70 percentage points to 0.14. Against second-quarter growth of 2.2%, the gross figure would be roughly a third of all growth and the net figure about 6%. Wells Fargo puts AI's share of growth over the past year at about a quarter after imports. The estimates differ mainly in what they count as AI investment and in how much of the hardware they treat as imported, which is why serious estimates still range so widely.
For a stop, the consequence is that much of the spending would disappear from both sides of the ledger at once. Fewer servers would be bought, and fewer would be imported. The direct hit to GDP comes mainly from the domestic part: the buildings, the electrical and cooling systems, the grid connections, and the people who install them. On the estimates above, losing it would cost somewhere between a few tenths of a percentage point and about half a point of annual growth, depending on what is counted.
Construction is where the jobs are
A data center needs a large workforce to build and a small one to run. Virginia's legislative audit commission, which studied the industry in the state that hosts the most of it, put peak construction at around 1,500 workers per building over a twelve- to eighteen-month build. Permanent on-site employment for a typical 250,000 square foot facility was around 50, about half of them contractors. Electricians, pipefitters and mechanical crews are hired for the build and released once the facility is energized.
So a stop would not look like a national labor market recession, because operating jobs barely depend on new construction. It would look like a sharp downturn concentrated in specific trades and specific counties, the places where these campuses cluster. And it would arrive while the rest of commercial construction is already shrinking and has little work to absorb those crews.
Grid upgrades cannot be undone
Utilities build ahead of demand. A new substation or transmission line is planned around the load that customers say is coming, and its cost is recovered from customers over decades. If a large load is promised and never arrives, someone still pays for the wires.
Regulators have started writing contracts for exactly that case. In Ohio, the utility commission approved a tariff in 2025 under which large new data centers pay for at least 85% of the capacity they reserve, whether they use it or not, for twelve years including a four-year ramp-up. Leaving early costs three years of those minimum charges. The stated purpose was to stop infrastructure built for promised demand from being stranded on other customers, and the order is now under appeal at the Ohio Supreme Court.
Where tariffs like that exist, a stop is paid for largely by the companies that asked for the capacity. Where they do not, the cost of unused grid investment is spread across everyone else's electricity bills. Which of those applies is being decided now, state by state, in rate cases that get little attention.
The bigger channel runs through share prices
A meaningful part of the stock market's value rests on the expectation that this spending continues. When people's holdings rise in value they tend to spend more, and when holdings fall they pull back, which economists call the wealth effect. The Flossbach von Storch analysis makes this point directly. American households are saving 2.7% of disposable income, against an average of 5.2% between 2000 and 2019, so consumer spending is running with little cushion. A sustained fall in the shares tied to AI would reach the economy through that spending rather than through construction.
That is also where a stop would show up first. A cut in capital spending is announced on an earnings call before a single site goes quiet. Share prices move on the announcement, and construction data follow months later.
What the last buildout left behind
The closest precedent is the fiber-optic boom of the late 1990s. Carriers laid far more cable than the market needed, and in mid-2001 an industry estimate put the share of fiber actually in use at around 5%. Capital spending collapsed, many of the companies that built the networks failed, and their investors lost most of what they had put in.
The fiber itself stayed in the ground. Other companies bought it cheaply and used it for the traffic that arrived over the following years, so most of the eventual return went to the buyers rather than to the companies that laid it.
A stop today would split the buildout along a similar line, with one important difference. Buildings, power connections and cooling plants behave like the fiber did. They last for decades and can sit partly idle until demand catches up. Chips lose most of their value within a few hardware generations, so equipment bought for demand that arrives late is worth far less by the time it arrives.
Construction, which counts most in American GDP, would mostly hold its value. Hardware would lose the most, but it is largely produced abroad, so those losses would land on the balance sheets of the companies and lenders that financed it, not on domestic output.
Where a stop would land
How much US growth depends on this spending has no single answer. It turns on what counts as AI investment and how much of the hardware is treated as imported, and serious estimates run from about 6% of growth to roughly a third. We would put less weight on any one of those figures than on what holds across all of them, which is that a stop would land unevenly. Specialized construction crews and the counties that host the campuses would feel it first, along with electricity customers in states without protective tariffs. The wider effect would come through household wealth and the stock market.
A full stop is unlikely as a single event. A slowdown at the edges, through permits, power and financing, is already visible, and the same analysis applies to it in proportion.
Our reading depends most on the import share. If more of the hardware starts being made in the United States, as several manufacturers have announced plans to do, the domestic share of each dollar rises and the economy's direct exposure to this spending grows with it. That shift would show up in the trade data before anywhere else.
On October 29 the Bureau of Economic Analysis publishes its first estimate of third-quarter growth, and the largest operators report results in the same week, which is where any change to next year's spending plans will appear first. Census construction figures for September follow in early November and will show whether the gap between data centers and everything else is still widening.