The hyperscalers are devouring the industrial base Washington is trying to rebuild. Data-center construction represents real economic activity at a time when not much else is being built — Census Bureau figures the Wall Street Journal reported showed data-center spending up 23 percent in May from a year earlier, while manufacturing-building construction fell 22 percent to a seasonally adjusted annual rate of $174 billion. The concrete is real. The steel is real. The trouble is what the concrete and steel are displacing, and what happens when you look past the aggregate to who is consuming the scarce inputs any manufacturer needs to put up a plant.
Mitchell Metal Products in northern Wisconsin wanted to add robotic welders and automated assembly equipment. A similar expansion in 2018 had cost $727,000. When bids came back last fall, the lowest was $2.1 million — nearly three times the prior cost, against a $1.2 million budget. The project was abandoned. Universal Metal Products in Ohio wanted to double its Texas stamping plant to offer more engineering and assembly work. At roughly $200 a square foot — about 60 percent above South Texas industrial construction a couple of years ago — the company scaled the project from 60,000 square feet to 40,000 and put it on hold for more than a year. “I never expected this to be such a pain in the neck,” said CEO Scott Seaholm.
These are not anecdotes. They are the documented consequence of a construction market where one category of buyer is absorbing inputs at a scale the rest of the industrial base cannot match. Data centers accounted for 8 percent of private nonresidential construction spending in May. Manufacturing, by contrast, accounts for nearly a quarter. One segment is being built at a rate that exceeds its share of the total by a factor of three; the other is contracting at a rate proportional to its share.
The mechanism is not complicated. Materials costs for nonresidential construction are up more than 55 percent since early 2020, per the government’s producer price index. Prices for transformers — the electrical-equipment components that step voltage up or down for industrial power distribution — are up roughly 70 percent in five years. Order backlogs for some electrical equipment run past two years. In a tight market, the buyer with the deepest pockets and the most urgent schedule gets the gear first. The data center, by its nature, requires electrical infrastructure at a scale and on a timeline no single factory can match. Whatever equipment the market can produce gets routed first to the buildout that has decided to consume the most of it. Thomas Murphy, a vice president at a Rochester-area electrical-infrastructure contractor, told the Journal: “Between data centers and solar projects, they’re taking a huge bite of any equipment.”
In such a market, the manufacturer’s options narrow. Toyota last week said it planned to spend $3.6 billion by 2030 to shift some Tacoma pickup production from Mexico to San Antonio, a move driven largely by tariff-driven cost-minimisation for its largest market. But this is not greenfield capacity. It is an incumbent foreign automaker reassigning existing production lines to defray tariff exposure — a rearrangement, not an expansion of the domestic manufacturing base. The factories Washington has already committed to are struggling. Softening chip demand, technical-worker shortages, and cost overruns have slowed construction at some semiconductor plants — the plants the federal reshoring push was supposed to underwrite. The data centers keep accelerating. The fabs keep stalling.
And then there are the tariffs. The Trump administration’s tariffs raise domestic prices for steel, electrical gear, and other construction materials — the very inputs manufacturers need to build the factories the tariffs are supposed to incentivize. As the builders’ group economist told the Journal, for a manufacturer with overseas plants or suppliers, “Is this the time to move my supply chain to the United States? In many cases, the answer is no.” The tariffs raise the floor on what it costs to build anything in this country. The data-center buildout raises the ceiling on what those inputs cost. The manufacturer is caught between, and the hyperscaler walks through with a checkbook.
What the public discourse about AI consistently refuses to say plainly: the current generation of large language models is not an abstract technology that lives in the cloud. It is a logistics system. It requires concrete, steel, copper wire, transformers, diesel generators, and water for cooling. The data center is not a brain. It is a warehouse that happens to run software.
The extraction is not limited to electrical equipment. Land, energy, water, silicon, copper, the construction labor to assemble them — the hyperscalers are not merely outbidding American factories for transformers. They are consuming the material substrate of the industrial economy wherever the same input markets are tight. The transformer is the most visible piece because the manufacturer had a name and a project that got abandoned. The same procurement squeeze is being run, less visibly, on the rest of what an industrial economy needs to put up a plant.
The straightest lever is the one the administration is not pulling. A rollback of the tariffs on construction inputs — steel, electrical gear, copper — would lower the floor on factory-construction costs immediately. The federal government has direct authority over the tariff schedule; no permitting pause, no congestion charge, no priority-allocation framework requires comparable institutional invention. The harder interventions — a congestion charge on hyperscaler infrastructure that prices the externality into the data center’s cost of capital; a priority-allocation framework reserving a share of electrical-equipment production for domestic industrial use; a pause on new data-center permitting until the equipment backlog clears — are real interventions with real tradeoffs, and they should be argued for on their merits. A congestion charge raises the cost of capital for AI deployment, which has its own economic consequences and would be opposed by every firm positioned to capture the buildout’s rents. A permitting pause would invite litigation and would be ignored by firms that have already broken ground. A priority-allocation framework would require an industrial policy the United States has not seriously deployed since the second world war.
The simplest intervention is also the one that does the most for the manufacturer trying to put up a plant in northern Wisconsin. The administration is not deploying it.
The transformer that could have powered a new production line in Wisconsin is sitting in a server hall built for a different purpose. The manufacturer is still waiting. The wait is two years, if the next one comes at all.