The AI Boom Pays Trades, Schools, and a Hidden Tab

The AI boom is already paying electricians, school boards, and chatbot users, even as most of the rest of measured U.S. growth barely budges. Paul Krugman argued on Sept. 30, 2026 that this is a technology of oligarchs. Noah Smith spent Oct. 5 answering that jobs, taxes, and household use already show otherwise.

Both men are arguing about who the technology is for. The larger effect sits one step over: a buildout big enough to carry national growth, fatten a few county ledgers, and bid labor and power away from everyone else.

Americans Already Use the Thing They Say They Hate

Krugman, a Nobel economist, put the public’s sour mood in payroll terms. Data centers, in his telling, are banks of blinking servers that hire almost no one, and even the construction phase is thinner than the capital spending implies. He leaned on work by economist Stijn Van Nieuwerburgh showing that only about a third of a data-center bill is the building and the power kit, with the rest in compute hardware, much of it imported.

This is a technology of, by and for oligarchs, with hardly any of the benefits trickling down to regular Americans.

Paul Krugman, economist, Sept. 30, 2026 essay

He added that firms have never spent so much, while helping to push interest rates higher, to create so few jobs. Smith, an economist and writer, called that reading incomplete. Households have piled into chatbots. Trades payrolls have thickened since late 2022. Local governments are booking property tax they did not have a decade ago. And a demand shock this large, he argued, is what kept a year of tariffs, deportations, and policy whiplash from showing up as a slump.

Americans still tell pollsters they dislike the technology. They fear job loss, and some fear worse. Anger at a handful of billionaires is not usually the first reason they give. The split that actually organizes the fight is more mundane: people who pour concrete and pull wire, people who pay the power bill, and people whose first office job is already thinner.

One Spending Category Carried 92% of Early 2025 Growth

Harvard economist Jason Furman posted the accounting that still sits under this argument. Investment in information-processing equipment and software was 4% of GDP in the first half of 2025. It accounted for 92% of measured real GDP growth in that stretch. Strip the category out and the leftover grew at a 0.1% annual rate.

Investment in information processing equipment & software is 4% of GDP.

But it was responsible for 92% of GDP growth in the first half of this year.

GDP excluding these categories grew at a 0.1% annual rate in H1. pic.twitter.com/7p1eAI1aAa

— Jason Furman (@jasonfurman) September 27, 2025

Furman was careful the same day. The 92% is a decomposition, not a movie of the world without AI. Absent the boom, he wrote, interest rates and electricity prices would likely have been lower, and other sectors would have grown more. In very rough terms, he said, that extra activity might replace about half of what the boom added. Matthew C. Klein, a trade economist, replied that imported machines and parts should be netted out before anyone treats the spending as domestic output.

Later readings still do not agree on the size of the shock, because they are not measuring the same object. The 2026 Economic Report of the President said investment in that equipment-and-software category grew at a 28% annual rate in the first half of 2025, after 5.5% in 2024. Van Nieuwerburgh has projected $10.3 trillion of data-center and related AI infrastructure spending from 2025 through 2032, an average of 3.6% of GDP a year. Goldman Sachs has put U.S. AI investment at 1.9% of GDP in 2026. Jan Hatzius, Goldman’s chief economist, has said AI is adding only about 0.1 percentage point to measured U.S. GDP growth, in part because so much of the hardware is made in Taiwan and South Korea.

WHERE EXPERTS DISAGREE

• Jason Furman: In H1 2025, information-processing equipment and software, 4% of GDP, accounted for 92% of measured growth; the rest of the economy printed 0.1%, though lower rates and power prices might have replaced about half of that boost.
• Jan Hatzius: After import leakage, AI is adding about 0.1 percentage point to U.S. GDP growth, so a large share of the capex shows up in East Asian output rather than American output.
• Reid Hoffman: The LinkedIn cofounder said on Oct. 3, 2026, that the infrastructure buildout is the only reason the country is not in a recession, with the spending spread well beyond the towns that host the campuses.

Smith’s version of the same point is that this demand is the dog that did not bark. Tariffs, deportation of construction labor, and a fog of executive-order risk should have cooled hiring. The campuses kept crews on jobsites instead. A common rule of thumb, Okun’s Law, says shaving 1 percentage point off growth is associated with almost a million lost jobs. That is a textbook translation, not a headcount from the Bureau of Labor Statistics, and Furman’s half-replacement caveat cuts the implied rescue in half. Even then, a shock that large is not a rounding error.

The Paychecks Are Going to the Trades

Krugman asked readers to look at nonresidential construction spending, which he said had flattened even as AI budgets exploded. Smith said the right series for a jobs argument is employment, not the dollar line, and that construction work has risen as a share of the workforce since late 2022, including through a 2025 pause he tied to deportations.

The Commerce Department’s private data-center construction numbers, running through July 2026, show why the two series can diverge. Seasonally adjusted spending on those campuses hit $37 billion, about $9 billion more than in the first seven months of 2025. Other private construction was about $46 billion below a year earlier. The boom is real. It is also crowding the rest of the building trades’ book of work.

THE JOBS THE CREWS CAN COUNT

• Union electricians: Don Slaiman of IBEW Local 26 said union electrician ranks in the Washington area rose from 9,000 to 17,500 as the campuses went up.
• Campus payrolls: LinkedIn estimates employers have added 117,000 jobs at data centers since the start of 2024, on top of construction, which the firm does not fold into that count.
• Broader AI listings: The same firm puts more than 750,000 U.S. jobs from 2023 through 2026 to date as AI-related, with a median listed salary near $180,000 against about $80,000 for all listings.
• Grid workers still needed: Goldman Sachs has said the U.S. power and grid chain will need about 500,000 extra workers by 2030, roughly 300,000 in generation and 200,000 on transmission and distribution, where electricians take about four years to train.

Gavin Baker, managing partner at Atreides Management, has become one of the boom’s blunt defenders. After walking back a sharper post, he wrote on Aug. 30, 2026, that demand for electricians, plumbers, welders, HVAC techs, and contractors has gone vertical, and that the work is not a one-season pour because the buildings get upgraded. Building-trades unions are now treating statewide bans as a threat to hours. An IBEW statement on proposed New England moratoriums warned that the bans would block investment and cut job opportunities for members.

That is the payroll Krugman’s BLS category for data-processing work does not capture. The people getting hired to stand the buildings up do not clock in as server operators. They clock in as electricians.

Loudoun County’s Budget Now Runs on Data Centers

Jobs are not the only way a windowless warehouse pays a town. Jared Walczak’s Tax Foundation paper on how states tax data center equipment walks through sales taxes, real property taxes, and tangible personal property taxes on the servers themselves. In a model of a $1 billion facility across 12 jurisdictions, tangible personal property is 20.6% of the total tax burden and real property is 36.5%. At the national average combined sales-tax rate of 7.52%, taxing machinery and equipment would cost that model center $58.3 million in year one and $9.7 million each year after. Most states exempt those purchases, often with job and investment strings attached.

Loudoun County, Virginia, is the worked example. The county’s own accounting, current as of March 1, 2026, is the cleanest primary record in the country.

WHAT LOUDOUN’S CAMPUSES PUT ON THE BOOKS

The same $1.2 billion in property tax revenue is why a jurisdiction of about 450,000 people can talk about schools, roads, and libraries without loading the whole bill onto houses. Another 35.7 million square feet is approved but unbuilt. Twenty-one more applications cover about 20.7 million square feet. From 2022 through May 2026, the Board of Supervisors approved 32 of 39 legislative applications.

That is a fiscal machine, and it is a concentration risk. Virginia’s data-center sales-tax exemption cost an estimated $1.6 billion in fiscal 2025. Smith is right that towns can still raise fees, rewrite community-benefit deals, and tax a campus after it opens, when moving the servers is hard. Jared Bernstein has argued those deals should include cheaper local power and water, money for the substations and pipes the sites need, and cash for schools and parks. Plenty of counties have not driven that bargain. Loudoun did, early, and now 40% of next year’s budget is a forecast about servers.

The Same Boom Is Bidding Away Power and Labor

Brookings researchers Dany Bahar and Greg Wright assembled about 1,500 U.S. facilities and compared counties that built them with counties that announced projects and then canceled. Their finding is the cold water in the jobs argument. A typical treated county sees about 100 to 200 local jobs per county over a decade, with data-processing employment up 56% from a small base and telecommunications up 43% in hyperscale markets. Wages did not move. Home prices rose 2% to 5%. Industry reports that skip the canceled-project control overstate the hiring.

The costs on the other side of the ledger are easier to see from a kitchen than from a GDP table.

THE TAB THAT DOES NOT SHOW UP IN CHATBOTS

• Power auctions: In the PJM region, which serves 65 million people across 13 states, power-supply costs jumped from $2.2 billion to $14.7 billion in a single year, with data centers accounting for nearly two-thirds of the increase.
• Household rates: Residential electricity prices nationally rose about 32% between July 2020 and July 2025.
• Other builders: Private construction outside data centers ran $46 billion below year-earlier levels through July 2026, while the campuses ran $9 billion above.
• Factory power: A planned Mississippi aluminum smelter that would have employed about 1,000 people was reported to have lost its electricity allocation to a data-center project near Vicksburg and gone to Oklahoma instead.
• Local vetoes: More than 100 communities have enacted moratoriums. Lawmakers filed more than 300 state data-center bills in the first six weeks of 2026. Sen. Bernie Sanders and Rep. Alexandria Ocasio-Cortez introduced a federal pause on large AI campuses on March 25, 2026.

Baker says well-structured projects now force the campus to bring or pay for new generation, and that towns which skip that deal are right to object. Hoffman says operators need to cut fair bargains with neighbors. Both are describing a political tax, not a free lunch. The Tax Foundation’s own national sketch is that U.S. data-center investment will exceed $1 trillion over the next five years, and that hyperscalers may need another $1.8 trillion in capital by 2030. Money on that scale does not sit in one sector. It bids for electricians, transformers, gas, land, and grid connections until something else loses.

The office side of the ledger is quieter and uglier for people who did what they were told. Routine customer-service and data-entry work is the early loser. Stanford-linked work on young workers in AI-exposed white-collar jobs has already shown slower hiring for people in their early twenties relative to jobs a model cannot do. The boom can be good for a journeyman wireman and still be a closed door for a new graduate who trained for a laptop job.

Users Already Bank $172 Billion the GDP Misses

Krugman’s essay treats the technology and the construction binge as the same object, and jobs as the main way a normal household might benefit. That leaves out the product people already have on their phones. Erik Brynjolfsson, Avinash Collis, Felix Eggers, Sophia Kazinnik, and David Nguyen ran choice experiments in July 2025 and March 2026, asking a representative sample of U.S. adults how much they would have to be paid to give up chatbots such as ChatGPT, Gemini, Claude, or Copilot for a month.

Mean willingness to accept rose from $98 to $124.50, a 27% increase. The median rose from $3.4 to $11.40. The adult user base grew from 98 million to 115 million. Those inputs produce $172 billion in yearly consumer surplus, up from $116 billion. The mean is being pulled by heavy users. The typical user is closer to the median, which is still more than triple what it was a year earlier, and still a long way from a luxury good. The authors note that this surplus is larger than estimated U.S. generative-AI revenue, which is another way of saying households, not the vendors, are capturing most of the welfare so far, largely because so much of the product is free.

That is a real benefit, and it does not show up in GDP, in a Loudoun tax file, or in an IBEW dispatch. It also does not cancel a higher power bill in the PJM footprint, or a county that has bet 40% of next year’s budget on servers staying full. Smith is right that Krugman’s oligarch line does not survive contact with those numbers. Krugman is right that a capital binge this large, with so much of the kit imported and so few permanent operators per building, is a strange way to run a full-employment machine.

Loudoun has already booked $1.3 billion in data-center tax for fiscal 2027. That number assumes the racks stay plugged in, the crews keep pulling wire, and the rest of the economy can live with the power and the interest rates the boom is using.