Jennifer Harris published a piece in the New York Times last week that everyone in our world should read, and then read again. In “The Generational Force Hollowing Out the Economy,” Harris makes the case that AI investment has grown so large, so fast, that it is starting to choke off capital from the rest of the economy. She’s talking about a buildout on pace to top a trillion dollars a year, money that used to flow into housing, manufacturing, energy infrastructure, and the thousands of ordinary, non-AI startups that used to make up the bulk of venture activity. Harris backs this up with real sourcing, and she deserves full credit for putting a name on something a lot of people in tech have felt for a while but hadn’t quite articulated.
She’s right about the capital. I just think she’s describing the first symptom of something much bigger.
Money is only one scarce resource, and once you start looking at AI through that wider lens, the crowding-out story Harris tells starts to look like just the opening chapter. AI isn’t simply pulling investment dollars away from other sectors. It has become the dominant gravitational force in the economy, and it is absorbing nearly every scarce resource we have.
Start with the physical world, because that’s where the evidence is easiest to see and hardest to argue with. Industrial real estate that used to go to warehouses and light manufacturing is now data center real estate, and it’s getting bid away at prices that make traditional developers wince.
Construction spending on data centers hit roughly $49.5 billion through April of this year, nearly four times what it was over the same stretch last year, and the category has now surpassed spending on the country’s entire transportation infrastructure, airports, rail terminals and marinas included. Steel, concrete, switchgear, transformers, cooling systems- all of it is being routed toward hyperscale campuses at a pace the supply chains weren’t built for.
Electric power is the tightest constraint of all, and it’s not close. A single AI training facility can draw 100 to 500 megawatts continuously, which is what a small city uses. Data centers could account for somewhere between 9 and 17% of total U.S. electricity consumption by 2030, according to EPRI’s most recent modeling, up from roughly 4 to 5% today. In Virginia, the country’s data center capital, facilities already consume more than a quarter of the state’s electricity, and that could climb past half by the end of the decade. Utilities are raising rates to fund the buildout, wholesale capacity prices in the PJM market jumped more than tenfold in two years, and grid planners are approving projects on timelines that used to be reserved for public works.
Transmission and generation equipment are backed up for years. High-voltage transformers now carry lead times of three to five years, and the shortage has pushed hyperscalers into building their own power behind the meter, natural gas turbines, fuel cells, small modular reactors, anything that lets them skip the queue. Semiconductor fabrication capacity is booked out just as far, with foundries prioritizing AI accelerators over the chips that go into cars, appliances and industrial equipment. None of this is theoretical. It’s showing up in freight schedules, permitting queues and capital budgets across a dozen unrelated industries.
Then there’s labor, and this is where the story gets personal for a lot of people who have nothing to do with AI. The construction industry needs roughly 350,000 net new workers this year just to keep up with demand, and data centers are eating an outsized share of that need. One recent industry estimate puts the data center labor shortfall alone at close to half a million workers, with electricians at the center of it. Electrical systems account for somewhere between 45 and 70% of total data center construction costs, and Microsoft’s own leadership has called the electrician shortage the single biggest bottleneck slowing the entire buildout. Wages for electricians, HVAC technicians and pipefitters on these projects are running 30 percent above traditional construction pay, which is pulling skilled trades away from hospitals, schools and ordinary commercial projects that can’t compete on price.
Engineering talent is following the same gravity. Every AI lab, every hyperscaler, every well-funded startup is bidding for the same small pool of machine learning researchers and infrastructure engineers, and salaries have moved into territory that makes it nearly impossible for climate tech, biotech or fintech companies to compete for the same graduates. Venture capital has followed the talent. A shrinking share of new funds is willing to back a startup that isn’t wrapped in an AI thesis, and founders have figured out that the fastest way to raise money is to make sure the word appears early in the pitch, whether or not it belongs there.
Executive attention has moved in the same direction, and this one is harder to put a number on, but not hard to observe. Sit through a few earnings calls or board meetings this year and count how many minutes go by before someone mentions AI. Government policy has followed too, with export controls, chip legislation, energy permitting reform and workforce programs all being written with AI as the organizing principle rather than one input among many. Media coverage, conference agendas and research priorities across academia have shifted in the same way. I run publications that cover this world for a living, and I can tell you plainly that AI now sets the agenda for nearly everything else we cover, not because we decided it should, but because that’s where the readers, the sponsors and the news are.
Every hour a brilliant engineer, scientist, founder or policymaker spends thinking about AI is an hour they are not spending on something else. Society only has so much elite human cognition available at any given time, and an extraordinary share of it is currently pointed at one problem. I’m not arguing that’s wrong. Some of the smartest people I know believe AI is worth every ounce of attention it’s getting, and they may turn out to be right. But cognition is scarce the same way capital and electricity are scarce, and right now it’s being allocated the same way everything else is: toward AI, almost by default.
None of this makes AI a bad investment. I want to be clear about that, because it would be easy to read what I’ve written so far as a warning shot, and it isn’t one. Much of this spending is probably justified. AI may produce returns over the next decade that dwarf what’s being poured in today, and the people making these bets aren’t fools. The issue isn’t whether AI deserves the resources it’s getting. The issue is opportunity cost, and every technological revolution creates opportunity cost. What makes this one different is that AI is competing for financial capital, physical infrastructure, skilled labor, engineering talent, attention and cognition all at once, in the same years, in overlapping markets, with no other era’s technology having done quite that.
This is where I want to draw a line between two ideas that get used interchangeably but aren’t the same thing. Economists talk about crowding out, which describes scarce inputs flowing from one sector into another. Capital that would have funded a biotech startup instead funds a data center. A steel order that would have gone to a bridge project instead goes to a hyperscale campus. That’s real, and Harris documents it well.
But something is happening underneath the input story, and I think it’s the more consequential one. Crowding out is about where the inputs go. Hollowing out is about what happens to the institutions those inputs used to sustain, once enough of that flow gets redirected somewhere else. A startup ecosystem still technically exists, but funding increasingly concentrates around a narrow AI thesis, and the founders working outside it are finding the well drier than it’s been in years. A software company still exists, but it now needs fewer engineers to ship more product, and the org chart that used to define a mid-size tech company is quietly getting thinner. Universities still exist, but AI is performing more of the functions that used to justify tuition, from tutoring to research assistance to first drafts of scholarship. Consulting firms still exist, but the junior analyst tier that used to train the next generation of partners is shrinking fast. Media companies still exist, mine included, but the economics of how information gets produced, distributed and paid for are being rewritten underneath us in real time.
Nothing on the outside looks different yet. The logo is the same, the building is the same, the name on the door is the same. But the economic substance inside these institutions is migrating somewhere else, quietly, one budget cycle at a time, and that’s a very different phenomenon than money simply moving from sector A to sector B. Crowding out shows up in an investment table. Hollowing out shows up ten years later, when someone asks why an entire category of company or institution doesn’t function the way it used to and nobody can point to the exact moment it changed.
We’ve been through resource-hungry technology cycles before, and it’s worth being honest about what history actually shows rather than reaching for a tidy parallel. Railroads absorbed a huge share of American capital in the 1850s, and when the promised returns didn’t show up on schedule, the country sank into a depression that wiped out thousands of businesses. Electrification in the 1920s and the dot-com buildout of the late 1990s followed similar patterns of aggressive investment followed by painful correction. Central bankers who study these cycles have drawn exactly that comparison to the current AI buildout, and they’re not wrong to. But every one of those earlier revolutions was primarily a capital story, with labor and infrastructure effects that followed downstream. AI is the first one I can point to that is simultaneously bidding for capital, electricity, transmission infrastructure, industrial real estate, skilled trade labor, engineering talent and human attention, all in the same three- or four-year window, which means the hollowing out, if it’s happening, is happening on every front at once rather than one at a time.
There’s one more piece of this worth naming before I close, because it ties everything else together. Attention is itself an economic resource, and we don’t usually treat it that way. Every hour of airtime an earnings call gives to AI is an hour it doesn’t give to something else. Every keynote slot at a conference that goes to an AI product demo is a slot some other category of innovation doesn’t get. Every board meeting that opens with an AI update is one where some other strategic question gets pushed to next quarter. That’s not a cultural observation. It’s a resource allocation decision, made dozens of times a day, across thousands of companies, almost entirely without anyone deciding to make it on purpose.
So here’s where I land, and I don’t think it’s a doom-and-gloom place to end up. Jennifer Harris identified something real and important, and the capital story she tells is the most measurable piece of a much larger pattern. But measurable isn’t the same as complete. The bigger challenge in front of us isn’t whether AI is crowding out other investments, because it clearly is, and the case for a lot of that spending is genuinely strong. The bigger challenge is whether we’re going to make these tradeoffs on purpose, with our eyes open, deciding deliberately how much capital, power, labor, talent and attention we want AI to consume versus everything else competing for the same finite pool. Or whether we’re just going to let AI’s economic gravity keep making that decision for us, one earnings call, one budget cycle, one hiring freeze at a time, until we look up in a few years and realize the shape of the economy changed while we were all busy talking about something else.
