In a field outside Manassas, Virginia, a transformer the size of a small house is being lowered into place by a 400-ton crane. When the data center it serves comes online next year, it will consume more electricity than the town next door. Two counties away, the Board of Supervisors has stopped approving new construction. The grid has run out of capacity.
This is what artificial intelligence looks like at the scale we are currently building it. Not chatbots. Not robots. Not the sleek interfaces on your phone, but steel, silicon, and concrete. And — more than anything else — electricity.
In 2026, five American companies will spend roughly $700 billion on building this infrastructure. That amount exceeds the entire Swiss economy. Almost none of it will be spent in Europe. About three-quarters of the money will go directly to AI components: chips, servers, cooling systems, the buildings to house them, and the power plants to keep them running.
It’s the largest, fastest, and most concentrated infrastructure buildout in modern history. And the people building it cannot stop.
The question worth asking is whether they should.
That is not a moral question. It’s a diagnostic one. When five companies are spending more than they earn, borrowing the difference, and lending to each other in arrangements complex enough that even Wall Street analysts are debating whether the companies are propping each other up — something is straining. When the most reliable predictor of these companies’ stock prices a year ago was their competitors’ stock prices, and now that correlation has dropped from 80% to 20%, the herd is splitting.
The pattern is familiar. In the late 1990s, telecom companies built fiber-optic networks well ahead of demand. Equipment manufacturers helped finance their customers’ purchases. Markets believed every projection. When demand fell short, the financing structure collapsed, and the industry failed. Most of those companies are gone. The fiber is still in the ground.
Senior figures in AI are aware of all this. Senior figures in the industry have defended these arrangements as appropriate, given the technology's transformational potential, when asked about them during earnings calls. The typical response is that this time is different.
Maybe. Some things are genuinely different. The technology works, and people use it. The product reached widespread adoption faster than any consumer technology before it. Whether “this time is different” turns out to be true depends on something nobody can yet measure: whether the buildout can be maintained until the technology begins to pay for itself.
That payment is overdue. Microsoft, the most disciplined among the spenders, generated about $25 - $30 billion in AI revenue compared to $100 billion in capital spending. The math doesn’t add up yet. The bet — and it is a bet — is that it will.
There is a particular kind of mistake that an outsider might make when examining all of this, and it is worth naming because insiders are making it too.
When you read the headlines, four different sources seem to confirm the same story. The frontier AI labs report their safety evaluations. The hyperscalers — led by AWS, Microsoft Azure, and Google Cloud — project their enterprise revenue. The chipmakers announce their order books. The Wall Street analysts publish their forecasts. Four independent sources, four different industries, four sets of qualified experts — and they all generally agree on the same story about the AI future.
The convergence feels like confirmation. Four sources agreeing means the picture is probably right.
It does not mean that. It means the four sources might be reading from the same spreadsheet. The labs are estimating their growth based on hyperscaler commitments. The hyperscalers are estimating their commitments based on lab capability promises. The chipmakers are estimating their orders based on hyperscaler capital expenditure plans. The analysts are estimating it all based on the projections from the first three. If you trace the dependencies back, every figure in every projection relies on the same fundamental assumption: that the capital committed today will eventually generate enough revenue to justify itself.
When four supposedly independent sources keep agreeing, the key question is whether they are truly independent or whether all four are conditioned on the same upstream bet. In the current AI transition, the truth is mostly the latter. The convergence is real; the independence is not.
This matters because it’s the mechanism by which a system this large can stay confidently on course even while heading toward a wall. Each source feels like external validation. None of them truly is. The system is processing its own bet back to itself in four different ways and interpreting the result as confirmation.
Meanwhile, in a different room, a 24-year-old who graduated last year with a finance degree is sending out her fortieth resume. She gets fewer callbacks than her older sister did at the same age. She doesn’t know, because she can’t see, that her resume is being filtered by an AI tool before any human reviews it — a tool that learns from past hires and therefore looks for the kind of person who has already been hired. She also doesn’t realize that the entry-level analyst work she trained for — the grunt work that taught her sister how the job actually worked — is exactly the work the new AI tools are beginning to perform.
She’s angry, and she doesn’t yet fully understand why. But the polls are starting to show it. Among Americans her age, the percentage who say they feel excited about AI dropped from 36% to 22% in just one year. The percentage who say they feel angry increased from 22% to 31%. The reason is that the door to her career is closing just as she was about to walk through it.
Two things can be true at once: the technology is real, and the buildout is overheated. The workers are anxious, and the leaders are convinced. The public no longer trusts the companies, the companies no longer trust the regulators, and the regulators no longer trust each other across borders. Trust is what greases a system this complex. To say that it’s running low is an understatement.
Look at the AI complex carefully, and you find five things approaching each other all at once.
The grid is the first. Power has shifted from “we’ll figure it out” to “we can’t build fast enough.” Loudoun and Prince William counties are at or near capacity. Microsoft has signed a contract to restart a reactor at the Three Mile Island nuclear plant. Amazon has bought a billion dollars of dedicated solar in Texas. Whatever the AI race looks like in five years, the binding limit between now and then is electrons.
The financial structure is the second. Capital is now being borrowed, not earned, to fund the buildout. Vendor financing — where companies invest in their own customers — is making a comeback. These arrangements are complex, and they’re also fragile. Even Nvidia, the company at the center of the boom, has quietly informed its shareholders that one of its major customer deals may not come to fruition.
The labor signal is the third. Hiring for entry-level white-collar roles has been tightening for two years. The unemployment rate hasn’t moved much yet because the people who already have those jobs still do. But the door to the room is closing behind them. In five to seven years, when those who were supposed to fill that room are not there, the consequences will become structural — and irreversible on any short timeline.
The trust signal is the fourth. Two-thirds of Americans say AI companies are not being honest about their actions. Two-thirds also think the government is not doing enough about it. Trust in the U.S. government to regulate AI is the lowest among any country surveyed — 31%, compared to 81% in Singapore. Once trust falls that far, it doesn’t recover on its own.
The geopolitical signal is the fifth. The United States is in the middle of an incoherent moment with China — restricting some advanced chip exports while allowing others under a 25% revenue tariff, with Congress moving to restore the restrictions the executive branch eased. “We’re allowing them to do it,” the President said, “but the United States is getting 25% of the chips, in terms of the dollar value.” Meanwhile, Beijing is spending hundreds of billions to build its own version of the same infrastructure. Both countries are investing heavily in something that should not have two incompatible standards. It currently does.
These five thresholds are not independent. They are connected through the same loop that drives the buildout — capability, expectation, capital, construction, more capability — and pushing one harder also pushes the others closer. When more than one boundary is reached simultaneously, the boundaries interact. They don’t fail one after another; they fail together, or not at all.
Energy is the constraint that forces the question. It’s the barrier the buildout will hit first, the physical limit that no one has figured out how to challenge. But energy isn’t the element the system can’t bypass. Vendor financing is a workaround. Behind-the-meter natural gas is a workaround. Federal preemption of state regulation is a workaround. Each is a way the mechanism is currently expanding its options without changing its core actions.
The thing the system cannot work around is the underlying logic — the belief, established in late 2022 and now embedded in every frontier lab, that someone has to build this, and it must be us. That belief is not a strategy; it is part of their identity. Asking the labs to slow down is not asking them to change their plans. It is asking them to change who they are. They can’t do that, and the request itself triggers exactly the defensive reaction that makes the conversation impossible.
This is why energy reform is the leverage point that matters between now and 2031. Not because energy is the most critical issue — it isn’t. Because energy is the constraint that the system can be pushed against without requiring it to renegotiate its core identity. The more profound discussion happens later, after easier interventions have rebuilt enough trust and signal clarity for the deeper dialogue to be receivable. Right now, the deeper conversation is one the participants cannot have.
What probably happens is this. Sometime between late 2026 and 2028, a financial event — like a missed earnings call, a downgraded bond, or a deal that quietly falls apart — exposes how much of the current AI revenue is being recycled among the same five or six companies. Markets correct themselves. Smaller AI startups and chip firms fail. The big firms survive, though leaner, and start to consolidate. Construction slows down. Some data centers remain half-built. Trust doesn’t fully recover, but it doesn’t completely collapse either. By 2031, AI is everywhere — more like electricity in 1900 than the internet in 2000.
That is the most likely path, but not the only one. A lower probability suggests a trigger event — such as a major safety incident, a politically explosive labor announcement, or a Taiwan crisis — could force a faster, messier resolution. An even smaller probability indicates the buildout works as advertised, revenue catches up, and the smooth transition the industry promises actually occurs. The optimistic scenario requires every threshold to hold for five years simultaneously. That’s a lot of coin flips.
The public has more leverage than it knows, but the leverage is in the boring places.
State-level reforms on how power lines are permitted and how data centers connect to the grid will influence the AI expansion more than any AI law. Disclosure standards on financing arrangements between AI companies and their customers — quiet rules in obscure SEC committee rooms — would help credit markets distinguish real demand from recycled revenue. Independent evaluations of AI’s capabilities and limits, conducted by groups that do not profit when AI succeeds, would bridge the gap between marketing claims and reality. Educational and labor pipelines designed for the world that is actually coming, not the one that is leaving, would slow the collapse of the entry-level on-ramp before the senior pipeline downstream runs dry.
These are early-stage moves. They sound small because they are small, and they are also the only moves the system can currently handle. The larger moves that might be needed later — such as independent evaluation with binding authority over deployment, genuine public stakes in the infrastructure, and an effective international governance regime with enforcement powers — are not yet possible. The trust required to make them work hasn’t been built, and the signal integrity needed to verify their effectiveness hasn’t been restored. The system isn’t ready, and pretending otherwise would only lead to the bigger moves being used as reasons for the system to defend itself more aggressively.
The first-stage moves are how the bigger moves become possible. Skipping them and jumping directly to bigger moves — which a frustrated public might be tempted to do — leads to backlash, not progress. The sequence is important. Not because the small moves are sufficient — sometimes they are — but because they lay the groundwork for the larger moves to succeed when the time comes.
None of this will trend on social media. All of it is within reach of ordinary democratic action — state legislators, federal agencies, school boards, regulators, voters. The most consequential decisions about the AI transition will be made in rooms most people will never enter, by people most of us will never meet.
In the field outside Manassas, the transformer is in place. The cooling systems hum. A circuit closes, and the data center comes online. It will run for thirty years.
The choices being made right now — about how it is paid for, who buys what it sells, who gets hurt by it, who profits from it, and what it’s allowed to do — are being pressed into wet cement. Five years from now, much of that cement will have set.
We’re not too late, but we’re running out of slack.