The numbers have stopped making sense. The five largest American cloud and AI companies are guiding toward somewhere between $635 and $690 billion in capital spending this year alone — more than double what they spent in 2024. Counting the fourteen biggest data center operators worldwide, the figure approaches three-quarters of a trillion dollars in a single year. By some estimates, AI infrastructure now represents roughly five percent of U.S. GDP, making it the largest infrastructure commitment in modern American history — bigger than the fiber boom, bigger than the peak of electrification. Power, land, chips, concrete, transformers: all of it is going berserk.
And this month, one of the companies at the center of it, Oracle, watched its bonds start trading like junk. That is the kind of detail that should make everyone stop and look up.
We are going to say plainly what a lot of people in and around this industry will only say quietly: this is a bubble. But we want to be equally clear about what that word does and does not mean — because the loudest voices on both sides are getting it wrong.
A Bubble Is Not a Fraud
When we say bubble, we do not mean the technology is fake. Artificial intelligence is real, useful, and in some domains genuinely transformative. We use these tools ourselves, and we intend to keep taking advantage of every tool we can. The mistake people make is assuming that if the technology is real, the prices must be rational. Those are two entirely different questions.
History is full of real, world-changing technologies wrapped inside irrational financing. The internet was real in 2000, and it still took the dot-com crash to sort the survivors from the pretenders. Houses were real, and people genuinely needed them, in 2006 — which did not stop the housing market from nearly taking down the global economy two years later. Oil is as real as a commodity gets, and it has boomed and busted on a roughly predictable cycle for over a century. In each case, the underlying thing had lasting value. The prices, the leverage, and the timing did not. That is exactly where we are with AI infrastructure today.
The Warning Lights Are Blinking
Look at the evidence honestly and the picture is hard to miss.
Start with the bond market, which tends to panic earlier and more soberly than the stock market. In July, S&P cut Oracle's credit rating to one notch above junk, citing an uncertain path to profitability. The company has taken on roughly $130 billion in debt to build AI data centers, its free cash flow has turned deeply negative — it burned through nearly $24 billion after capital spending in its last fiscal year — and the cost of insuring its debt against default has climbed to levels not seen since the 2008 financial crisis. Oracle is not alone in the direction it is heading; it is just further down the road. The hyperscalers as a group have shifted from funding their buildouts with operating cash to funding them with borrowed money, and Morgan Stanley expects their debt issuance to top $400 billion.
Then look at the gap between spending and revenue. By various estimates, the AI sector is burning something like $400 billion a year while generating perhaps $50 to $60 billion in revenue. OpenAI has reportedly committed on the order of $150 billion against a small fraction of that in sales. Adoption is wide but shallow — the overwhelming majority of companies are experimenting with AI in some corner of their business, but fewer than half have scaled it past the pilot stage into something that actually pays for itself. That gap between what is being spent and what is being earned is the defining feature of every bubble ever inflated.
In June, the Bank for International Settlements — the central bank for central banks, not exactly a hype machine — warned that an AI capex bust, the opaque circular financing deals knitting these companies together, and record sovereign debt were three of the most likely triggers for the next crack in the global financial system. When the BIS and the bond market are both nervous, it is worth being nervous with them.
Who Pays for the Power?
There is a second story tangled up in the first, and it hits ordinary people more directly than any stock price: the power bill.
AI data centers are extraordinarily hungry for electricity, and that demand is running straight into a grid that was not built for it. Projections have AI-driven power demand reaching well over 150 gigawatts by the end of the decade, requiring trillions in cumulative investment. Someone has to pay for all those new lines, substations, and generators. And so far, too much of that cost has landed on regular ratepayers. In data-center-heavy Virginia, residential electricity prices have climbed as much as 267 percent over five years. Nationally, data centers have already added billions of dollars to household power bills.
To their credit, some utilities and state regulators have started pushing back — creating special "large load" tariffs to force data centers to pay for the infrastructure they require, and in some cases to post protection against simply walking away from a half-built project. More than thirty states have proposed or adopted measures like this, and federal regulators have ordered at least one major grid operator to make developers pay their fair share. The tech companies have signed a nonbinding pledge to cover their own power costs.
But here is the hard part, and it is the part you identified correctly: data centers are mobile, and utilities are not. A developer choosing where to build can shop across dozens of utilities and dozens of states, and it will tend to land wherever the fees are lowest and the concessions richest. That creates a race to the bottom, in which the jurisdiction that charges the least wins the project — and its ratepayers inherit the risk. A pledge here and a tariff there will not solve that; only broad, consistent standards will. Otherwise ordinary families end up quietly subsidizing a speculative boom they never invested in and will not profit from.
The Correction Is Coming — and That Is Not a Tragedy
Capital cycles are old and reliable. Money floods toward the hot thing at the top, builds far more capacity than near-term demand can use, and then — when the returns disappoint — financing pulls back, projects stall, and the weakest, most leveraged players fail. There is no reason to think AI has repealed that law. Analysts are already penciling in corrections of twenty to thirty percent in the most AI-exposed stocks, most likely as a gradual deflation over 2026 and 2027 rather than a single dramatic crash.
The earliest cracks are already visible, and this is where we want to be careful and fair rather than alarmist. The clearest ones are at the consumer end: the run-up in memory and flash prices has been so extreme — with some contract prices nearly doubling in a single quarter — that ordinary buyers are hitting an affordability wall, and those breakneck price increases have begun to cool. On the server side, demand is still strong for now, but it is increasingly constrained by supply rather than looking insatiable, and the classic early tells of a turning semiconductor cycle — softening unit demand and inventories quietly rebuilding — are exactly the signals worth watching from here. None of that guarantees the timing. It does say the pressure is building in the familiar places.
This Does Not Mean AI Goes Away
Now the other half of the truth, the half the doomsayers miss. When this corrects, artificial intelligence does not disappear. The data centers do not evaporate. The power lines stay in the ground.
Think about what happened to all that fiber-optic cable laid during the dot-com frenzy. The companies that installed it mostly went bankrupt, and for years much of the fiber sat unused — "dark," in the industry's word. But it did not vanish. It became the cheap backbone that carried the internet, streaming, and the cloud through the next two decades, bought up for pennies on the dollar by the survivors. Housing did not stop mattering after 2008. Oil did not stop mattering after any of its busts. The pattern is consistent: the infrastructure outlives the financing that built it. It simply changes hands, usually at a steep discount, and the disciplined survivors go on to build the durable industry on top of it.
AI will be no different. The chips, the buildings, and the megawatts will still be here the morning after the correction, and they will still get used. What will not survive intact is a great deal of the capital — and the reputations — of the people who bid the hardest at the top.
The Common Sense Position
So what do prudent people and prudent policymakers do with all this?
For anyone with money at stake, the message is the one we always come back to: do not let the fear of missing out make the decision. Keep using the tools — we certainly will — but be careful with your investments, and assume that today's prices embed an overbuild and an eventual correction. FOMO is precisely the emotion that bubbles use to separate ordinary people from their savings, right at the moment when caution would serve them best. The excitement is not a strategy.
For policymakers, the right role is neither to inflate the boom with subsidies nor to try to pop it by decree, but to let the market do its necessary, clarifying work while keeping the damage contained to the people who actually took the risk. That means letting over-leveraged, over-optimistic projects fail rather than bailing them out. It means insisting that data centers pay for the power infrastructure they demand, through broad and consistent rules, so that families are not left holding the bill for capacity built on a speculative bet. It means utilities planning their buildouts soberly, with the near-certainty of an overbuild and correction in view, so they do not strand ratepayers with the cost of a boom that busts. Some projects will stop. Some should. The task is to make sure the decisions about power and construction are made as prudently as possible, precisely because we can see the correction coming.
We are, on this one, both optimists and skeptics at the same time — optimistic about the technology, skeptical of the price. Both can be true, and holding them together is the whole point. The bubble will deflate. AI will still be here when it does. The only real questions are who ends up absorbing the losses, and whether we were wise enough to make sure it is not the families who never got swept up in the frenzy to begin with.