Every infrastructure boom starts with a demand forecast that nobody can really check. How much traffic would the railways carry? How fast would the internet grow? How much compute will AI need?
The forecast is plausible, so capital commits to it. The money moves, the ground breaks, and the concrete sets long before anyone can prove the demand was ever there. It often ends the same way. The technology was right, the scale and timing were wrong, and the ones who financed it were long gone before it paid off.
Today we are rushing to build AI data centers. The forecast may hold, but if it doesn’t, the bill lands somewhere, and probably not where we would look for it.
The Debt That Is Not There
A Nikkei investigation published in July 2026 found that five US tech giants — Alphabet, Microsoft, Amazon, Meta and Oracle — now carry an estimated $1.65 trillion in off-balance-sheet liabilities. This is eight times more than four years ago, and it is more than their combined reported debt of about $1.35 trillion. Meta alone has around $420 billion off its balance sheet, close to three times what it shows on its books.
Moody‘s looked at the same companies from a different angle and found a similar picture: $662 billion in data center lease commitments that have not started yet, equal to about 113% of the five companies’ combined adjusted debt.
That gap exists because of how these deals are structured. In most cases, hyperscalers do not build data centers, infrastructure funds and developers do. And what makes the project bankable is a lease agreement signed by a hyperscaler and the party that actually owns the data center. Under current accounting rules, a lease for a building that is not operating yet is not a liability on the hyperscaler’s balance sheet. The capacity gets built and the money gets committed, but the risk sits in a place where investors do not really see it.
History Has a Pattern
This is not the first time money has rushed ahead of proof. Bubbles appear every time a technology revolution meets a lot of available capital and a fear of being left behind.
In Britain in the 1840s, private investors were putting roughly twice as much money into railway construction as the government was spending on its military. Parliament approved thousands of miles of new track. Then the crash came, and most of those investors lost their money. The railways were built and Britain still uses that network today. But the people who financed it did not get the return.
In the late 1990s, telecom companies laid more than 80 million miles of fiber in the US, expecting internet traffic to keep doubling every few months. By 2002, after the dot-com crash, only about 2.7% of that fiber was being used. Global Crossing and WorldCom went bankrupt. The fiber stayed in the ground until cloud computing and streaming finally needed it, years later.
In China between 2009 and 2013, solar manufacturers doubled their production capacity on cheap state credit, expecting global demand to keep growing. When European subsidies were cut after the 2011 Eurozone crisis, solar panel prices dropped about 40% in a single year. Suntech, at the time the largest solar panel maker in the world, went bankrupt.
Three centuries, three technologies, one pattern. The technology was real, and so was the long-term demand. But the capacity arrived ahead of it, and the first wave of capital was wiped out before the demand it bet on ever arrived.
How Solid Is the Forecast?
The buildout of AI data centers is being justified by very aggressive forecasts. Forecasts of compute demand, forecasts of electricity demand, forecasts of shortage. Everyone is committing capacity now because nobody wants to be the one without it later. This is how the previous cycles started.
But there is a second trend that does not get enough attention. AI models are getting more efficient, not only bigger. DeepSeek showed that you can get close to frontier performance without the best chips, at a much lower cost, and the Chinese labs keep pushing in that direction because they have no other option. And for most day-to-day work, you might not need a frontier model at all. Small models, sometimes running locally on your own machine, are enough. The very large models will still matter, but for a narrower set of applications.
I want to be fair to the counter-argument here. Historically, when a technology becomes cheaper, we do not use less of it, we use more. Cheap bandwidth after the fiber glut did not stay unused, it gave us YouTube and streaming and everything after that. The same could happen with cheap inference: efficiency creates new uses instead of reducing demand.
Both forces are probably working at the same time. The long-term demand is likely real. Whether it arrives at the size and speed these deals assume is a different question, and the one that matters — because the commitments are already signed. The AI data centers are full today, but that measures demand now, not the demand these fifteen-year leases assume.
Follow the Money Trail
So the question is not whether an AI demand slowdown hurts, but who it hurts. And it does not hurt everyone the same way.
Alphabet, Microsoft, Amazon and Meta have a profitable business underneath the AI spending. They would take the writedown, get repriced, and continue. The pure-AI companies do not have that. The model labs live on billions in funding, and if that flow slows, some of them do not survive it. For them, a slower demand curve is not a bad two years, it might be the end.
But none of them own the data centers. The labs create the demand, the hyperscalers sign for the capacity, but the buildings belong to someone else. So the demand for compute can disappear with the companies that caused it, and still leave the asset behind. That is the layer most people never look at: the infrastructure funds holding data centers built for tenants who no longer need them. The lease commitments that were never visible on a balance sheet suddenly become very visible to whoever is on the other side.
Now follow the money. Infrastructure funds have backers. Those backers are often banks and pension funds. This is how a sector correction becomes something wider, and it is the same way railway shares and telecom debt spread the pain far beyond the companies that failed.
I would not go as far as saying these assets become worthless. AI data centers are physical assets with power, land, cooling and grid connections, and history says that capacity eventually gets used — the fiber was used, the solar capacity was used. But the people who owned the assets in the meantime usually did not survive to see the returns.
So, Is History Doing It Again?
I am not saying AI is a bubble in the sense that companies with no product were a bubble in 2000. AI is real and it will stay. The question I keep coming back to is narrower than that.
Is history doing it again? A real technology revolution, large infrastructure built ahead of an inflated demand, and financing structured in a way that keeps the risk out of sight until it cannot be kept there anymore?
I do not know the answer. But the pattern has appeared often enough, in enough different industries and enough different centuries, that it seems worth asking the question now.