AI Boom Is Buying From Itself
Let's start with a thought experiment that sounds too stupid to be real.
You own a factory that makes shovels. A guy wants to start a gold mine but he has no money. So you hand him 10 billion dollars. He immediately spends all 10 billion buying shovels from you. You record 10 billion dollars in sales. Your stock price goes up because sales are booming. Analysts write reports about surging demand for shovels.
Now ask yourself the awkward question. Did anything actually happen?
Some money moved in a circle. A number on your income statement went up. The guy has a hole in the ground and a lot of shovels. Whether any gold ever comes out of it is a question for later, and by then you have already banked the revenue and the stock gain.
This is not a hypothetical. This is roughly how a meaningful chunk of the AI boom is currently being financed, and by 2026 estimates the total value of these looping deals is north of 800 billion dollars.
Nobody is doing anything illegal. Everyone involved has a reasonable explanation. But the shape of the thing is worth understanding, because the shape is what tells you who gets hurt if the gold never shows up.
How the loop actually works
Here's the basic circuit, stripped down.
A chip maker invests billions into an AI lab. The AI lab uses that money to commit hundreds of billions to cloud providers for computing capacity. Those cloud providers turn around and spend that money buying chips from the chip maker. Everybody in the chain books revenue, or at least backlog, off the same original pile of money.
One dollar, counted three times, in three different press releases.
The specific deals are staggering in size. In 2026, Nvidia has been weighing a plan to guarantee 250 billion dollars for an OpenAI data center project in Ohio. Separately it has considered financing OpenAI's purchase of 350 billion dollars worth of its own chips. In that Ohio project, OpenAI leases the facility for up to 20 years, Nvidia is the exclusive chip supplier for the first phase, and Nvidia also put 1.5 billion into the energy company that owns the site.
So Nvidia is, in various forms, the investor, the supplier, the guarantor, and a stakeholder in the landlord. That's not a customer relationship. That's a company building its own demand.
And the company at the center of it all, OpenAI, is losing money at spectacular scale. It is on track to lose around 14 billion dollars in 2026, nearly triple its losses from the year before. It projects 100 billion in revenue by 2029, which may well happen, but "may well happen in three years" is doing an enormous amount of load-bearing work in a structure this large.
The tell
If you want one detail that captures the whole problem, here it is.
In August 2026, one of these giant OpenAI data center arrangements with Nvidia was finalized. When the actual numbers landed, the deal came in roughly 145 billion dollars lower than what had been reported earlier.
That gap is the size of a large national economy. It vanished between the announcement and the paperwork.
That's what makes these numbers so hard to trust. A lot of what gets reported as an AI megadeal is a letter of intent, a framework, a multi-year commitment with escape hatches, or a headline figure that assumes every optional phase gets exercised. It generates a press cycle, a stock move, and a talking point about explosive demand. Then the real contract quietly turns out to be a fraction of it.
The concern experts keep raising is simple: when a supplier funds its own customers, the incentive stops being is this capacity actually needed and becomes can we book this deal. Those are very different questions, and only one of them is about reality.
Now the part hidden off the books
The circular deals are the visible half. The debt is the invisible half, and it's bigger.
J.P. Morgan puts hyperscaler capital spending at around 697 billion dollars in 2026, roughly triple what it was in 2024. That money is not all coming out of profits. A lot of it is borrowed, and increasingly it is borrowed in ways that don't show up where you'd look for them.
Private credit funds are now the main lenders here. Their loans to AI-related companies went from basically zero to over 200 billion dollars in a handful of years, and Morgan Stanley projects private credit will supply another 800 billion in data center financing over the next two years.
Then there's the accounting. Companies have used special purpose vehicles, essentially separate entities that hold the assets and the debt, to move more than 120 billion dollars of data center spending off their balance sheets in about eighteen months. The typical setup: a joint venture owns the data center, outside investors fund it, the tech giant takes a minority stake plus a long-term lease and sometimes a guarantee. The obligation is real. It just doesn't appear as debt on the tech company's books.
How much is out there? Roughly 662 billion dollars in hyperscaler data center lease commitments sit off balance sheet under standard accounting rules. That figure is 113 percent of those same companies' reported adjusted debt. In other words, there is more obligation hiding outside the balance sheet than sitting on it. Total lease exposure runs closer to 969 billion.
Independent estimates of Big Tech's total off-balance-sheet AI commitments land around 1.65 trillion dollars. Meta alone accounts for something like 420 billion of that, which is nearly triple its officially reported debt.
None of this is fraud. It's all disclosed somewhere, in footnotes, by companies with real profits and real products. But if you looked only at the headline balance sheets, you would badly underestimate how much of this boom is running on borrowed money.
We have seen this exact movie before
Here is where it stops being a vibe and starts being a documented historical pattern, because vendor financing has already blown up a technology boom once, and it happened in living memory.
In 1999 and 2000, the telecom equipment giants had a problem. Their customers were new companies laying fiber optic cable with no revenue and no cash flow, and banks had stopped lending to them. So the equipment makers lent the money themselves.
Lucent committed 8.1 billion dollars in vendor financing. Nortel extended 3.1 billion. Cisco promised 2.4 billion in customer loans. The programs peaked in the fall of 2000, precisely as the underlying customers were getting shakier. Lucent publicly described its 2 billion dollar financing arrangement with a company called Winstar as a long-term strategic relationship.
Winstar went bankrupt in April 2001, owing Lucent more than 800 million dollars.
Then the whole thing came apart, and the numbers are brutal. Look at what happened to bad loans as a share of total loans between the end of 2000 and the end of 2001:
- Motorola went from 6.7 percent to 57 percent
- Lucent went from 2.6 percent to 60 percent
- Nortel went from 25.5 percent to 80 percent
Within twelve months, most of the money these companies had lent their customers was uncollectable. The revenue they had booked was real on paper and fictional in practice.
Now here is the comparison that should stop you cold. Lucent's vendor financing commitments at the peak were about 8.1 billion against 33.6 billion in revenue, roughly 24 percent of annual sales. Nvidia's direct investments run around 110 billion against roughly 165 billion in trailing revenue, which is about 67 percent.
Relative to its own revenue, Nvidia's customer exposure is about 2.8 times larger than Lucent's was right before the telecom crash wiped Lucent out.
The honest counterargument
To be fair, there is a real defense here, and it deserves stating properly rather than as a strawman.
Demand for AI computing is genuinely, verifiably enormous. This isn't fiber optic cable being laid for internet traffic that hadn't arrived yet. Hundreds of millions of people use these tools daily. Enterprises are spending real money. The capacity being built is largely being used.
In that light, the deals look less like a shell game and more like ordinary supply chain financing. Asset managers defending the structure call it a virtuous circle rather than a circular one: if you're a chip maker and you know demand is exploding, locking in long-term buyers by helping finance them is a rational way to secure your order book. Manufacturers have done this forever.
That argument is not stupid. It might even be right.
But notice what it depends on. It depends entirely on demand continuing to grow fast enough, for long enough, to service an enormous stack of debt and lease obligations built on the assumption of that growth. If AI demand merely grows well instead of explosively, the loop still unwinds. The customers were funded on the assumption of a specific curve. Miss the curve and the loans go bad, exactly like Winstar.
The bull case isn't wrong. It's just extremely narrow.
Who actually holds the bag
Here's the part that matters if you don't own any of these stocks and think this is somebody else's problem.
The Magnificent Seven now make up more than a third of the entire S&P 500's value. At the absolute peak of the dot-com bubble, the leading tech companies were around 15 percent. Technology as a sector accounts for over half the index's footprint and more than 15 percent of the corporate bond market.
Which means it's in your retirement whether you chose it or not.
State and local public pension systems in the US held about 6.5 trillion dollars in assets, with roughly 42 percent in equities on average. Anyone in a standard index fund or global tracker is heavily exposed to the same handful of names, because that's what an index-weighted fund does. You didn't pick Nvidia. Your pension picked it for you, by definition.
Both the Bank of England and the European Central Bank have publicly flagged this. The Bank of England warned that a sudden deflation of the AI boom would hit the pension pots and portfolios of ordinary savers, noting that UK share valuations were near their most stretched levels since 2008 and US valuations were starting to resemble the run-up to the dot-com crash.
And there's a nasty second-order effect. When funds take losses and savers pull money out, the funds have to sell to meet withdrawals. They sell the liquid stuff first, which pushes prices down further, which triggers more withdrawals. The structure that spread the gains around is the same structure that transmits the losses.
The real problem is that you cannot see any of it
Step back from the numbers and the actual issue is not that this boom is debt-financed. Most booms are. The issue is opacity.
Nobody, including the participants, can currently draw a complete map of who owes what to whom in this system. Obligations sit inside special purpose vehicles. Revenue gets counted at multiple points along the same dollar. Commitments are announced at headline figures and quietly signed at fractions. Guarantees create exposures that show up nowhere obvious until they're triggered.
That's not a moral failing, it's a plumbing failure. Our financial infrastructure lets enormous, interlocking obligations exist without any shared, verifiable record of them. Every crisis in modern financial history has had this same fingerprint: the exposures were knowable in theory and invisible in practice until the moment they mattered.
This is the unglamorous problem that the serious end of the crypto world has been chipping at for years, and where projects like Ozak AI are building. Not as a trading story, but as infrastructure: when value and obligations move on rails that anyone can audit, circular exposure becomes something you can actually see rather than something you reconstruct from footnotes after the fact. There's a second point too. Compute that's spread across many independent operators isn't financed by five companies lending each other money in a loop, which means it doesn't share the loop's single point of failure.
That won't stop a bubble. Nothing stops bubbles, they're a feature of humans, not of technology. But being able to see the wiring is the difference between a correction and a surprise.
How to actually think about this
Separate the technology from the financing. AI is real, useful, and not going away. That is a completely different question from whether the current capital structure funding it is sound. Both things can be true: the tools work and the money is arranged badly.
Treat announced deal sizes as marketing. When you see a 500 billion dollar AI partnership headline, assume the real number is smaller, possibly much smaller. The 145 billion dollar gap in one deal this August is your reference point.
Know your own exposure. If you have a pension or index fund, you own a lot of this. Not a reason to panic, but a reason to know. "I don't invest in tech stocks" is almost certainly not true anymore.
Watch the debt, not the revenue. Revenue in a circular system can be manufactured. Lease commitments and loan obligations cannot. When you see off-balance-sheet obligations exceeding reported debt, that's the number that tells you the real story.
The uncomfortable truth is that we don't know yet. This could be a rational supply chain financing a genuine industrial revolution. It could be the telecom bubble with better products. Most likely it's some of both, and the sorting happens after the fact, when the ones holding uncollectable loans find out which category they were in.
The people building this are not fools. Lucent wasn't run by fools either. It was the largest telecom equipment maker in North America with 157,000 employees, and it lent money to customers who couldn't pay it back because everyone around them was doing the same thing and the growth story was intoxicating.
Money moving in a circle looks exactly like money moving forward. Right up until it stops.
