The superintelligence hasn't arrived yet, but the super bill has? AI infrastructure debt and bubble risks

Fewer people are discussing the Korean stock market, and it's not because the Korean market's wild swings have become so frequent that everyone has been desensitized.

Rather, recently the whole world has been falling, falling, falling, falling, falling—from US stocks to Korean stocks to A-shares, none have been spared...

A very fitting Halloween costume

In the last week of October, the market value of the "Nasdaq 100" evaporated by about $2 trillion, equivalent to wiping out the GDP of several countries in a single week.

In particular, last week's "Black Friday" saw all three major US stock indices close sharply lower, with tech stocks and the semiconductor sector leading the decline.

It could even be said they were the absolute main force being sold off by the market.

In this round of the crash, Google's parent company Alphabet plunged 12.64%, with about $500 billion in market value evaporating in a single day.

Even worse was Intel, which closed down 28.58% that day, with $150 billion in market value evaporating—its second-largest single-day drop since the company went public.

Our A-share semiconductor sector has also recently suffered a sharp correction, with consecutive steep declines on October 29 and 30.

Buffett and his Coca-Cola remain firm

For stock investors, although the broader market hasn't had the kind of violent roller-coaster ride Korean stocks had before, this kind of rapid plunge concentrated in the hottest sectors and even leading companies stings like a needle and bores in like an awl—and now that it has landed on your head, it really has become a mountain.

So it's not that people don't care about other countries' stock markets; it's that everyone is equally miserable now, with no time at all to care about other countries' markets—their own accounts are like clay Bodhisattvas fording a river.

But Buffett wasn't spared this time either

Wait, that's not right?

Including this time, why is it that the better the performance, the harder the fall?

Take Google and Intel, the hardest hit this time: both companies' results for the quarter actually looked quite good, but the market just wouldn't buy it.

First, Google: third-quarter revenue was $127.6 billion, up 25% year over year, and Google Cloud revenue doubled. Yet the company's free cash flow for the quarter was actually negative—it made so much money and still couldn't fill the hole it had dug: the AI infrastructure gamble.

Intel is a similar case: third-quarter revenue was $16.513 billion, up 21% year over year, and data center and AI business grew 78% year over year, its strongest growth in 15 years. But the CFO warned that the company still won't achieve positive free cash flow in 2027, and future capital investment will keep increasing.

Two brothers in misfortune, both taking a beating

Behind this tale of fire and ice is Silicon Valley entering an era of great repression after great money-throwing!

How to put it? Even if this money were really set on fire, that would definitely be much, much slower than the crazy way it's being used now.

Because right now, tens of billions is the entry requirement, hundreds of billions is normal, and trillions is the sea of stars.

Alphabet, Microsoft, Amazon, Meta, and Oracle are expected to spend $769.2 billion in capital expenditure in 2026, nearly double 2025, and it could reach $1 trillion in 2027. Add SpaceX, and according to S&P estimates, these leading companies' capital spending on data centers and other AI projects from 2025 to 2030 will exceed $7 trillion—truly astronomical figures.

Hundreds of billions for the "Stargate" data center project—built just like that; tens of billions for computing units—bought just like that. Everyone knows Silicon Valley's elite are flush with cash, but even a fortune of ten thousand gold can't withstand spending like this, and in the end even the landlord's family has no surplus grain...

Look at Amazon first: in the latest quarter it generated $42.4 billion in operating cash flow, but then spent $77.3 billion buying land, buying chips, and building data centers, ultimately ending up with negative free cash flow of $33.5 billion—four times worse than analysts' expected negative $7.7 billion!

How to fill the gap? Borrow! This quarter it added $48.1 billion in long-term debt, and long-term debt has soared from $65.6 billion at the start of the year to $173.1 billion.

Alphabet also had negative free cash flow, raising $12.4 billion by selling stock in the same quarter and borrowing another $29 billion.

Meta is even more extreme: it earned $32.8 billion in a single quarter and dared to spend $34.9 billion, driving free cash flow straight to negative $2.1 billion—the first time since it went public.

If you look at Microsoft, Alphabet, Amazon, Meta, and Oracle—the "royal five"—together, total operating cash flow in the first three quarters was $542.9 billion, up 29% year over year and a record high. But capital expenditure doubled to $556.9 billion.

Crazy bond issuance

Obviously, they're spending more than they earn—and this is the first time since data began in 2011.

This means the "royal five's" free cash flow has also turned from a surplus of $143.5 billion in the same period last year into a deficit of $14 billion.

Often the money they've earned hasn't even arrived yet, but the money in their accounts has already been spent in advance—one could say the "cash cows" have collectively contracted "milk-production anemia."

When money is insufficient, the first to be sacrificed is naturally shareholders.

The five companies' buybacks were cut by 83%, from $81.8 billion to $13.6 billion. Alphabet and Meta didn't buy back a single share this year, whereas in the same period last year the two together bought $66.5 billion. The money now used to reward shareholders accounts for only 9.3% of the cash earned, versus 27.4% last year—and that's not even counting what's being spent now.

In the past, tech giants could invest using the money they earned themselves, but now using shareholders' money as a sacrifice isn't enough; the next step is large-scale bond issuance to raise funds.

By this year, leading AI companies had issued about $244 billion in bonds, compared with just $108 billion for all of last year. JPMorgan expects tech-sector bond issuance to reach $540 billion for the full year, a record high.

It has pushed up global bond issuance overall.

Market attitudes are shifting too: spending used to be seen as "confidence in building," but now spending makes you a "wastrel squandering the family fortune."

So no matter how good the results are, or how vigorously construction is being pushed, none of it can reassure investors anymore.

This distrust spares even the star large-model companies.

Raise money, raise money, and raise™ more money!

Anthropic, which is preparing for an IPO, has had its credit rating revealed for the first time: just BB+, the highest tier within speculative grade, one notch below investment grade.

Put it this way: in the bond market, this is basically what many people would call a "junk bond."

It means capital markets still have clear doubts about Anthropic's long-term debt-servicing capacity, the stability of its cash flow, and the sustainability of its business model.

02

The Invisible Giant Shadow

What's more terrifying than discovering a pile of bubbles is that even more bubbles are hidden in the shadows.

When the light finally comes, the bubbles in the sunlight won't be colorful—they will burst tragically and in a chain reaction, and nothing will be left.

Last year, some media combed through the financial statement footnotes of Silicon Valley's Big Five and found something counterintuitive: their debt not listed on the balance sheet had risen to about $1.65 trillion, more than the $1.35 trillion on the books, and it had grown eightfold in four years.

Off-balance-sheet debt is staggering.

Follow-up media reports expanded the scope to nine companies, and the figure reached an astonishing $3 trillion!

That said, this doesn't mean they owe $3 trillion today.

These are contractual commitments to be fulfilled over the next 5 to 10 years, but that doesn't mean these Silicon Valley giants are saints.

Brown is on-balance-sheet, blue is off-balance-sheet.

The bulk comes from two parts: about $1.2 trillion in lease commitments that have not yet begun to be paid, and about $1.9 trillion in long-term purchase commitments, including chips, servers, and other data center infrastructure.

The problem is that these future payment obligations will not all appear directly on the balance sheet in the form of traditional debt; instead, they appear partly as lease obligations and purchase commitments in items such as contractual commitments and lease disclosures in the financial statement footnotes.

In other words, if you only stare at the liability figures on the balance sheet, you may underestimate the real funding pressure these tech giants will have to bear in the coming years.

On one hand, they hide debt in fancy ways so it doesn't show on the books; on the other, they raise financing in fancy ways, constantly making the debt bigger and bigger.

Lease, don't buy.

Data centers cost tens of billions of dollars apiece, and the giants don't build them all themselves—they lease heavily on a long-term basis.

Under current accounting standards, leases are not counted as liabilities before they begin and are only disclosed in the footnotes. Oracle is the most extreme example.

It supplies computing power for OpenAI's "Stargate," relying on leased data centers. According to its latest 10-Q disclosure, its lease commitments that have not yet begun have reached $288 billion, compared with just $7.1 billion in November 2022—a roughly 40-fold increase in less than four years!

Bring in partners.

For its Hyperion data center in Louisiana, Meta brought in Blue Owl to form a joint venture, with the other party holding 80% and Meta 20%, with development costs of about $27 billion. Once Meta leases back the completed campus, that debt won't count as Meta's liability.

Meta is currently also considering using the same trick to plan a new data center in Texas on the scale of $13 billion. But according to reports, Meta's off-balance-sheet debt is now about $420 billion, nearly three times its on-book debt.

Beyond the giants, new cloud providers lack the cash flow of the big companies and can only pledge the GPUs themselves as collateral.

CoreWeave once borrowed about $2.3 billion using H100s as collateral, with a coupon of about 14%. GPU-collateralized financing by such AI cloud startups has cumulatively exceeded $10 billion.

Even more subtle is the "internal circulation" among the giants.

The Wall Street Journal once reported that Nvidia discussed a financing guarantee of about $250 billion with OpenAI.

This is mainly to help OpenAI lease the 10-gigawatt mega data center that SoftBank is developing in Ohio, with total project investment expected to exceed $500 billion. Other reports say Nvidia is also discussing providing up to $350 billion in financing for OpenAI to buy chips.

The three who play the game best

If that makes your stomach burn a little, take a look at the intuitive path of the money:

Nvidia guarantees → SoftBank builds the data center → OpenAI leases computing power → OpenAI uses the money propped up by the guarantee to go back and buy Nvidia chips → Nvidia recognizes revenue, then goes off to guarantee more projects.

Good grief—the one selling the shovels is guaranteeing the people buying the shovels.

Bond issuance is even more wildly imaginative.

Alphabet's sterling century bond (maturing in 2126) is the first by a tech company since 1997, and Amazon's dollar bonds run as long as 2076, pushing repayment pressure decades into the future—truly a matter of "trusting the wisdom of future generations."

JPMorgan estimates that total AI infrastructure investment will reach $5.5 trillion by 2030, of which about $4.1 trillion will need to be financed through debt.

Columbia Business School professor Van Nieuwerburgh has warned about this: because this kind of debt is increasingly flowing through pension funds and private credit vehicles, few people realize that this risk has already seeped into their own retirement savings and insurance savings.

On overseas forums, ordinary investors' worries about this can only be described as growing more severe.

Related discussions on the AI Debt entry

The news reports every day on all kinds of massive AI infrastructure buildouts, and in the US, Trump has even demanded that AI be renamed Super Intelligence—but where the money comes from and how it is used is a murky account???

Especially the most intuitive feeling: so much money has been poured in, so why hasn't much of a stir been heard?

Some might say, what if this is similar to the telecom bubble in the early 2000s? Infrastructure gets the money first, and only afterward can results bloom.

In 2000, the capital expenditure of US-listed telecom operators was about $121 billion, roughly 1.2% of GDP that year.

By Goldman Sachs' measure, US AI investment as a share of GDP is expected to rise from 1.8% in 2026 to 2.5% in 2027, and continue further in the future.

So this time the money being poured in is truly too aggressive—not on the same order of magnitude at all.

The Enron Scandal is one of the most serious corporate financial fraud incidents in US history. It not only left thousands of employees unemployed and caused investors to lose tens of billions of dollars, but also dealt a heavy blow to employees' retirement savings.

As the most well-known financial fraud case involving a listed company, its complex financing, off-balance-sheet commitments, heavy reliance on future earnings, and the difficulty of identifying true leverage all have similarities with the current AI Debt......

This Reddit user, whose post received more than 9,000 upvotes, described it even more as a combination of "Enron-style financial fraud + 2008 financial crisis financing model."

Enron-style risk—complex structures make the true financial condition opaque.

2008 financial crisis-style risk—whether credit expansion has exceeded final demand.

After the Enron incident, the 2008 financial crisis still occurred

In the face of a bubble, everyone seems to want to be the lucky one who "rides the bubble."

"Ride The Bubble"—rational professional investors (such as hedge funds and institutional investors), even when they clearly know that the price of an asset is already severely bubbled and far exceeds its intrinsic value, will not choose to short or sell immediately. Instead, they choose to keep buying or holding the asset, attempting to cash out with a profit at the last moment before the bubble bursts.

The pessimistic camp: if AI demand cannot be delivered, data centers, GPUs, and long-term lease commitments will form huge sunk costs, which in turn will trigger contagion of credit risk.

Citadel founder Ken Griffin said this round of AI enthusiasm has "almost certainly" formed a capital bubble.

Just let him fight the guy mentioned earlier

BlackRock CEO Larry Fink warned that if data center construction is delayed, it could push up computing costs and threaten economic growth.

Invest too little, and if demand really explodes, customers and market share will be ceded to rivals; invest too much, and the worst case is depreciation, impairment, and pressure on stock prices, which is not fatal to cash-rich giants. Because the costs of the two kinds of mistakes are asymmetric, the only option is to keep raising the stakes and keep burning money.

The moment of rupture is unknowable to anyone

The strong rely on cash (inheritance) to tough it out; the weak rely on leverage (fate) to gamble recklessly.

This article comes from the WeChat public account "Coolplay Lab"; author: Coolplay Lab; published by 36Kr with authorization.