Circular AI Financing Relocates the Downside — It Doesn't Remove ItAI 循环融资把下行转嫁给厂商，而不是消除它

Lede

Circular financing does not remove the risk in the AI boom. It decides who eats the loss. In mid-October the IMF counted the money moving in circles — 14 deals since 2023, about $300 billion of equity and $650 billion of debt for roughly $1.4 trillion in promised future revenue — and a day later the credit market began pricing who actually pays if those promises fall short.

That is the whole argument in two sentences. The rest of this piece is the mechanism behind it, the strongest case that it is wrong, and the few observations that would change our mind.

On October 13, 2026, the IMF's Global Financial Stability Report listed an AI-driven market correction as one of three key risks to financial stability, alongside high bond yields and cyberattacks. The next day, the stress stopped being abstract. A batch of CoreWeave bonds worth $1.25 billion briefly traded at a yield above 15%. CoreWeave stock fell more than 15.5% in a session, Oracle 8.8%, Dell about 11.3%, AMD 5.73%. The common thread was not a single bad earnings report. It was a sudden doubt about a financing structure that ties chipmakers, cloud builders and AI labs into one loop.

The question

The concept here is simple and old: leverage amplifies both ways. It magnifies verified methods and unverified mistakes with the same indifference. The guiding question for AI's circular deals is therefore not "is this a bubble" but something more precise: when a chipmaker or hyperscaler finances the very customers who buy its products, is the downside being removed, or only moved somewhere else — and if it is moved, to whom?

That question matters because most people reading this own the answer whether they looked for it or not. Broad index funds are heavily weighted toward exactly these names, and their credit sits in bond funds and money-market instruments that feel safe. If the risk has been relocated rather than removed, the "safety" of a diversified index is partly an illusion about where the loss would land.

The default read

The comfortable interpretation runs like this. The mid-October selloff was volatility, not a verdict. Vendor equity stakes and backstops are a feature, not a flaw: a cash-rich chipmaker investing in its customers and guaranteeing part of their obligations makes the buildout safer, because it signals confidence and gives lenders someone solid to stand behind. On this read, circular financing de-risks the system.

There is real evidence for that view, and we will give it its full weight below. But the default read quietly assumes the conclusion it needs to prove — that a guarantee makes risk smaller. A guarantee does not make risk smaller. It makes risk someone's. The question is whose, and what happens to that someone if the demand underneath never shows up.

Evidence

We know the shape of the money. The IMF reviewed 14 circular AI transactions since 2023 and found about $300 billion in equity investments and $650 billion in debt exchanged for roughly $1.4 trillion in future revenue commitments. In plain terms: investors put in $300 billion of ownership, lenders put in $650 billion of loans, and in return the deals promise $1.4 trillion of revenue that has not yet arrived. The IMF's own description of the danger is worth quoting, because it is the mechanism stated by the body least prone to hype: stress at key companies receiving investment would lower revenue forecasts and investment valuations, weakening the value of the accumulated debt in those deals.

We know what it looked like when confidence slipped. On October 14, a CoreWeave bond batch of $1.25 billion briefly yielded over 15% — a rate that says lenders suddenly wanted to be paid like they were holding distressed debt, not an investment-grade loan to a fast-growing cloud company. CoreWeave had expanded from 3 data centers to 51 in four years and carried roughly $50 billion of obligations. A Wells Fargo analyst called its balance sheet "relatively bad" compared with other neoclouds. These exact bond and spread levels are media-reported and should be confirmed against FINRA's TRACE bond-price data before anyone treats them as precise; what is not in doubt is the direction. Credit moved first, and it moved hard.

We know the named example everyone points to. In September 2025, Nvidia agreed to invest up to $100 billion in OpenAI to help fund at least 10 gigawatts of data-center capacity — enough power for a major city — while OpenAI committed to deploy millions of Nvidia GPUs in those facilities. The framework was later scaled back; Jensen Huang said the original $100 billion was "probably not in the cards," with about $30 billion more likely as Nvidia's last investment before an OpenAI IPO. Whatever the final number, the structure is the point: the seller of the chips is helping finance the buyer of the chips.

How the system works

Follow one dollar around the loop. A chipmaker or hyperscaler takes an equity stake in a neocloud — a company whose business is renting out racks of GPUs. The neocloud now has cash, and it borrows more against those GPUs as collateral. It spends the combined pile on chips, which it buys from the same vendor that just invested in it. That spending lands on the vendor's income statement as revenue. Higher revenue lifts the vendor's valuation, which strengthens its balance sheet, which lets it fund the next equity stake in the next neocloud. The loop feeds itself.

Notice what the loop does to information. In a normal market, a chipmaker's revenue is a signal: customers paid real money because they had their own demand. The sales inside the loop are booked the same way and may be perfectly proper as accounting — the point is not that anyone is cooking the books. The point is economic: part of the demand behind those sales was funded by the vendor's own capital, so the revenue is a weaker signal of independent, outside demand than it looks. The structure amplifies demand it has not verified, and because leverage amplifies both ways, it will amplify the reversal just as efficiently if end-demand disappoints.

Now add the guarantees. Vendors do not only invest; they backstop. A residual-value guarantee, for instance, promises that if a customer defaults and the GPUs have to be re-leased or sold, the vendor covers the gap between a pre-set minimum value and what the hardware actually fetches. While demand is strong, that promise costs nothing and reassures lenders. The incentives all point one way: the vendor wants the buildout to happen because the buildout is its revenue, so it has every reason to keep the loop spinning and the guarantees flowing.

The cost of that arrangement is not removed. It is parked. It sits as a contingent liability, invisible on good days, and it routes the downside back to the vendor on bad ones. If AI-native end-demand cools, the vendor is hit twice: once as slower hardware sales, and again as guarantees get triggered on the customers it financed. The feedback that was virtuous on the way up runs backward on the way down. Dearer or frozen GPU-backed credit forces unsold capacity to be sold, which triggers the residual-value guarantees, which pulls the loss onto the vendor's own free cash flow — and makes the next round of financing more expensive, slowing the buildout that the revenue depended on.

This is why we infer that the credit market, not the stock market, is the place to watch. Equity holders own the upside and are slow to mark down a growth story. Lenders own a fixed claim and a hard maturity date; they reprice the instant they doubt they will be repaid. The October 14 bond move was the loop's early-warning system doing its job.

Who gains, who pays

On the way up, almost everyone gains, which is exactly why the structure is hard to resist. The chipmaker books revenue and a rising valuation. The neocloud gets capital and growth it could not have financed alone. The AI lab gets compute. Index investors get the gains concentrated in the largest names. The gains are real while the music plays.

On the way down, the parties separate sharply. Neocloud creditors pay first, because they reprice first — that is what the 15% yield was. The vendor pays next and pays most, because the guarantees it wrote come due and the demand it financed evaporates on the same quarter. The AI labs, often privately held, are insulated from public marking until a funding round or an IPO forces a number. And the index investor, who never saw a loop and only bought "the market," pays last and least knowingly, because the loss arrives as a drawdown in a fund they were told was diversified.

Put plainly: vendor financing doesn't remove the risk; it decides who eats the loss. A backstop is not a cushion for everyone. It is a decision about whose balance sheet absorbs the fall.

The strongest counter

Now the serious case that we are wrong, stated at full strength.

The vendor can carry it. By one careful analysis, Nvidia holds about $99.4 billion in cash and marketable securities against roughly $165 billion in total customer commitments — about 1.3 times its $127 billion of trailing-twelve-month free cash flow, and only about half of expected free cash flow two years out. The much-quoted $105 billion OpenAI figure is not a revenue floor and not a loan; it is a residual-value guarantee that triggers only on default, terminates automatically if OpenAI reaches a satisfactory credit rating, and does not even begin until the buildings are ready, expected around 2028. Nvidia's CFO put it bluntly: "We're not making loans." On this view the structure is vertical integration without an acquisition — a chipmaker securing demand for a product it is uniquely able to supply.

The demand can be real. CoreWeave reports a contracted backlog of roughly $104 billion as of June, led by Microsoft, Meta and OpenAI — among the most creditworthy buyers on earth. If most of that backlog is genuine end-demand from customers who are not themselves vendor-funded, then the circularity is a small ring around a large base of real orders, and the October credit move was a repricing of one weak balance sheet, not the first crack of a cascade.

This counter is strong enough that the honest position is conditional, not certain. If the backlog is durable non-vendor-funded demand and the vendor stays solvent, circular financing is just trade credit with a confident supplier, and the selloff was noise. One caution even here: a guarantee that triggers "only on default" is less protective than it sounds, because the moment a customer defaults is also the moment used-GPU resale values are weakest — the trigger and the loss tend to arrive together, not apart. Our claim is not that the loop must break. It is narrower and, we think, harder to dodge: the backstop relocates the risk onto the vendor, so the question that decides everything is how much of the demand is the loop financing itself — and that is precisely the number the disclosures do not cleanly break out.

There is a longer shadow behind the counter, too. In the late-1990s telecom boom, Lucent Technologies used vendor financing to help customers buy its equipment, committing on the order of $8.1 billion, with about $2.1 billion outstanding. It booked the revenue on the way up. When customers like WinStar went bankrupt, Lucent wrote off about $700 million on that account alone, took bad-debt provisions of roughly $2.2 billion in 2001 and $1.3 billion in 2002, and posted a $16.2 billion loss for fiscal 2001. Vendor financing did not remove the risk of weak end-demand. It delayed the recognition and then delivered the bill to the vendor. The AI names today have far stronger balance sheets than Lucent did, which is the counter's best point. The mechanism that turned Lucent's revenue into Lucent's write-down is the same one running now, which is ours.

What would change our mind

We hold this as a medium-confidence judgment, and we would revise it on specific, observable evidence, not on a change in mood.

We would back away if, over roughly the next three months, two things happen together. First, neocloud debt service is demonstrably covered by contracted, non-vendor-funded customer revenue — shown in issuer 8-K and 10-Q filings that break out how much of the backlog comes from customers who are not themselves being financed by the chip vendors. Second, CoreWeave and Oracle bond yields and CDS spreads normalize without new vendor backstops or fresh equity injections propping them up. If debt gets repaid by real outside customers and credit calms down on its own, the relocation thesis loses its force and the structure really is ordinary trade credit. The check-by date is January 31, 2027, and the scorekeepers are FINRA TRACE, the issuers' own filings, and the next IMF GFSR update.

We would also weaken the claim if a direct, primary breakdown showed that the circular portion — vendor equity funding the demand for vendor chips — is a small share, call it under roughly 5%, of total AI infrastructure spending, confirmed by the next IMF GFSR update or a comparable primary breakdown by mid-2027. The IMF's $300 billion / $650 billion / $1.4 trillion counts the deals; it does not by itself tell us what fraction of all end-demand is self-funded. That gap is the honest limit of what anyone can claim today.

What an individual can do

This is not advice to buy or sell anything, and nothing here is a forecast of price. It is a way to read your own exposure more honestly.

If you hold a broad index fund, a chipmaker, or any AI-infrastructure credit, do one small, checkable thing before you treat the October dip as a discount: separate contracted, non-vendor-funded demand from vendor-funded demand. In practice that means reading the next backlog disclosure not for its headline size but for its source — how much of it comes from customers who are not being financed, directly or indirectly, by the companies selling them the chips. Watch the credit market as your early indicator, because it will move before the equity story does. And size any conviction to what you can actually verify, not to the confidence of the loop, because the loop is designed to feel confident. Leverage amplifies both ways; the discipline is to add exposure only as fast as the demand is verified, not as fast as the financing is available.

The buildout may well be real, and the strongest AI names may carry their commitments without strain. But "the backstop de-risks it" is the one sentence to distrust. The backstop does not shrink the risk. It tells you whose problem it becomes. Knowing whose is most of the work.

This is research, not investment advice.