Deep Dive The Guarantor Falls, the Guaranteed Rise: Circular Financing Is Now Priced in the Bond Market · Jul 29, 2026
First, the background. "Circular financing" is the name for a funding structure that has become pervasive in the AI industry: a chipmaker or hyperscaler invests in its own customers, or lends its credit to guarantee their leases and debts, and those customers turn around and buy chips or computing power from it. Money makes a loop among insiders, the business gets bigger, and outsiders find it ever harder to tell whether the demand is real. On July 27, this storyline produced a landmark event: The Wall Street Journal reported that Nvidia is in talks to provide up to $250 billion in financing guarantees for OpenAI's planned 10-gigawatt data-center campus in Ohio. Nvidia closed the day down 4.99%, roughly $250 billion of market value gone, and ceded the title of world's most valuable company to Apple. Gary Marcus, a scholar and longtime AI bear, wrote the same day that the very kind of announcement — a supplier backstopping its customer — that sent Oracle up 43% intraday nine months ago is now sending Nvidia down. His conclusion: the market has flipped how it prices these announcements; announcing a mega-deal no longer lifts the stock, it reads as a confession of stress. This essay takes that judgment and checks it, announcement by announcement, against the stock-price reactions to every comparable deal since September 2025. Three conclusions came out of the audit:
>
- The "flip" claim fails as stated, because it compares the wrong parties. Oracle, which soared nine months ago, was the company receiving a giant order; Nvidia, which fell last week, is the company extending the guarantee — opposite positions in the transaction. And on July 27 itself, on the very same news, while guarantor Nvidia fell, the compute providers CoreWeave and Nebius rose in premarket trading — the two companies that would get access to cheap money precisely because of this guarantee. The dividing line between winners and losers is not time; it is which side of the deal you stand on. Going by the quotes we could obtain, the market treats a backstop as a subsidy flowing from the guarantor to the guaranteed, and prices each side according to its position.
- Align the positions, though, and the guarantor side's announcement premium really did evaporate over nine months. In September 2025, Nvidia announced up to $100 billion of investment in OpenAI: the stock rose 4% to a record high that day, gave it back the next, and the "investor of last resort" skepticism — nobody else will fund OpenAI, so only Nvidia can — was in print within the week. By November, when it and Microsoft invested in Anthropic, the market had gone nearly numb to the news; last week, the same kind of announcement bought a decline. But the books for that down day don't close cleanly: AMD fell about 5% the same day, and AMD has nothing to do with this backstop; semiconductors were broadly weak. On stock prices alone, "the market now punishes backstops" doesn't hold up.
- The genuinely clean signal is not in stock prices; it is in the bond market. Nvidia's five-year credit default swap — the insurance you buy on a company's debt — jumped to a record 82 basis points that day: that is the price of Nvidia's own credit, and AMD's decline cannot explain it. On the Oracle side, one retrospective analysis notes that back on September 10 last year, the day the stock leapt 36%, bond markets were already starting to price rising default risk. If that description holds, the stock market took roughly nine months to arrive where the credit market stood on day one. The Bank for International Settlements listed a circular-financing unwind among the top three risks to global financial stability at the end of June. To monitor this structure's health, the dial to watch now is credit spreads, not the stock market's same-day reaction to announcements.
Start by laying out what happened on July 27. Before Monday's open, The Wall Street Journal reported that Nvidia is negotiating to guarantee up to $250 billion of debt for OpenAI's 10-gigawatt data-center campus in Pike County, Ohio, covering leases and part of the construction cost, but not the GPUs themselves; a separate discussion covers a $350 billion financing package for chip purchases — two distinct transactions (SiliconANGLE). The campus's developer is SB Energy, a SoftBank company, with the first roughly 800 megawatts due for delivery in 2028. Nvidia declined to comment on the report; OpenAI did not respond to requests for comment. As of this essay's July 29 press time, the deal is neither confirmed nor denied; the terms could change, and it could fall apart (Datacenter Dynamics).
The scale deserves a moment. Google's guarantees on other companies' data-center rents total roughly $44 billion; Nvidia's rumored backstop would be nearly six times that (The Next Web). The problem it solves is equally plain: OpenAI has no investment-grade credit rating and cannot borrow this cheaply on its own. With Nvidia's guarantee, lenders underwrite Nvidia's credit, not OpenAI's income statement.
The market's response: Nvidia closed at $196.51, down 4.99%, its biggest one-day drop since February this year, with about $250 billion of market value erased in a session (Benzinga). The broad market was essentially flat that day: the S&P 500 rose 0.02%, the Dow gained 0.5%, and the Nasdaq slipped only 0.2% — with Nvidia itself supplying most of that decline. Apple rose 1% and, at $4.95 trillion of market value against Nvidia's $4.77 trillion, ended the day as the world's most valuable company (GuruFocus). The financial press mostly compressed the day into one line: the company building the compute got punished; Apple, which builds nothing and rents instead, got rewarded.
Gary Marcus's reading traveled farthest. He quoted an investor's email calling the deal one that "seems incredibly desperate by both parties. Two drunks leaning on each other to stand up." His frame: on September 10, 2025, Oracle's $300 billion mega-order sent the stock up as much as 43% intraday, and now the same structure's news is sending Nvidia down — the market has learned to read these announcements as bad news (Marcus on AI). It is an easy contrast to pass along: nine months, one structure, two numbers with opposite signs. What this essay does next is check whether it survives the tape.
Begin at the nine-months-ago end, because Marcus's version of it contains factual errors. After the close on September 9, 2025, Oracle reported earnings showing its backlog — contracted business not yet delivered — had exploded 359% to $455 billion; Oracle itself never named the customer. The next day, September 10, The Wall Street Journal identified the bulk of that backlog as OpenAI: a five-year, $300 billion contract (Reuters, via Yahoo Finance). Oracle rose as much as 43.2% intraday and closed up about 36% at $328.33, its best single day since 1992, adding $244 billion of market value (CNBC). The broad market was nearly flat; the move was entirely a single-stock event. Marcus's claim that the stock "stood at $307 per share at the close of trading on September 10, and briefly closed as high as $328 ten days later" does not match the record: $328.33 was the September 10 close itself — the all-time closing high — and the stock gave back about a tenth of it over the following two sessions.
Set that day beside July 27 and the problem is immediate: Oracle was the party receiving the order; Nvidia is the party extending the guarantee. A company handed a $300 billion order can rally, and a company promising to shoulder $250 billion of contingent liabilities can fall, in any era — and neither tells you the market has "seen through" circular financing. That is the original comparison's most substantive defect, and the test for it is equally direct: collect every comparable announcement since September 2025, compare guarantee providers with providers and order recipients with recipients, and see whether the direction of the reaction at the same position changed over time.
This essay assembled the major announcements since September 2025 in both categories — suppliers putting up money, credit, or backstops for customers, and customers landing giant orders — along with each one's same-day stock reaction. The samples differ in deal type and size (direct investments, credit guarantees, joint ventures), so the comparison here is about the direction of the reaction at a given position, not its magnitude. The result: with positions aligned, what emerges is not one flip but two separately evolving lines.
The guarantor line: the announcement premium evaporated within nine months, and the skepticism was in the room in week one. On September 22, 2025, Nvidia announced plans to invest up to $100 billion in OpenAI. The stock rose about 4% to a record high that day — and fell 2.8% the next, with CNBC's market coverage already describing Nvidia as possibly "OpenAI's only option to get the capital it needs right now, an investor of last resort" (CNBC, market report). In other words, "backstop equals desperation" is not a 2026 invention; it was in print the day after the first mega-deal, but it just couldn't yet outweigh the optimism. By November 18, when Microsoft and Nvidia announced a combined investment of up to $15 billion in Anthropic, the two stocks barely moved — one up less than 1%, the other slightly down (CNBC). The guarantor side's announcement premium had, in effect, reached zero by late autumn 2025; July 27 was simply the step from zero to negative.
The recipient line: the announcement-layer reward ran all the way into last month; the break came at the earnings layer. On November 3, 2025, Amazon landed a $38 billion, seven-year compute contract from OpenAI and rose 4% to a record close (CNBC). As recently as July 21, 2026, news of Nvidia taking stakes in compute providers such as Nebius sent Nebius up 16%, CoreWeave up 8%, and Oracle up 5% (as relayed by 24/7 Wall St.). The break didn't come from an announcement at all — it came from a filing. On June 10, 2026, Oracle delivered a fiscal fourth quarter that beat expectations across the board — revenue up 21%, cloud-infrastructure revenue up 93%, backlog reaching $638 billion — and the stock fell 10% after hours anyway (CNBC). The market wasn't looking at the orders; it was looking at the books: capital spending up 162% to $55.7 billion, free cash flow at negative $23.7 billion, and more debt still to be raised. The same company, delivering the same kind of good news, rose 36% nine months earlier and fell 10% this time. Lay these announcements out by position and date, one cell per event, and this is the cleanest positive-to-negative cell in the table — and it sits at the earnings layer, not the announcement layer.
Two more cells cut hardest against "the market now reads backstops as bad news." On February 2, 2026, media reported that Nvidia's $100 billion investment talks with OpenAI had stalled — five months without a signature — and Nvidia fell about 3% (per CNBC). On May 7, The Information reported that financing for Broadcom's $18 billion custom-chip project with OpenAI had hit a snag, because Broadcom wanted Microsoft to commit to roughly 40% of the chips before putting up the money — and Broadcom fell 4% (Investing.com). If the market truly priced backstops as a negative, a deal collapsing should trade as good news; both times it traded as bad news. The market punishes backstops falling through, not backstops. Even in 2026, the equity market still prices these transactions as necessary conditions for the business to keep turning — nowhere near a knee-jerk discount the instant a backstop appears.
Which brings us back to July 27's most direct contrast: same day, same news — guarantor Nvidia closed down 4.99%, while the compute providers CoreWeave, Nebius, and IREN rose in premarket trading, up 4%, up 6%, and higher respectively — the three companies that would get cheap funding precisely because of this guarantee (CoinCentral). Those three numbers are premarket readings; closing figures were not obtainable, and premarket quotes trade thinly — this is the weakest cell in the whole essay. But its direction matches the recipients' closing rally on the Nvidia-stake news six days earlier. Bloomberg Intelligence's read (which we have only secondhand; we could not obtain the original) was that the structure shows Nvidia is prepared to stand behind larger infrastructure projects, and that "demand concerns may be overblown." On the evidence available, one backstop got two prices pointing in opposite directions: a liability for the guarantor, a subsidy for the guaranteed. The pricing function has not flipped to "discount whenever you see a backstop"; it is doing what it always did — allocating the gains and losses of the transfer according to position.
Even looking only at the guarantor itself, too much else is mixed into July 27's decline. Four pieces of evidence dilute the equation "stock drop equals the market's pricing of the backstop."
First, AMD fell about 5% the same day, nearly identical in size, and AMD has nothing to do with this backstop; Teradyne fell 4%, Intel more than 3%, and the semiconductor ETF over 2%. A sector-wide pullback was running that day, and the Nasdaq had already closed lower the previous Friday on AI-spending worries. Second, Nvidia traded 153 million shares — almost exactly its 20-day average, with no volume spike: when a market is digesting a fact that changes a valuation, it does not usually look like this. Third, the weekend actually brought two big stories: besides the backstop report, Nvidia had officially announced its more-than-$500-billion partnership with Korea's SK Group after Friday's close. Bloomberg attributed Monday's decline to "$750 billion of deals" reigniting circular-financing fears — the backstop rumor and the SK deal counted together (Bloomberg); there is no basis for booking the whole drop to the backstop alone. Fourth, the next day Nvidia neither rebounded nor kept selling off — it edged up 0.25% — while the semiconductor rout's theme had already rotated to memory, with Micron down 9%.
So the honest version reads: Nvidia fell 5% that day, and how much of that was the pricing of the backstop cannot be separated using the public data available. Separating it requires an event study that strips out market and sector factors, and to date nobody has done one. This essay searched the academic literature: formal event studies on AI-compute or supplier-financing announcements — none exist. The closest peer-reviewed work is a 2025 paper by Qian, Peng, and Li in the journal Production and Operations Management, which measures suppliers' stock reactions when firms announce generative-AI initiatives — and the direction there is still positive.
But one number the sector pullback cannot explain: Nvidia's own five-year credit default swap jumped to 82 basis points that day — the highest since the contract began trading actively in November 2025, and its largest single-day intraday rise (Bloomberg, via TradingKey). A credit default swap is insurance bought on a company's debt; a jump in the premium means the bond market judged the company's default risk to have risen. It responds only to Nvidia's own credit, untouched by whether AMD fell or which sector rotated. At the start of the month the number sat near 40 basis points; by month's end, 82. To be precise about the accounting: this jump cannot be booked to the backstop alone either — the debt market was pricing the entire package of new commitments Nvidia laid out that weekend, the SK Group's $500 billion included. But however you decompose it, Nvidia's own credit got repriced, and that fact does not move. Manish Kabra, head of U.S. equity strategy at Société Générale, gave the phenomenon its one-liner: "For hyperscale computing companies, it's CDS, not EPS, that matters now." Meaning: the scale on which these companies are graded has switched from earnings per share to credit spreads (Investing.com).
That is what was genuinely new on July 27. The stock's 4.99% is a muddled ledger; what left circular financing's first measurable mark on a guarantor's own credit was those 82 basis points.
Pull the lens back to Oracle and the same layering shows even more clearly. A retrospective analysis from July 2026 writes that on September 10, 2025 — the day the stock leapt — "bond markets immediately began pricing in an increased risk of default for Oracle's bonds" (The Motley Fool). That is a single-source, after-the-fact description; this essay could not find first-hand bond-spread data for that day, so file it at that grade. The path afterward, though, can be audited hard. Oracle borrowed furiously to build against those orders; in June 2026 it beat expectations across the board and fell 10% anyway; the stock lost 35% over the month, and the last week of June was its worst since the dot-com bust of 2001. On July 9, S&P cut its credit rating to BBB-, one notch above junk, citing climbing debt and concentration in OpenAI. On July 28 the stock closed at $119.96, down 63.5% from its peak. Oracle itself added new risk language to its annual report: some customers are highly leveraged, and the company may have overestimated customer demand.
If that retrospective description holds, the timeline reads like this: the bond market asked the right question on day one; the stock market took nine months to price it in. To state the calibration precisely: equities were not silent — the stock gave back about a tenth within two sessions of the rally, and drifted down about 25% over the first half of 2026. But the repricing that actually followed the credit market's day-one question — can the books carry these orders — came only after the June 10 earnings report, a 35% decline in a single month. From September 2025 to June 2026 is nine months. And Nvidia's CDS record on July 27 is the same script opening on the guarantor's side: the risk did not disappear; it moved off OpenAI's balance sheet, through the guarantee, into Nvidia's credit curve — and there it got priced.
This layered reading also resolves the most important contradiction in the debate this essay opened with. The strongest warning this newsletter has on file since it began tracking this trend comes from the investor Bill Gurley: this structure pollutes demand signals, so outsiders reading the filings cannot tell real demand from insiders propping each other up — which means any glut gets discovered later, and more suddenly. Marcus's flip claim, if true, would amount to declaring that the masking has already failed: the market can price these deals the moment they are announced. Our audit's answer: both are right, on different layers. At the announcement-day stock layer, no clean price discovery is happening — the moves are set by position and sector noise, and the masking thesis is never even tested there. At the filings-and-credit layer, the masking Gurley described is real: how much of Oracle's backlog will ever become revenue, whether OpenAI can pay — the market waited nine months before it began voting on those questions with its feet, and the first to speak up was not shareholders but creditors. Price discovery never moved forward to announcement day. It moved to a different market — the credit market. When this newsletter adjudicated the circular-financing split two weeks ago, the judgment it left on file was that credit spreads would become the leading thermometer, ahead of stock prices. This month supplied hard readings at both ends: Oracle's downgrade, and Nvidia's CDS record.
Open ?
Also on the record: analyst readings of this same structure are, at this moment, publicly split. On the bull side: the industry research firm SemiAnalysis spent a long essay in early July casting Nvidia's backstop program as the "Central Bank of AI," stating flatly that "Without the backstop, there would be no cluster," and projecting an AI debt market with over $7 trillion outstanding by 2029 (SemiAnalysis); the asset manager Janus Henderson calls these mega-deals a "virtuous circle" (Benzinga); and the investment research firm Morningstar's headline on July 27 itself was that circular-deal fears were rising — but Nvidia was undervalued. Not one of Goldman Sachs, Morgan Stanley, or JPMorgan has cut its price target over the backstop. On the bear side: the veteran short-seller Jim Chanos likens Nvidia's practice directly to the customer financing that telecom-equipment maker Lucent ran in the late 1990s — a case that ended in bad debt and collapse; and the Bank for International Settlements, the central bank of the world's central banks, in its June 28 annual report listed a circular-financing unwind, alongside an AI capital-spending retreat, among the top three risks to global financial stability, flagging specifically that these arrangements' terms are "often poorly disclosed, creating the risk of the same underlying asset being pledged multiple times" (BIS annual report). Nvidia CEO Jensen Huang's position has not moved: he calls the circular-financing framing "ridiculous," arguing the company's investments are small relative to its partners' overall fundraising (Bloomberg). When one structure is being read simultaneously as credit enablement and as the eve of a Ponzi scheme, and central-bank institutions have started writing it into financial-stability reports, the split is itself the signal: this structure's pricing is nowhere near converged, and it is the credit market that will reveal first who was right.
The stock market's reaction to a circular-financing announcement is not a gauge of whether the structure is healthy: on the same backstop news, the guarantor fell and the guaranteed rose — the stock layer allocates the subsidy's gains and losses by position, and it never flipped. The gauge that matters is in the credit layer. Oracle's path demonstrated the sequence: the bond market prices first, the ratings follow, and the stock market spends nine months catching up; Nvidia's CDS setting a record on July 27 is the same sequence opening on the guarantor's side. For companies still planning to run the announce-a-mega-deal-and-lift-the-stock playbook, the premise has changed: on the guarantor side, the announcement premium has gone from zero to negative, and every new guarantee now gets booked directly onto your own credit curve.
>
What would prove this wrong (observation windows are this newsletter's own settings except where noted, with verdict dates falling when they close): One — if the $250 billion backstop formally signs and Nvidia's stock rises that day while its CDS falls, the "backstops weigh on the guarantor's credit" reading gets marked down: it would mean what the market wanted was simply certainty. If, instead, signing day still trades as guarantor-pays, guaranteed-gains, this essay's position-allocation reading gains. Two — if Nvidia's five-year CDS falls back near its start-of-month level of 40 basis points by the end of 2026, the credit-layer alarm was a false one and "the pricing moved to the bond market" gets downgraded; if it holds above 80 or keeps setting records, it gains. Three — if the recipients' announcement premium disappears at the next billion-dollar order — a recipient announcing a mega-deal and falling that day, with no sector co-decline to blame — that, and only that, is real evidence for Marcus's full flip, and this essay's position-allocation reading yields to it. Four — if someone completes a formal event study stripping out market and sector factors and it overturns the direction of "the guarantor side's announcement premium has decayed," this essay's aligned comparison defers to it: our table uses unadjusted raw returns, and that is its methodological ceiling.
Companies queuing to announce the same kind of deal — and their CFOs. For the past two years, announcing an astronomical compute contract was itself a capital-markets operation: the announcement lifted the stock, and the higher stock lowered the cost of the next financing round. That playbook now splits by side. If you are the recipient, the announcement-layer reward is still there — but Oracle demonstrated that its shelf life ends at your next earnings report: the speed at which orders convert to revenue, and your free cash flow, will set your stock price nine months out. If you are the guarantor, the announcement premium has already gone from zero to negative, and every guarantee will show up directly in your credit spread; disclosure timing and deal structure now need to be checked against your bond desk, not your investor-relations department.
Investors. Keep three things apart. First, do not use the same-day stock reaction to announcements to judge circular financing's health: that number mixes position, sector, and sentiment, and this essay has shown how dirty it is; if you want a gauge, watch the guarantor's CDS and the guaranteed side's bond spreads. Second, the cell worth memorizing is "the market punishes backstops falling through": in the market's eyes, these deals are still necessary conditions for the business to keep going. Bulls lean on that as support — but "the whole business stops turning without one party" is systemic risk said politely. Third, sell-side price targets have not moved at a single house; if the credit-layer alarm persists and equity analysts begin marking down, that lag is your window.
Anyone whose product roadmap rides on compute expansion. Your roadmap is a bet on the cost and priority of downstream compute access, and both are now written first in the credit layer: if the guarantor's CDS and the guaranteed side's bond spreads keep widening, the financing cost of compute supply is being repriced. Watch those two numbers as leading indicators for your roadmap's assumptions — you will see it earlier than the capacity announcements.
Lenders and risk officers. The two things the BIS flagged — thin disclosure and re-pledged assets — are precisely the wall this essay hit while auditing: backstop terms sit scattered across rumors, 10-Q footnotes, and declined comments; even Broadcom's $29 billion lease backstop first surfaced in the subsequent-events footnote of a quarterly filing. Writing "counterparty must disclose a full schedule of contingent liabilities" into your terms beats waiting for the 8-K by nine months.
Policymakers. When a supplier's credit becomes the financing foundation for its entire downstream, a single company's CDS stops being mere company news: Nvidia's credit curve is becoming the systemic interest rate of the AI buildout. The handle for financial-stability monitoring belongs on the guarantee network's total exposure and its cross-holding structures — not on waiting for the stock to break.
Written from the same research and judgments as the Traditional Chinese edition; every claim links to a primary document.