On July 27, a data point flashed red on Wall Street traders' screens: NVIDIA's five-year credit default swap (CDS) spread surged 14 basis points in a single day to 82 basis points. This marked the largest single-day increase for the contract since it began trading in November 2025.
On the same day, NVIDIA shares plunged 5%, erasing roughly $250 billion in market value. Apple seized the opportunity to reclaim the title of the world's most valuable company, reaching a $4.93 trillion market capitalization.
The trigger for this sell-off was ironically what NVIDIA itself announced as "good news": a deal with South Korea's SK Group for over $500 billion in AI infrastructure collaboration, and ongoing discussions to provide OpenAI with up to $250 billion in financing guarantees to help it lease a 10-gigawatt super data center being developed by SoftBank in Ohio. The total value of these two initiatives exceeds $750 billion, roughly 3.5 times NVIDIA's projected fiscal 2026 revenue.
The market isn't rejecting AI. The market is asking a more fundamental question: why would a company that sells chips use its own credit to guarantee a customer's purchase of those same chips?
The true sharpness of this sell-off isn't about whether AI has demand. It does. Hyperscale cloud service providers have a massive $1.45 trillion in contractual commitments, and the supply-demand imbalance for computing power remains unchanged.
The problem lies elsewhere: when NVIDIA simultaneously acts as a chip supplier, equity investor, and debt guarantor, every dollar of AI investment circulates within the same closed loop. The credit market has asked the most basic question on everyone's mind: if OpenAI fails to generate products that can repay its debts, who will step in? Will it be the guarantor NVIDIA, SoftBank which lent to OpenAI, or the banks that securitized this debt?
So, this isn't panic over an AI bubble. This is a repricing of the financing structure. When a Societe Generale analyst stated, "Now we need to look at CDS, not EPS," the translation is clear: the AI narrative is shifting from revenue growth to credit risk. From this moment on, the protagonist is no longer computing power, but the ledger.
How the "Circular Financing" Loop Operates
Let's break down this closed loop.
First, NVIDIA provides OpenAI with a $250 billion financing guarantee. OpenAI currently lacks an investment-grade credit rating; it is still burning through cash. With NVIDIA as the backstop, creditors are willing to lend at a lower cost for this data center project.
Second, the project is being developed by SoftBank's SB Energy in Piketon, Ohio, with a planned capacity of 10 gigawatts—enough to power 8 million US homes. Total investment is estimated to exceed $500 billion. Based on industry estimates for GPU costs in hyperscale AI data centers, the GPU procurement alone could reach hundreds of billions of dollars.
Third, NVIDIA is also separately negotiating with OpenAI for up to $350 billion in chip procurement financing. Combined with the $250 billion guarantee, NVIDIA's total exposure to this single client could reach $600 billion—compared to NVIDIA's own annual revenue of $216 billion.
In a full circle, the flow of money is: NVIDIA guarantees → SoftBank builds the data center → OpenAI leases computing power → OpenAI uses the money backed by NVIDIA's guarantee to buy NVIDIA's chips → NVIDIA books revenue and orders → then guarantees even more projects.
This is the closed loop the market calls "Circular Financing."
Michael Burry, famous for shorting subprime mortgages, posted a single sentence on X: "Round and round and round. NVIDIA to guarantee $200 billion that ChatGPT will spend on NVIDIA chips."
Famous short-seller Jim Chanos was more direct: "We've reached the stage of this cycle where NVIDIA must provide financing guarantees for two-thirds of the cost of chips it sells to data centers?! Lol, okay."
Aleksandar Tomic, Associate Dean at Boston College, compared this to the 1999 internet bubble era, where companies purchased each other's products to create an illusion of booming demand. "OpenAI is essentially using NVIDIA's money to buy NVIDIA's chips, rather than using revenue generated from customers to support its expansion."
Jensen Huang's response was blunt. In January, when discussing NVIDIA's investment in CoreWeave, he stated, "This is just a small fraction of the total capital they will eventually need to raise. Calling this circular is ridiculous." He emphasized that investments in companies like OpenAI and Anthropic would both drive industry growth and yield investment returns.
The problem is that the market is already repricing this "ridiculousness."
The Credit Market Flashes a Full Red Alert
The surge in NVIDIA's CDS is not an isolated event. It's the loudest alarm in a series of warnings.
The 14 basis point single-day jump in NVIDIA's CDS was not only the largest in the contract's history, but also crucially, during the same period, TSMC's CDS edged up only 0.02%, and ASML's even narrowed slightly. NVIDIA is a "lonely island" in this credit tremor.
However, this doesn't mean other companies are safe. Zooming out:
Oracle's CDS has risen to 203 basis points—not just a rise, but a surge. It was at 144 basis points at the beginning of the year and hit an all-time intraday high on July 21. The trigger was S&P downgrading its credit rating from BBB+ to BBB-, just one step away from junk status. The core reason: AI capital expenditure far exceeded market expectations, and free cash flow remains negative.
Even excluding Oracle's extreme case, the average CDS for Meta (65 bps), Amazon (58 bps), Alphabet (48 bps), and Microsoft (42 bps) is approximately 49 basis points. This is the highest level since 2018 and has doubled since the start of 2025.
This is not a credit problem for a single company. This is a wholesale repricing of credit across the entire industry.
The driving force behind this is the scale of bond issuance. Since 2026, companies including Amazon, Google, NVIDIA, Meta, Oracle, and SpaceX have issued a combined $182 billion in investment-grade bonds—a 1300% increase year-over-year, accounting for about 15% of all US corporate bond issuance.
This is not a normal pace of financing. This is the debt acceleration created by compressing trillion-dollar capital expenditure plans into a 12-month period. Morgan Stanley data shows that the overall leverage ratio for hyperscale cloud service providers has soared from 0.9x to 1.8x in just two quarters.
Moody's statistics are even more startling: the top five tech companies alone have $662 billion in off-balance-sheet lease-related commitments. A Nikkei survey shows that the total off-balance-sheet obligations of major hyperscale cloud service providers are approximately $1.65 trillion—an eight-fold increase in four years.
SoftBank's Bet: A $40 Billion Bridge Loan
SoftBank plays a crucial role in this story that cannot be ignored.
SoftBank's cumulative investment in OpenAI exceeds $60 billion, including a $40 billion bridge loan—one of the largest bridge financings in Asian history, maturing in March 2027. S&P has revised its credit outlook for SoftBank from stable to negative, citing "its AI-related investments primarily involve startups and private companies, facing significant AI innovation risks and intense competition. OpenAI is one of its weakest credit quality investments."
A more subtle signal comes from financing difficulties. SoftBank initially planned to borrow $10 billion using its OpenAI stake as collateral. The amount was later reduced to $6 billion, and ultimately—negotiations stalled. Lenders find it difficult to value a private company without a public market price.
The market is waiting for OpenAI's IPO—already confidentially filed with Goldman Sachs and Morgan Stanley as lead underwriters—to price this layer of risk. But before that, whether SoftBank can repay its $40 billion bridge loan depends on the valuation OpenAI can achieve and how much trust the market still has in AI.
The first phase of the Ohio data center is scheduled to go live in 2028. NVIDIA's $250 billion guarantee has several years before it becomes a real risk exposure. But the credit market never waits for settlement; it prices risks in advance.
Not a Bubble, but a Flawed Financing Structure
NVIDIA is not Enron.
Over the past 12 months, it generated $96.6 billion in free cash flow and held over $13 billion in cash on its balance sheet as of the end of April. Its AA credit rating presents almost no near-term concern. Its $25 billion bond issuance in June attracted $85 billion in subscriptions, an oversubscription of more than three times.
AI demand is not fabricated. Hyperscale cloud service providers have $1.45 trillion in contractual commitments. AWS CEO Matt Garman stated that "current AI infrastructure investment is by no means speculative." He says this not just to address criticism, but because the orders are genuinely there.
The core problem is that real demand is being amplified through an opaque financing structure.
When NVIDIA is simultaneously a chip supplier, equity investor, and debt guarantor, the market cannot distinguish the interests of these three roles. For a company to use its own credit to guarantee a customer's procurement, it inevitably invites comparisons to "alchemy" in any industry and any cycle. Sal Naro, Chief Investment Officer of Coherence Credit Strategies, uses precisely this term: "opaque financial structures, off-balance-sheet transactions, and complex relationships between related parties could foster 'financial alchemy,' ultimately leading to a credit rating downgrade."
The credit market doesn't bet on whether corporate revenue will grow. It bets on whether there is a buffer if things go wrong. At this critical juncture where AI is transitioning from infrastructure construction to commercialization, the thickness of this buffer depends on a more fundamental question:
If the commercial returns from products like OpenAI's ChatGPT, Anthropic's Claude, Mistral, or xAI's models cannot support the $1.4 trillion in spending commitments—as Vested Finance CEO Viram Shah points out, OpenAI's current spending commitments are about $1.4 trillion while its revenue is about $13 billion, an imbalance that unsettles the market—then every guarantee, every bridge loan, and every CDS contract will be triggered simultaneously within the same time window.
This is not a question of a bubble. This is a question of concentration.
History offers two reference points. Cisco in 2000 was a story of "demand not existing"—the internet fiber was laid, but there wasn't enough traffic. AI will not repeat this script because the demand for computing power is real.
The subprime mortgage crisis of 2007 was a story of "risk being diversified, but no one knew where the risk was." CDOs sliced up MBS, obscuring risk visibility. The current AI financing problem is close to this logic: NVIDIA guarantees → SoftBank builds the data center → OpenAI leases computing power → OpenAI uses NVIDIA's money to buy chips. The risk appears distributed across four entities, but every link is bound on the same credit chain. If any one link breaks, all four dominoes will fall together.
Who Profits, Who Bears the Cost, and Who Pays When Defaults Occur
At this point, it's worth laying out the bill to see what each party in this cycle is earning and what they are underwriting.
NVIDIA: Earns GPU gross margins and ecosystem control. Bears guarantee exposure and credit spreads. In June, it borrowed $25 billion in bonds at a cost of 50 basis points in credit spread—benefiting from its AA-rated interest rate, but guaranteeing the debt of OpenAI, which doesn't even have a rating.
SoftBank: Earns potential equity premiums from OpenAI's IPO. Bears the $40 billion bridge loan and S&P's negative outlook. Masayoshi Son is betting that this gamble can be converted into real cash before March 2027.
OpenAI: Earns a market for AGI that does not yet exist. Bears the structural deficit of $1.4 trillion in spending commitments against $13 billion in revenue—and a postponed debt problem.
Banks: Earn underwriting fees and spreads. Bear the risk that if this wave of confidence reverses, the $182 billion in AI debt might find no buyers in the short term.
Apple: Bears nothing. And that is why it became the world's most valuable company.
Apple's capital expenditure has been declining for the past three quarters. Jay Woods, Chief Market Strategist at Freedom Capital Markets, said: "Apple was criticized for underinvesting in AI, but now it appears to have successfully avoided the capital expenditure trap." This doesn't mean Apple is not doing AI—Apple Intelligence's progress is admittedly slow—but it shows that the market is voting with its feet: an asset-light AI strategy is currently more trusted.
Three Key Points to Monitor
First, will NVIDIA's CDS break through 100 basis points? 82 is not a dangerous number—the crisis threshold for an investment-grade company's CDS is typically above 200. But the speed of the move is more critical than the absolute level. From November to July, the spread rose from 35 to 82. If this slope is maintained, it won't be long before it crosses the psychological barrier. At that point, NVIDIA's borrowing costs will rise, its guarantee capacity will decrease, and the "circular" nature of the circular financing will itself become a problem.
Second, the window for OpenAI's IPO. If OpenAI goes public within the year, SoftBank's $40 billion bridge loan will have a clear exit path, and the deepest risk in the credit chain will be distributed to the vast stock market. If it is delayed—and market sentiment is already unfavorable—pressure will transmit backwards along the guarantee chain. The first casualty won't be NVIDIA's balance sheet, but market confidence.
Third, can Seoul stop the bleeding? On July 27, SK Hynix ADR fell 7.47% to $143.02, down over 38% from its June high. On July 13, the day of a market circuit breaker, the Bank of Korea issued a report to calm the market, listing three reasons: "tight AI infrastructure demand," "not a traditional inventory cycle," and "difficult HBM capacity ramp-up." These reasons themselves are sound. But the logic of credit doesn't rely on reasons; it relies on who is willing to catch the falling knife.
In 1873, British financial writer Walter Bagehot left behind a quote that, over a century later, is printed on the instructions of nearly every credit default swap: "In a boom, everyone can borrow (everyone trusts each other). The real question is, when the time of repayment arrives, who can actually pay?"
The AI boom is approaching an inflection point. From now on, the protagonist is no longer computing power, but the ledger.
Comments