Is the free lunch of AI over?

For the past two years, markets have assumed someone would always foot the bill for AI infrastructure. That assumption has now been rejected by credit markets.

On July 28, South Korea's KOSPI index fell by 10.84%, triggering its eighth circuit breaker of the year.$NVIDIA (NVDA.US)$ down 5%,$SK hynix (SKHY.US)$Japan declined by 7.47%,$Kioxia Holdings (285A.JP)$down 18%.

On the same day, $Apple (AAPL.US)$ up 1.2%, regaining the title of world’s most valuable company. The Dow Jones Industrial Average rose 0.51%, and Europe’s three major equity indices all closed higher.

On the same day, within the same news cycle, market movements were completely divergent. There was no panic—markets were voting. The assets voted out were those deeply tied to AI infrastructure; the assets voted in were those that did not rely on the AI infrastructure narrative. The signal triggering this vote originated not in equity markets, but in debt markets.

Who Is Saying No

On Monday, NVIDIA’s five-year credit default swap (CDS) spread surged by 14 basis points in a single day to 82 basis points, marking the contract’s largest one-day increase.$Oracle (ORCL.US)$ CDS spreads widened to 215 basis points, and Alphabet’s CDS rose to 67 basis points,$Broadcom (AVGO.US)$、$Amazon (AMZN.US)$、 $Meta Platforms (META.US)$ all setting new records.$SpaceX (SPCX.US)$ Newly issued bonds in June traded below par in the secondary market.

CDS pricing logic is fundamentally different from equities. Equity valuation focuses on growth potential, narratives, and future prospects. CDS buyers—pension funds, insurance companies, and hedge funds—do not read stories; they analyze cash flows: whether the cash you generate today can cover principal and interest obligations at maturity.

Over the past two years, equity markets have held full pricing power over AI-related assets. Equity investors operated under the assumption that capital expenditure equaled a moat—the greater the capex, the greater the competitive advantage—and that every GPU purchased was a down payment on future market value. Under this narrative, the combined capital expenditures of the Big Five tech firms are projected to reach $737 billion by 2026, up from $261 billion two years ago.

Credit market participants remained silent because they do not trade 'stories.' They were simply waiting.

The immediate catalyst for Monday’s market turmoil was NVIDIA’s announcement on July 25 of a$SK (034730.KR)$ $500 billion AI infrastructure partnership, along with OpenAI’s $250 billion financing guarantee and a $350 billion chip procurement financing arrangement. OpenAI’s potential exposure alone exceeds $600 billion—more than Thailand’s annual GDP. Markets spent the weekend digesting the implications of this figure, leading to an immediate jump in CDS spreads at Monday’s open.

Prior to this, signals had been accumulating for three weeks. On July 9, Oracle was downgraded by S&P from BBB to BBB—just one notch above junk status. Alphabet reported negative free cash flow of $5.855 billion in its earnings last week—the first time in its two decades as a public company. When either event occurred in isolation, markets chose to absorb rather than panic—Oracle’s situation could be attributed to excessive customer concentration, while Alphabet could still be seen as being in an investment phase; each issue appeared distinct.

NVIDIA’s $600 billion revolving financing facility connected the dots. Oracle’s largest customer is OpenAI, which is a primary beneficiary of NVIDIA’s revolving credit arrangement. Alphabet, Meta, and Amazon all operate under the same 'borrow-to-build-infrastructure' financing model. These are not isolated incidents. They represent the same balance sheet, repeatedly pledged under different borrower names.

Credit markets spent three weeks digesting signals, establishing linkages, and forming a judgment. On Monday, that judgment crystallized.

How Credit Entered a Closed Loop

The semiconductor industry traditionally operated on a linear model: NVIDIA sold chips to cloud providers, who then leased computing power to AI companies, with capital flowing upstream from end users layer by layer—each entity bearing its own credit risk.

NVIDIA’s current model has turned this into a loop: NVIDIA extends credit to customers → customers use that credit to purchase chips → chip sales confirm NVIDIA’s revenue growth → revenue growth supports its credit rating and valuation → a stronger balance sheet enables larger guarantees → and the cycle continues.

Within this closed loop, part of NVIDIA’s revenue growth stems not from end-user demand but from credit it injects itself. Sal Naro, Chief Investment Officer at Coherence Credit Strategies, refers to such arrangements as 'financial alchemy.' John Flood, head of Goldman Sachs’ trading division, wrote in a client memo that multiple traders expressed 'discomfort' with this dynamic, noting that credit markets have become the critical window through which the AI narrative is now assessed.

This is not a problem unique to NVIDIA. The entire AI infrastructure financing model rests on one assumption—that the cost of capital approaches zero and that markets will always be willing to pay for the 'future.' For the first time, that assumption is being tested as credit default swaps (CDS) begin pricing in risk.

Chain Reaction

Credit market skepticism does not require verifying each company individually—it propagates automatically down the financing chain.

$Oracle (ORCL.US)$ It was the first to breach the threshold. On July 9, S&P downgraded its rating from BBB to BBB-, citing three reasons: AI-related capital expenditures are burning cash faster than expected (FY2026 capex is projected to surge 162% year-over-year to $56 billion, with free cash flow at negative $23.7 billion, and further rising to $95 billion in FY2027); excessive customer concentration (OpenAI accounts for roughly half of its remaining performance obligations); and insufficient internal workloads to absorb excess computing capacity. Oracle’s CDS spread has widened from 144 basis points at the start of the year to 215 basis points. For a company with annual revenue of $67.4 billion and cloud infrastructure revenue soaring by 77%, the cost of bond insurance has doubled.

This risk points directly to OpenAI itself. According to Fortune and the Financial Times, OpenAI generated $13.07 billion in revenue in 2025 but reported an operating loss of $21 billion and a net loss of approximately $39 billion after accounting for one-time charges. A company losing nearly $40 billion annually is Oracle’s largest customer and also the primary beneficiary of NVIDIA’s revolving financing facility.

The chain continues. Alphabet’s free cash flow has turned negative, and Meta’s $12.55 billion bond issuance for its Texas data center carries a coupon rate of 7.534%—a level typically seen only in high-yield (junk) bonds. The offering attracted just 1.6x subscription, less than half the average book-to-cover ratio for U.S. corporate bonds this year. Across the tech sector, average borrowing costs for AA/A-rated firms have risen from 3.9% a year ago to 5.1%, while those for BBB-rated firms have jumped from 4.7% to 7.0%. Unsecured loan rates for lower-rated, AI-related companies have already reached 10.5%–12%. Meanwhile, capital expenditure guidance from the top five tech giants continues to rise.

This is not an issue confined to a single company. It reflects a systemic shift in the financing model—from 'infinitely cheap' capital to market-based pricing. Every widening spread cascades down to the next tier, the next borrower, and the next market.

The Bill from 2000

This script isn’t new.

In March 2000,$Cisco (CSCO.US)$it became the world’s largest company by market capitalization at $555 billion. Its growth engine was called "vendor financing"—lending money to unprofitable internet companies to purchase Cisco equipment. At its peak, it had an exposure of approximately $4 billion across more than 700 clients. Chambers described it as a "once-in-a-century revolution."

After the Nasdaq peaked, clients began collapsing en masse. In May 2001, Cisco wrote off $2.2 billion in inventory and recorded $800 million in bad debt provisions. Its stock price plummeted from $79 to $9.50. Cisco’s share price finally surpassed its 2000 high in December 2025—25 years later. However, due to continuous share buybacks that reduced its outstanding shares, its current market capitalization remains more than 40% below its 2000 peak—its stock price recovered, but the company’s scale did not.

Between 1996 and 2000, vendor financing was framed as 'ecosystem enablement' and a 'demand catalyst.' After the 2000 Nasdaq crash, as clients collapsed en masse, Cisco’s bad debts surfaced—and the same numbers were suddenly reinterpreted as: 'When exactly are you going to repay us?'

The shift in logic is abrupt: before the break, equity markets provided narrative legitimacy for vendor financing; after the break, the narrative vanished, leaving only the numbers on the bill. As Jim Bianco, president of Bianco Research, puts it: 'Until credit markets assert pricing power, equity valuations can always be debated. Once credit markets begin pricing risk, the debate ends.'

Cisco has an exposure of USD 40 billion. NVIDIA alone has potential arrangements related to OpenAI exceeding USD 600 billion. This time, the CDS market and debt buyers reacted much faster than in the previous cycle—Cisco took 14 months from peak to write-off, whereas NVIDIA’s CDS began repricing within a single trading day after the news disclosure.

The Rewarded 'No'

Apple reclaiming the title of the world’s most valuable company today is no coincidence.

Apple’s latest quarterly capital expenditure was only USD 1.97 billion, down 36% year-over-year. In the same period, Amazon set its full-year capex guidance at USD 200 billion, while Alphabet raised its guidance to USD 205 billion. Apple opted instead to return cash to shareholders—announcing an additional USD 100 billion in share buybacks for the second quarter and increasing its dividend. Cook explained this approach by stating that the company has 'always taken a prudent and deliberate stance on capital expenditures.'

Apple is not avoiding AI investment altogether. The A19 chip includes a built-in neural engine, Apple Intelligence is now embedded across its entire operating system suite, and the new Siri has been rebuilt using Gemini. The distinction lies in its strategy: on-device inference combined with access to external models, eschewing the construction of large-scale in-house training clusters.

From 2024 to 2025, Wall Street repeatedly criticized Apple for insufficient investment, with the prevailing view at the time being that building proprietary computing infrastructure equated to a competitive moat. By summer 2026, however, credit markets had reversed that judgment—massive capex shifted from being seen as a competitive advantage to a debt burden, and a restrained balance sheet transformed from appearing conservative to strategically flexible. Jay Woods, Chief Strategist at Freedom Capital Markets, remarked: 'Apple was previously criticized for underinvesting in AI, but in hindsight, it successfully avoided the capital expenditure trap.'

This reversal was price-confirmed on July 28. The market is not only penalizing those with excessive leverage but also rewarding companies with sound balance sheets. These two outcomes represent opposite sides of the same strategic choice.

Eighth Circuit Breaker of the Year

The reassessment conducted by credit markets on Monday reached Korean retail investors’ accounts within just one day.

The KOSPI triggered its eighth circuit breaker of the year. $Samsung Electronics (005930.KR)$ and$SK Hynix (000660.KR)$Together accounting for 60% of the index weight, these two stocks contributed approximately 70% of the KOSPI’s gains over the past two years.

In May, regulators approved 16 leveraged ETFs offering 2x exposure to their underlying indices. Within two months, assets under management surged from KRW 4.9 trillion to over USD 9 billion, with retail investors holding 92% of positions. On the upside, leverage amplifies gains, attracts incremental capital, and pushes up the prices of the underlying stocks. On the downside, the same mechanism works in reverse: a decline in underlying stocks triggers passive ETF deleveraging, which further depresses those stocks, leading to margin calls, forced liquidations, and continued downward pressure. During the first two weeks of July, 1.2 million leveraged accounts received margin calls, and between 320,000 and 460,000 accounts were fully liquidated. Retail investors aged 20–30 accounted for 62% of those incurring losses.

This group could not possibly have known what was happening before markets opened on Monday. They were unaware of NVIDIA’s USD 600 billion revolving financing structure, the reason for Oracle’s credit downgrade, or that credit markets were repricing AI-related assets using a framework entirely different from their own. All they knew was that Samsung Electronics and SK Hynix were falling—and that their leveraged accounts had received margin calls.

This is precisely the structural characteristic of the entire transmission chain: information flows from professional credit-market buyers to equity markets, from equity markets to the global semiconductor supply chain, from the supply chain to indices, from indices to leveraged ETFs, and ultimately reaches retail investor accounts. At each layer, the signal becomes more delayed, additional leverage is layered on, and the magnitude of the shock absorbed exceeds that of the preceding layer. A 14-basis-point spike in CDS spreads on Monday translated into a trading halt on the KOSPI index—and into account wipeouts for leveraged retail traders.

What credit markets are repricing is not a single stock, but an entire pricing framework that had taken 'perpetually cheap capital' as a default assumption. The recalculation begins at the root, and information propagates along the leverage chain—stage by stage, magnification by magnification. Those who knew first took directional positions; those who knew last paid the price.

Editor/Deng