AI的免費午餐吃完了？

Original Source: Wall Street Insight

Over the past two years, the market has defaulted to the assumption that someone would always foot the bill for AI infrastructure. That assumption has now been rejected by the credit market.

On July 28, South Korea's KOSPI plunged 10.84%, triggering its 8th circuit breaker of the year. Nvidia fell 5%, SK Hynix dropped 7.47%, and Japan's Kioxia slid 18%.

On the same day, Apple rose 1.2%, reclaiming the title of the world's most valuable company. The Dow Jones gained 0.51%, and all three major European indices closed higher.

On the same day, within the same news cycle, gains and losses were completely reversed. The market wasn't panicking — it was voting. The assets being sold off were all companies deeply tied to AI infrastructure. The assets being bought were all companies not reliant on the AI infrastructure narrative. The signal triggering this vote didn't come from the stock market; it came from the bond market.

Who is Saying No

On Monday, the five-year credit default swap (CDS) for Nvidia surged 14 basis points in a single day to 82 basis points, marking the largest single-day increase for the contract. Oracle's CDS rose to 215 basis points, Alphabet's to 67 basis points, while Broadcom, Amazon, and Meta all hit new records. SpaceX's bonds issued in June broke below their issue price in the secondary market.

The pricing logic of CDS is fundamentally different from stocks. Stock pricing looks at growth potential, narratives, and the future. CDS buyers are pension funds, insurance companies, and hedge funds. They don't read stories; they only read cash flow — the cash you generate now, and whether it can cover the principal and interest when due.

For the past two years, the pricing power of AI assets has been entirely in the hands of the equity market. The assumption of equity investors was: Capital expenditure equals a moat; whoever has the larger capex wins. Every GPU is a down payment on future market capitalization. Under this narrative, the total capital expenditure of the five major tech giants for 2026 is projected at $737 billion, up from $261 billion two years ago.

The buyers in the credit market stayed quiet, because they don't trade "stories." They were just waiting.

The direct trigger for Monday's trading was Nvidia's July 25 announcement of a $500 billion AI infrastructure partnership with SK Group, as well as OpenAI's $250 billion financing guarantee and a $350 billion chip procurement financing arrangement. The potential exposure on just the OpenAI side exceeds $600 billion, more than Thailand's annual GDP. The market spent the weekend digesting the implications of this number — CDS jumped directly at Monday's open.

Before this, signals had been accumulating for three weeks. On July 9, S&P downgraded Oracle from BBB to BBB-, just one notch above junk status. Last week, Alphabet's free cash flow turned negative by $5.855 billion, the first time in its twenty years as a public company. When these events occurred in isolation, the market chose to digest rather than panic — Oracle's issues could be attributed to overly concentrated clients, Alphabet's to still being in an investment phase. Each was seen as its own problem.

Nvidia's $600 billion revolving financing plan connected the dots into a line. Oracle's biggest client is OpenAI; OpenAI is the core beneficiary of Nvidia's revolving financing plan. Alphabet, Meta, and Amazon are all operating on the same "borrow money to build infrastructure" financing model. These are not several independent events. This is the same balance sheet, multiply collateralized under different borrowers' names.

The credit market took three weeks to digest the signals, make connections, and reach a judgment. On Monday, the judgment landed.

How Credit Entered a Closed Loop

The traditional model of the semiconductor industry is unidirectional: Nvidia sells chips to cloud providers; cloud providers rent computing power to AI companies; funds flow from end-users upwards through the layers, with each layer bearing its own credit risk.

Nvidia's current model has turned this into a cycle. Nvidia provides credit to clients → Clients use this credit to purchase chips → Chip sales confirm Nvidia's revenue growth → Revenue growth supports its credit rating and valuation → A stronger balance sheet supports larger guarantees → The cycle continues.

In this closed loop, a portion of Nvidia's revenue growth doesn't come from end-user demand, but from the credit it injects itself. Sal Naro, Chief Investment Officer at Coherence Credit Strategies, calls these arrangements "financial alchemy." John Flood, head of the trading desk at Goldman Sachs, wrote in a client memo that several traders feel "uneasy" about this, pointing out that the credit market has become a key window for judging the AI narrative.

This is not a problem unique to Nvidia. The entire financing model for AI infrastructure is built on an assumption — that the cost of capital is near zero, and the market is always willing to pay for the "future." When CDS started pricing this in, the assumption was tested for the first time.

Chain Reaction

The credit market's skepticism doesn't need to verify each company one by one. It propagates down the financing chain naturally.

Oracle was the first to hit the wire. On July 9, S&P downgraded its rating from BBB to BBB-, citing three reasons directly: AI business cash burn exceeded expectations (FY2026 capex up 162% YoY to $56 billion, negative free cash flow of $23.7 billion, projected to rise further to $95 billion in FY2027); excessive client concentration (OpenAI accounts for about half of its remaining performance obligations); and insufficient internal load to absorb excess computing capacity. Oracle's CDS rose from 144 basis points at the start of the year to 215 basis points. For a company with an annual revenue of $67.4 billion and cloud infrastructure revenue surging 77%, the cost of its bond insurance doubled.

This risk points directly back at OpenAI itself. According to Fortune and the Financial Times, in 2025, OpenAI's revenue was $13.07 billion, with an operating loss of $21 billion. After accounting for one-time expenses, its net loss was around $39 billion. A company losing nearly $40 billion a year is Oracle's biggest client and the core beneficiary of Nvidia's revolving financing plan.

The chain continues. Alphabet's free cash flow turned negative. Meta's $12.55 billion bond for its Texas data center carried a 7.534% interest rate — a coupon typically seen only in the junk bond range, with a subscription multiple of just 1.6 times, less than half the average subscription multiple for US corporate bonds this year. The average borrowing cost for AA/A-rated companies across the tech industry rose from 3.9% a year ago to 5.1%, and for BBB-rated companies from 4.7% to 7.0%. Unsecured loan rates for lower-rated AI-related companies have hit 10.5%-12%. Meanwhile, the capital expenditure guidance for the five major tech giants is still being raised.

This is not the problem of a single company. This is the process of an entire financing model switching from "infinitely cheap" to "market pricing." Every widening of spreads at one level transmits to the next level, the next borrower, the next market.

The Bill from 2000

This script isn't new.

In March 2000, Cisco became the world's largest company with a market capitalization of $555 billion. Its growth engine was called "vendor financing" — lending money to unprofitable internet companies to buy Cisco equipment. At its peak, the exposure was about $4 billion, covering over 700 clients. John Chambers called it a "once-in-a-century revolution."

After the Nasdaq peaked, clients went bankrupt en masse. In May 2001, Cisco wrote off $2.2 billion in inventory and set aside $800 million in bad debt. Its stock fell from $79 to $9.5. In December 2025, Cisco's stock price finally broke above its 2000 high, 25 years later. But due to continuous share buybacks reducing share count, its current market cap is still over 40% lower than its peak in 2000 — the stock price recovered, but the company's size didn't.

From 1996 to 2000, vendor financing was priced as "ecosystem empowerment" and a "demand catalyst." After the Nasdaq crash in 2000, clients died off en masse, and Cisco's bad debts surfaced — the same stream of numbers was reread as "when exactly are you going to pay us back?"

The shift in logic is abrupt: Before the break, the equity market provided narrative legitimacy for vendor financing; after the break, the narrative vanished, and the bill was just numbers. Jim Bianco, President of Bianco Research, argues that as long as the credit market doesn't have pricing power, equity valuations can always be debated; once the credit market starts pricing risk, the debate is over.

Cisco's exposure was $4 billion. Nvidia's potential arrangements related solely to OpenAI exceeded $600 billion. The CDS market and debt buyers reacted much faster this time around — it took Cisco 14 months from its peak to the write-downs. Nvidia's CDS was repriced just one trading day after the news broke.

The Rewarded "Opt-Out"

It's no accident that Apple reclaimed the title of the world's most valuable company today.

Apple's latest quarterly capital expenditure was only $1.97 billion, down 36% year-over-year. In contrast, during the same period, Amazon set its full-year capex guidance at $200 billion, and Alphabet raised its to $205 billion. Apple's choice was to return cash to shareholders — increasing its buyback by $100 billion in the second quarter, while also raising its dividend. Tim Cook's explanation was "taking a careful and deliberate approach to capital expenditure."

It's not that Apple isn't investing in AI. The A19 chip features a built-in neural accelerator, Apple Intelligence is embedded across its entire system, and the new Siri is rebuilt based on Gemini. The difference lies in the path: on-device inference plus external model integration, refusing to build massive training clusters in-house.

From 2024 to 2025, Wall Street constantly criticized Apple for underinvesting, with the prevailing view being that building proprietary computing power equaled a moat. By the summer of 2026, the credit market had reversed that same judgment — massive capex went from being a competitive advantage to a debt burden, and a restrained balance sheet went from being conservative to flexible. Jay Woods, Chief Strategist at Freedom Capital Markets, said: "Apple was once criticized for inadequate AI investment, but in hindsight, it successfully avoided the capital expenditure trap."

This reversal was confirmed in price action on July 28. The market is not only punishing the over-leveraged side but also rewarding the side with a healthy balance sheet. These two things are two sides of the same coin, the same multiple-choice question.

The 8th Circuit Breaker of the Year

The revaluation done by the credit market on Monday took only one day to reach the accounts of South Korean retail investors.

The KOSPI saw its 8th circuit breaker of the year. Samsung and SK Hynix account for 60% of the index weighting, and these two stocks contributed about 70% of the KOSPI's gains over the past two years.

In May, regulators approved 16 2x leveraged ETFs linked to them. These ballooned from 4.9 trillion won to over $9 billion in two months, with retail investors holding 92% of the shares. On the way up, leverage amplified gains, attracted additional capital, and pushed the underlying stocks higher. On the way down, the same mechanism works in reverse: Underlying stocks fall → ETFs are forced to de-lever → Underlying stocks fall further → Margin calls → Liquidation → Further declines. In the first two weeks of July, 1.2 million leveraged accounts received margin calls, and between 320,000 and 460,000 were fully liquidated. Investors in their 20s and 30s accounted for 62% of the total losses.

This group couldn't possibly have known what was happening before the market opened on Monday. They didn't know about Nvidia's $600 billion revolving financing structure. They didn't know the reasons behind Oracle's downgrade. They didn't know the credit market was using a framework completely different from theirs to reprice AI assets. They only knew that Samsung and SK Hynix were falling, and their leveraged accounts were receiving margin calls.

This is precisely the structural characteristic of the entire transmission chain: Information flows from professional buyers in the credit market to the equity market, from the equity market to the global semiconductor supply chain, from the supply chain to the indices, from the indices to leveraged ETFs, and finally to the accounts of retail investors. Each layer is more lagging, each layer has its own leverage, and each layer bears a larger shock than the one before. A 14-basis-point jump in CDS on Monday translated into a circuit breaker on the KOSPI. Translated into leveraged accounts, it meant being wiped out.

The credit market isn't repricing just one stock. It's repricing an entire pricing framework that had taken "ever-cheap capital" as its default assumption. It's recalculating from the root, and information is transmitted and amplified down the leverage chain, level by level. Those who knew first made their directional bets; those who knew last paid the price.