Debt risk across the AI supply chain is migrating from large cloud computing companies toward chip suppliers.
According to four credit reports published by Goldman Sachs' sales and trading division last Sunday, the credit default swap (CDS, which can be understood as the market's pricing of corporate default risk) spread of a basket of chip companies comprising Broadcom (AVGO), Advanced Micro Devices (AMD), NVIDIA (NVDA) and Intel (INTC) has, for the first time in nearly three months, surpassed the CDS basket of hyperscalers such as Oracle (ORCL), Meta Platforms (META), Amazon, Alphabet (GOOGL) and Microsoft (MSFT).
This means investors are becoming more concerned about the risks chip companies must bear: on one hand, the accelerating replacement cycle of high-priced graphics processing units (GPUs) is increasing depreciation pressure; on the other, chipmakers are supporting downstream customers' expansion through financing, guarantees and other means. The market is asking: if returns on computing power investment fall short of expectations, who ultimately absorbs the equipment depreciation and financing losses?
Meanwhile, weekly trading volume in AI-related default risk contracts has risen to US$8 billion to US$10 billion, roughly seven times the level a year earlier. Goldman Sachs credit trader Alisha Pasi put it bluntly: "Risk is migrating."
Why are chipmaker risks suddenly rising?
The most direct signal comes from changes in default risk pricing.
In early August this year, the CDS spread for chip companies was about 80 basis points, roughly 20 basis points lower than that of large cloud providers; three months later, that spread has doubled to about 160 basis points and has exceeded cloud providers for the first time. A widening spread typically means the market believes companies' future debt repayment risk is increasing.
At the individual stock level, Broadcom's CDS spread has widened by about 70 basis points from its low two months ago, AMD's by 55 basis points and NVIDIA's by 50 basis points. By contrast, Meta Platforms and Alphabet widened by only about 15 basis points, while Microsoft showed almost no notable change.
Last week, among the US investment-grade corporate default risk index, the three companies with the largest spread widening all came from the chip sector: Intel widened 40 basis points to 141 basis points, AMD widened 26.5 basis points to 147 basis points, and Broadcom widened 17.6 basis points to 186 basis points.
This reflects the market's repricing of the AI risk chain: although large cloud providers have enormous capital expenditures, they have more stable cash flows, stronger balance sheets and greater financing capacity; chip suppliers, however, face more complex risks — they must withstand depreciation pressure from rapid iteration of high-end chips and may also indirectly bear downstream demand risk through customer financing, order commitments and other means.
The core question has changed: who will absorb the ever-increasing chips?
Goldman Sachs believes 2027 could become a key stress test for AI financing risk. By then, the market is expected to see more than US$150 billion in computing power financing, primarily through special purpose vehicles.
So-called special purpose vehicle financing typically packages assets such as GPUs, data center equipment or power contracts and finances them based on future lease cash flows. The problem is that the collateral for such financing is not traditional factories or land, but GPUs whose technology iterates extremely quickly.
Once a new generation of chips is launched and computing power prices fall, the residual value of old equipment may shrink rapidly; if downstream customers' orders, lease income or financing capacity run into problems, risk could further transmit from data center operators to chip suppliers.
As a result, the market's focus is shifting from "how much more will cloud computing giants borrow" to "who ultimately bears the depreciation and collection risk on the chips these giants and data center operators buy."
Cloud providers are still issuing large amounts of debt, but financing costs are clearly rising
Risk does not mean large cloud providers are already safe. On the contrary, they remain the main issuers in the AI financing wave; the market simply believes for now that their risk resistance is stronger than that of chip suppliers.
Last week, US investment-grade corporate bond issuance totaled about US$76 billion, nearly twice market expectations, making it the fourth-largest issuance week of 2026 and the largest week on record for November. Year to date, total US investment-grade bond issuance has surpassed US$2.079 trillion, setting a record.
Among them, hyperscalers were the main driving force. Since October 29 alone, five separate bond deals from cloud computing giants have appeared in the market, totaling about US$67 billion. Global hyperscaler bond issuance for the year has reached US$296 billion, clearly exceeding Goldman Sachs' previous forecast of US$250 billion.
However, the cost for investors to absorb this wave of supply is rising. Year to date, the average new issue spread was 3.4 basis points, but it has recently risen to 7.8 basis points; in November, the average new issue spread for AI-related bonds exceeded 18 basis points, significantly higher than for non-AI corporate bonds.
Yields on some investment-grade AI bonds are even higher than those on 83% of bonds in the high-yield index. This means that although these issuers still have relatively high credit ratings, investors have begun demanding higher returns to compensate for the uncertainty brought by AI capital expenditures and debt expansion.
Cloud providers are unwilling to lock in long-term high-cost debt
Another notable change is that cloud providers are shortening bond maturities.
For example, in Meta Platforms' recent US$15 billion bond issuance, 30-year bonds accounted for only US$2.75 billion, and for the first time it did not issue 40-year bonds; Oracle's US$10 billion issuance had a maximum maturity of only 20 years, with no 30-year or 40-year tranches; Apple's US$7.5 billion issuance also did not include 30-year bonds.
This shows that both issuers and investors are more cautious about ultra-long-term AI debt. Issuers are unwilling to lock in long-term financing costs when rates are high, and investors are unwilling to be overly exposed to far-term risks from AI investment returns that have not yet been fully validated.
The high-yield bond market is already feeling the pressure first
The spread of risk to chips and data centers is also more evident in the high-yield bond market.
Goldman Sachs data show that last week the 29 AI-related high-yield bonds it traded fell by an average of 1.67 points, while other high-yield bonds rose by an average of 0.36 points.
Among higher-rated BB bonds, AI and data center-related bonds fell by an average of 1.52 points, about 4.5 times the decline in BB energy bonds. Although data center bonds account for only 6.7% of the BB index, less than half of the energy sector's 14.3% share, their drag on index performance was twice that of the energy sector.
In addition, data center expansion is facing policy resistance. Goldman Sachs statistics show that last week 41 local elections in the United States involved data center construction bans, tighter regulation or specific project approvals, and 40 of those outcomes were unfavorable to data center construction. Goldman Sachs economists estimate that short-term construction bans could affect about 45% of planned US construction capacity.
This means data center projects face not only rising financing costs, but also practical risks such as construction approval, power supply guarantees, community resistance and project delays.
The stock market fell sharply, but the credit market was unusually calm
What is worth watching is that stocks and bonds have recently shown a clear divergence.
Last Thursday, the S&P 500 fell 4.4% in a single day, and the Nasdaq market lost about US$2.5 trillion in value; yet the investment-grade corporate default risk index widened by only 3.5 basis points that day, far below the normal reaction when stocks have seen similar declines historically.
For the full week, the S&P 500 fell 2.7%, investment-grade credit spreads widened by only 1 basis point, high-yield credit spreads even narrowed by 4 basis points, and the leveraged loan market recorded its best weekly performance since April this year.
Goldman Sachs traders believe this "stocks down, bonds stable" pattern does not necessarily mean risk has disappeared; instead, it may mean the credit market has not yet fully reflected the pressure accumulated during the AI investment cycle. On the surface, credit indices remain stable; but at the individual bond level, divergence across AI, chips and data centers has clearly intensified.
2027 may see concentrated release of financing pressure
Goldman Sachs expects that as hyperscalers' free cash flow turns negative, debt financing has already funded about 36% of their 2026 capital expenditures.
Looking ahead to 2027, the market is expected to see about US$440 billion in direct bond issuance from hyperscalers and about US$300 billion in AI project financing enter the market, covering multiple channels including investment-grade bonds, high-yield bonds, private credit and structured financing.
In addition, hyperscalers' off-balance-sheet commitments have reached as high as US$3.6 trillion, of which about US$685 billion will mature before the end of 2027.
This means the real test for AI financing has not yet arrived. If interest rates remain high, data center construction is blocked, growth in computing power demand slows, or GPUs are updated faster than assets can be recovered, risk may further transmit from chipmakers to project financing, private credit and the high-yield bond market.
Goldman Sachs trader Avishek Biswas believes the current stage is more one of "high uncertainty" than "broad high risk." But the market has clearly begun pricing tail risk in advance: indices may still look calm, yet pressure on chip suppliers and data center financing has already surfaced first.
The core question of AI debt risk is shifting from "how much more can cloud giants borrow" to "who will pay for rapidly depreciating chips, continuously expanding data centers and still-unproven computing power demand."
Comments