So far, AI companies have been willing to shoulder higher debt costs, but some investors say they are starting to worry about future financing.
At the same time, part of the risk exposure behind already-issued AI debt may be obscured.
As U.S. Treasury yields climb to their highest level since 2007, companies that rely on debt financing will have to face higher and rising borrowing costs.
This means the AI infrastructure buildout, which has already reached historic levels in this cycle, will become even more expensive.
JPMorgan previously estimated that by 2030, as data center companies and other firms tied to the AI boom race to expand capacity, these businesses will issue $4.1 trillion in AI-related debt.
So far, AI companies have been willing to bear higher debt costs, but some investors say they are beginning to worry about future financing.
At the same time, part of the risk exposure behind already-issued AI debt may be obscured.
Financing costs keep climbing
When AI companies with bond issuance needs return to the market, they find the 10-year Treasury yield has approached 5.17%, up about 1 percentage point since the start of the year, meaning companies issuing AI bonds will have to offer more attractive returns to lure investors.
The good news is that the market is not yet in a state of panic over this.
Heavily indebted CoreWeave shares are still performing well, rising nearly 8% this week.
But Oracle, which has also relied on bond issuance to expand its AI business, has had a tougher time, falling 7% this week and about 30% year to date.
Siebert Financial chief investment officer Mark Malek said, "The relevant companies, at least for now, are not sensitive to higher Treasury yields and the resulting higher bond issuance costs, and they can all accept higher financing costs. In my view, many of these companies indeed need to be price-insensitive because they need to obtain as much capital as possible to compete in this AI boom."
At the center of the AI boom are leading model developers OpenAI and Anthropic, both valued at close to $1 trillion in private markets.
To provide the infrastructure needed for advanced models, as well as models and services for many other companies, the hyperscalers among traditional U.S. tech companies — Amazon, Google, Meta and Microsoft — have committed to hundreds of billions of dollars in capital expenditure this year, with increases expected by 2027.
Malek said, "Debt financing provides healthy funding for these companies' development, but these tech giants all have investment-grade credit ratings and can access relatively cheaper funding sources compared with the broader market. But for many other AI companies, there are clearly greater challenges ahead."
KBRA global head of corporate, project and infrastructure finance Andrew Giudici said that even though rising Treasury yields may affect AI bond trading, he still does not expect a major impact on corporate bond issuance demand.
"In a normal environment, market participants might step back and pause bond issuance. But in the AI corporate bond issuance market, I don't think that will happen. We will continue to see relatively large AI corporate bond issuance."
Early warning signs appear
Although the relevant companies are not yet sensitive to changes in financing costs, a senior private credit investor who declined to be named said that in the future, large AI companies such as CoreWeave will also find it harder to raise financing because their buffers for absorbing costs will become smaller.
Mitsubishi vice president Riley Thompson said that even if companies willing to finance and borrow agree to pay higher interest costs, lenders themselves are becoming increasingly picky about the projects they are willing to fund.
"The market may not really be interested in 50 companies like CoreWeave, but only 20," he said.
CoreWeave, which went public last year, warned about rising interest rates in its filing with the U.S. Securities and Exchange Commission.
The company said in its latest quarterly report that as of June, based on its outstanding floating-rate debt balance, every 100 basis point (1 percentage point) increase in interest rates could raise its interest expenses by $30 million.
The latest warning signal in the AI bond financing market came from Oracle.
Oracle issued a "force majeure" notice related to its New Mexico data center project to protect itself from higher costs, after which Oracle shares fell.
Reports said that if the project fails to come online as expected in 2028, the company is considering delaying payments for the project.
Oracle said the project "remains on schedule."
In addition, rising financing costs brought by higher Treasury yields are not the only problem facing related corporate bond issuance.
Before Treasury yields surged, the CEOs of Anthropic and OpenAI had already begun urging a slowdown in the pace of AI development, after industry researchers publicly expressed concerns that advanced models could escape human control.
At the same time, strong opposition nationwide to AI data centers is growing.
In a recent poll supported by SurveyMonkey, 69% of respondents said they opposed building AI data centers locally.
On Monday, Texas Republican Governor Greg Abbott, after pausing grid approvals last month, directed the state's environmental regulator to suspend all data center-related approvals until the state's Electric Reliability Council completes a review of the existing queue of data centers seeking grid connection.
On one side are rising costs and accumulating risks in the financing market, while on the other side demand for AI services continues to surge.
The latest example is Meta's Muse personal assistant app, whose popularity has soared since its launch in early September.
Muse surpassed 2.5 million global downloads in its first two weeks, exceeding ChatGPT's previous record as the most downloaded app on Apple's App Store.
Evercore senior managing director Mark Mahaney said Muse is expected to reach 100 million users within 6 to 12 months.
Latham & Watkins vice chair of emerging companies and growth and a specialist in AI infrastructure financing Haim Zaltzman said, "There is no doubt that as bond issuance costs rise, someone has to bear them. But under the current demand structure, absorbing these costs is still quite easy because demand is so strong."
Hidden debt risks
More notably, beyond bond issuance, AI corporate debt risk comes more from hidden "off-balance-sheet guarantees."
According to media data, over the past year, U.S. tech giants have provided up to $300 billion in guarantees for AI data centers and chips.
These arrangements have helped support this construction boom, and most of the risk exposure has not been immediately recorded as debt because it does not appear in the main line items of any financial statement, but is hidden in the notes.
The formal name for this type of guarantee is a "residual value guarantee" (RVG).
It works by having a special purpose vehicle (SPV) borrow money, buy computing equipment or build data centers, while the tech company endorses a minimum future value.
If these assets are ultimately sold or leased at a price below that minimum future value, the tech company must make up the shortfall.
The structure can be broken into three layers: the SPV borrows to buy chips or machine rooms, and the borrowing is supported by the cash flow from a usage contract with an AI company.
Once the AI company stops paying, the assets are subleased or sold to repay debt.
If a gap remains, the guarantor — usually the chipmaker or cloud provider — covers it.
Zaltzman said that for investors, this structure allows tech companies to expand AI capacity while preserving credit ratings and conserving cash, because it hides potential liabilities off the statements.
But if AI demand falls short of expectations, hardware depreciates faster than expected, or the industry builds more computing power than customers need, this structure will trigger asset sales.
Meta, Broadcom and Nvidia all use this structure.
CreditSights views this "residual value guarantee" as "selling a put option."
The firm said this practice and structure cost almost nothing during the AI boom, but are fatal in a severe industry downturn when customers default and hardware values fall.
Moreover, this procyclical model amplifies the magnitude of industry booms and busts.
DoubleLine portfolio manager Mariya Entina said this is like exploiting regulatory loopholes and trying to get special treatment from rating agencies. "When a certain type of financial structure becomes prevalent, it obscures the true face of finance."
In any case, the "residual value guarantee" is not an ordinary investment-grade bond issuance model, and given its off-balance-sheet nature, tail risk is relatively high.
Comments