Are tech giants "secretly borrowing money" for AI, bringing back the specter of the subprime mortgage crisis?
Author | Wang Hanyu
Editor | Zhang Fan
Recently, Meta issued an announcement confirming that the company will invest $600 billion in the United States by 2028 for infrastructure construction such as building artificial intelligence data centers and talent recruitment.
One week earlier, this technology giant indirectly completed a financing of approximately $30 billion through the establishment of a Special Purpose Vehicle (SPV) for the construction of data centers.
Meanwhile, it was reported that Alphabet, the parent company of Google, will issue another euro - denominated bond this year, with a total scale of at least €3 billion, following the issuance of its first €6.75 billion bond.
On the other hand, the Credit Default Swap (CDS) of Oracle soared in September this year, indicating the market's concern about the potential debt - default risk caused by its large - scale borrowing for AI infrastructure investment.
Some even believe that this performance should be regarded as an early warning signal of the "AI bubble" - the 2008 subprime mortgage crisis started with the widespread rise of CDS.
As of the end of September this year, the financing scale of technology companies in the US bond market has reached $157 billion, a 70% increase compared with the same period last year. Since the fourth quarter, these companies' financing activities for AI infrastructure have only increased.
A series of debts have pushed up the leverage ratio of technology companies, expanded their credit risks, and may spread to their soaring stock prices - some voices are repeatedly warning about the "AI bubble".
So, does the market's concern about the debt risks of technology companies mean that the current huge AI spending is too aggressive? What signs can investors look for to identify the "AI bubble"?
The "AI bubble" is still in its early stage
On September 10th, Oracle signed a $300 billion computing power procurement contract with OpenAI, which finally boosted its stock price, causing it to skyrocket by 36% on that day and increasing its market value by $251 billion.
Also in September, Oracle's CDS also rose significantly. By the end of the month, the 5 - year CDS rate exceeded 60 basis points. In the first three quarters of this year, this indicator had been hovering between 30 - 45 basis points.
At the beginning of this month, Oracle's 5 - year CDS rate reached 87.7 basis points, almost doubling compared with before August, reaching the highest level in the past 18 months.
The trends of Oracle's 5 - year CDS and stock price since August this year
The significant increase in Oracle's CDS is mainly due to the market's concern that its large - scale AI spending will affect the company's financial health. This sentiment is even spreading to the huge debts borne by AI companies for infrastructure investment.
But does Oracle's situation represent the overall performance of US technology companies?
Judging from the debt - to - asset ratio, Oracle currently significantly exceeds other AI giants. As of the first quarter of fiscal year 2026, Oracle's debt - to - asset ratio is about 85%. In the same period, the debt - to - asset ratios of NVIDIA, Alphabet, Microsoft, etc. are between 25% - 45%.
As of September 16th, Oracle's debt - to - equity ratio is also significantly higher than that of many AI companies.
Debt - to - equity ratios of major AI companies
It seems that Oracle's high - leverage style does not seem to be universal, and the subsequent rise of CDS is also not universal.
In addition, considering the operating data of each company, the profit growth of many leading AI companies remains relatively strong.
For example, Alphabet's total revenue in the third quarter was $102.346 billion, a 16% year - on - year increase, and its net profit was $34.979 billion, a 33% year - on - year increase. Its growth momentum mainly comes from the strong demand for cloud services and AI business. The company specifically pointed out that the revenue from products based on Google's generative AI models has increased by more than 200% year - on - year.
In comparison, although Oracle's cloud business performance in the third quarter was slightly lower than analysts' expectations, its cloud revenue still increased by 25%. At the same time, its net profit increased by 22% year - on - year.
The underlying logic of this performance is that the continuous development of AI technology has indeed driven the efficient growth of productivity in many industries. In other words, the demand for AI actually exists.
This makes the current huge capital investment of technology companies in AI construction still within a reasonable range.
As the founder and managing director of Xin Hongrui Investment Management Company, Xia Yuchen's current investment business in the UK covers both primary and secondary markets, and AI is one of his research themes. He also analyzed for 36Kr that, referring to the historical experience left by previous major financial bubbles, currently AI is not in a classic bubble state.
"Although we can see some signs of a bubble in the stock prices of some companies, it is still in a relatively early stage," Xia Yuchen mentioned.
Compared with the Internet bubble in the early 2000s, at that time, the price - to - earnings (PE) ratios of many Internet companies had reached 100 times or even 200 times. Some unprofitable startups could obtain valuations of billions of dollars just based on concepts.
Currently, the valuations of AI giants are significantly lower than the above levels. For example, NVIDIA, the leader in AI chips, currently has a PE ratio of about 56 times; Microsoft has a PE ratio of about 36 times; and Alphabet has a PE ratio of about 28 times.
In addition, around 2000, there were fewer Internet users, and there were even fewer relevant application scenarios. The business models were still in the conceptual stage. Against this background, the radical investment style gave rise to the Internet bubble.
Today, it is easy for people to experience AI applications and feel the convenience it brings to work and life. This allows the market to obtain more data for reference when making judgments about this industry and assisting investment decisions.
This shows that most companies during the Internet bubble were based on concepts and expectations, while the valuations of current technology giants due to their increased investment in AI are more based on actual profitability and business fundamentals.
Therefore, to some extent, the cautious attitude of some people may be a kind of "PTSD" after experiencing the 2000 Internet bubble.
AI investment with "minority - stake control": Will the shadow of the subprime mortgage crisis reappear?
Although the technological transformation brought about by AI and the resulting market demand have been proven, the huge capital investment required for infrastructure and the debt pressure from borrowing still pose challenges to the debt and cash - flow management of technology companies.
A recent research report from Goldman Sachs China pointed out that large - scale borrowing for data - center construction is a typical model for US companies in the AI competition, while Chinese companies are much more conservative. It is estimated that by 2027, US cloud giants will invest nearly $700 billion in total in data - center construction, while Alibaba, Tencent, ByteDance, and Baidu in China will invest less than $80 billion in total.
In the benchmark test, the system performance on both sides is roughly the same.
It also stated that the next - stage growth of US companies may be based on a model similar to Oracle's, borrowing more debt and using more complex structured financing.
Meta's recent financing behavior confirms this prediction. In October this year, it established a Special Purpose Vehicle (SPV) to complete an indirect financing for the construction of a data center in Arizona, USA.
Previously, as an innovative financial tool, SPV was widely used in the real - estate sector before the subprime mortgage crisis. Its typical operation mode is as follows: Banks package housing mortgages into Asset - Backed Securities (ABS) and sell them to a third - party SPV. The SPV repackages these debts and issues Mortgage - Backed Securities (MBS) to obtain funds for taking over the bank's debts.
In this way, the housing - mortgage debts originally belonging to the bank are transferred to the SPV, and the mortgage - default risk originally borne by the bank is also transferred to the investors of MBS.
The complex structure of SPV makes it difficult for investors to penetrate the quality of the underlying assets - when the subprime - loan default rate exceeded 15% in 2007, the rating adjustment of MBS was delayed by 6 months - so this financing method successfully evaded supervision for a large number of subprime loans.
Back to the present, Meta obtained the funds for data - center construction by jointly establishing an SPV with Blue Owl Capital Inc. This SPV, as an independent financing entity, issued bonds and completed a financing of nearly $30 billion through Morgan Stanley.
Meta used this fund to develop and construct the data center, and then lease and operate it. Since Meta only retains 20% of the equity of this SPV, the $30 - billion debt borne by the latter does not need to be included in Meta's balance sheet.
In short, Meta invested 20% of the equity in this SPV to achieve "minority - stake control" in the construction of the data center, which significantly "beautified" its balance sheet while indirectly obtaining the funds for data - center construction - the $30 - billion debt is isolated within the SPV, and in the subsequent 16 - year lease, Meta will fulfill the actual principal - and - interest repayment obligation in the form of rent.
Like Meta, under the pressure of cash - flow and financing costs, more US technology companies are also starting to use the "minority - stake control" method to transfer debt pressure outside the company. For example, xAI of Elon Musk adopted the SPV structure in its latest $20 - billion fundraising; the AWS division of Amazon also established multiple SPVs to finance data - center projects in different regions.
The SPV model can prevent huge debts from affecting the credit rating of the main body and can also attract different types of investors such as pension funds. If AI companies follow suit in the future and design more complex structures for asset - securitization operations on the current basis, it will further increase the regulatory difficulty of project assets and the credit of the main body, and may even lead to the reappearance of the shadow of the subprime mortgage crisis.
However, since 2008, SPV has been set with stricter information - disclosure standards and risk - control mechanisms. On the other hand, it also meets the technology companies' need for risk isolation - usually, the investment scale of AI infrastructure is huge, and the failure of a single project may have a major impact on the parent company. In addition, the regulatory requirements of different jurisdictions also vary greatly.
Xia Yuchen told 36Kr that technology companies' use of the SPV structure for financing may also be for the purpose of reducing compliance costs.
Previously, several US companies received fines from local regulatory authorities for violating local special regulations on data management in the EU. Holding relevant projects through an SPV can avoid the parent company being dragged down by compliance issues.
How to identify the "AI bubble"?
From the perspective of the "early AI bubble" theory, how should investors judge when a "classic bubble" appears in the AI field?
Xia Yuchen proposed two quantitative indicators: First, how much of the new investment in the industry comes from loans, and whether it has exceeded the same period in previous years, such as the level during the Internet bubble. Second, the change rate of company stock prices and similar stock prices, and whether it has reached an unsustainable level.
"Linear or rapid growth is common, but a parabolic - like increase is not appropriate," he further explained.
Looking back at the current situation along this line of thought, currently, the debt levels of AI companies are significantly lower than those during the Internet bubble, and there is no trend of prevalent junk bonds or short - term debt dominance, which may also indicate that the debt structure and scale are still within a safe range.
However, Xia Yuchen also warned: "There is currently a bubble, but it has not reached a very radical stage. If this continues, there may be a bubble similar to the Internet bubble more than 20 years ago."
Compared with around 2000, a significant difference is that the trading efficiency of the stock market is now higher. This gives the market better self - adjustment ability.
During the Internet bubble, investors had to communicate with traders by phone to place orders, and any operation and cognitive change took a longer cycle. In the mobile era, the modern electronic trading system enables the market to react and adjust more quickly.
This means that even if a bubble appears, the adjustment cycle required by the market will be shorter.
Just like the "Internet bubble" more than 20 years ago, it took the market three years to deflate it. During the COVID - 19 pandemic, it only took one year to deflate the bubble formed by the Fed's interest - rate hikes.
At the same time, from 2004 to 2006, the Fed raised interest rates 17 times in a row to curb inflation, causing the mortgage - repayment interest rate to rise sharply and triggering a wave of homeowner defaults, which ultimately led to the burst of the "real - estate bubble". Currently, the AI investment boom is in an interest - rate cut cycle, and its macro - background is completely different from the subprime - mortgage period.
*Disclaimer:
The content of this article only represents the author's views.
The market is risky, and investment should be cautious. In any case, the information in this article or the opinions expressed do not constitute investment advice to anyone. Before making an investment decision, if necessary, investors must consult professionals and make careful decisions. We have no intention of providing underwriting services or any services that require specific qualifications or licenses to the trading parties.
Follow us for more information