WuBlockchain has republished a research note from the China International Capital Corporation research department that looks at debt risks tied to the US artificial intelligence buildout through the lens of Hyman Minsky’s financial instability hypothesis. The core conclusion is that AI debt risk is still within a controllable range, and that the market is still far from a true “Minsky moment.”
The note says the bigger shift is not that debt has already become unmanageable, but that financing structures are changing. As AI capital expenditure keeps rising, some large technology companies are becoming more dependent on external funding instead of internal cash generation.
AI infrastructure spending is pushing large tech firms toward heavier balance sheets
CICC says major US technology companies used to operate under a relatively light-asset, high-cash-flow model, with limited capital expenditure and investment largely funded by operating cash flow. As AI infrastructure construction accelerates, that model is shifting toward heavier assets and much higher capex, which in turn is lifting external financing demand.
According to the note, capital expenditure at the five major cloud providers now equals 97.4% of operating cash flow. Their combined bond issuance in the first half of 2026 reached about $170 billion, already 1.5 times the amount issued during all of 2025.
CICC places that trend inside Minsky’s framework, which holds that prolonged stability can breed instability. In long expansions with loose financing conditions, markets may become overly optimistic, risk appetite can rise, and corporate financing structures can gradually move away from more conservative forms toward models that rely more on leverage and refinancing. If expectations later fail and borrowing costs rise, defaults, deleveraging and asset sales can reinforce each other and trigger a “Minsky moment.”
The report uses three Minsky financing categories
The note follows Minsky’s classification of corporate financing into three types.
- Hedge finance: operating cash flow is sufficient to cover both principal and interest, so repayment does not depend on new borrowing.
- Speculative finance: cash flow can cover interest but not principal, which means the company must keep rolling over debt.
- Ponzi finance: operating cash flow cannot even cover interest expenses, so repayment depends on rising asset prices or asset sales.
CICC says this framework also works for AI investment. The key questions are what stage current AI financing has reached, whether it could create systemic financial risk, and what that implies for the durability of the AI investment cycle.
The five large cloud firms still cover debt service with operating cash flow
To measure debt-servicing capacity, the report uses two indicators. The first is the cash-flow interest coverage ratio, defined as operating cash flow divided by interest expense. The second is the debt service ratio, defined as the present value of principal and interest obligations divided by operating cash flow.
In the report’s framework, a cash-flow interest coverage ratio below 1 would imply Ponzi finance because the company could not even pay current-year interest. If the ratio is above 1 but the debt service ratio is above 1, the company can cover interest but not maturing principal, which points to speculative finance. If interest coverage is above 1 and the debt service ratio is below 1, operating cash flow can cover the full debt burden, which qualifies as hedge finance.
On the latest data, all five major cloud companies have cash-flow interest coverage ratios above 1. Oracle is the lowest at about 7x. Amazon is about 48x, while Google is the highest at about 82x.
Debt service ratios are also below 1 for all five. Oracle is about 48%, Amazon about 12%, and Google about 6%. CICC says that means the group can still rely on core businesses such as search, advertising, cloud services and SaaS software to generate stable operating cash flow and cover current principal and interest obligations. Bond issuance has increased sharply over the past two years, but overall repayment capacity still looks solid.
Relative comparisons show Oracle is the weakest of the group
The note says absolute indicators are not enough on their own, so it compares the five cloud firms with listed companies across sectors in the S&P 500. On that basis, Microsoft, Google, Meta and Amazon still rank ahead of the broader market in operating cash flow protection, while Oracle stands out as weaker.
Although Oracle is a technology company, its cash-flow interest coverage ratio has fallen below the technology sector average and now looks closer to sectors such as utilities and real estate, which are more typical heavy-asset industries. CICC says those sectors usually carry higher capex burdens and depend more on rolling debt, so debt risk tends to be higher as well.
Free cash flow is under pressure, with Amazon and Oracle already negative
The picture becomes more mixed under a stricter stress test based on free cash flow, defined as operating cash flow minus capital expenditure. Over the past 12 months, Microsoft, Google and Meta still generated healthy free cash flow, but Amazon and Oracle have already turned negative.
CICC notes that Amazon is classified under GICS as consumer discretionary rather than pure software or internet services. Its capital expenditure includes not just AWS data centers, but also logistics networks and warehouse construction. That naturally makes Amazon’s capex and free cash flow metrics larger than those of other cloud providers, so business structure needs to be taken into account in side-by-side comparisons.
The report also points to Google’s latest results for the second quarter of 2026, which showed quarterly free cash flow falling to negative $5.9 billion for the first time since the company was listed. CICC says that reflects a period in which AI capex growth has outpaced growth in cash flow from core operations. If revenue growth in those core businesses does not accelerate materially, reliance on external financing will increase.
Still, the note says negative free cash flow does not automatically mean debt risk is out of control. AI infrastructure spending is cyclical by nature. As data center construction gradually moves toward completion, capex growth may slow and free cash flow could improve again. The report also notes that many technology companies have gone through periods of negative free cash flow during fast expansion phases, and says heavy upfront spending is normal in the early stage of a new technological revolution.
Debt maturity profiles remain favorable, with low short-term refinancing pressure
Under the Minsky framework, another key question is whether debt depends on constant refinancing and whether repayment relies on higher asset prices or asset sales. To address that, CICC looks at the share of short-term debt in total debt and the size of cash plus short-term marketable securities.
The five cloud firms all show relatively low short-term debt ratios, well below the 27.4% average for S&P 500 constituents. Microsoft is the highest at about 14.7%, while Meta is the lowest at about 2.2%.
Bond maturities are also longer than average. Amazon is at about 8 years and Microsoft about 11.8 years, both above the roughly 5-year average maturity for US corporate bonds. CICC says longer-term funding lowers near-term refinancing pressure and better matches debt maturities with the investment cycle for AI infrastructure such as data centers, reducing maturity mismatch risk.
Effective borrowing costs remain around 2%
The report says the actual financing cost of debt for the five cloud providers is still relatively low. Even though US risk-free rates have risen sharply since 2022, the firms’ effective debt rates have not moved up in the same way and are currently around 2%, below the 4.4% yield on the 10-year US Treasury.
CICC gives two reasons. One is that these companies locked in debt during the low-rate period. The other is that some of them have borrowed in lower-rate non-dollar currencies, including Swiss francs, which has helped reduce overall financing costs. The note says this is another reason why longer debt tenors matter. They reduce exposure to a rapid rise in financing costs. That said, borrowing costs would still rise if credit risk worsened materially.
Cash buffers remain strong for Google, Microsoft and Meta
On liquidity, Google, Microsoft and Meta all hold substantial cash, cash equivalents and short-term marketable securities. After subtracting interest-bearing debt, they still remain in net cash positions. CICC says that means they retain solid debt repayment capacity even without relying on fresh borrowing, and their repayment does not depend on rising asset prices or asset sales.
Amazon’s net cash balance turned negative in the second quarter of 2026, but the note describes the size of that shortfall as limited and again points to the company’s different industry classification. Oracle, by contrast, shows a more visible net cash deficit, a higher dependence on external financing, and greater financial fragility.
Off-balance-sheet borrowing is a risk area the report flags
The note says one issue worth tracking is off-balance-sheet borrowing. Citing a report from the Bank for International Settlements, CICC says the use of “shadow borrowing” by large cloud providers to fund AI data centers is increasing rapidly.
That financing often takes the form of private credit. A common structure involves a special purpose vehicle, or SPV, or a joint venture set up to buy or build data center assets. Capital for the entity usually comes from a consortium of investors, while the cloud provider holds only a minority stake. The cloud company may commit to long-term operating leases or compute purchase agreements and provide credit support, with debt serviced through lease cash flow.
The risk, the report says, is that if AI compute demand after completion is not enough to cover long-term lease costs, those commitments could turn into an added liability and raise financial pressure further. Some institutions estimate this hidden debt at about $1.65 trillion.
CICC also points to an offset. Microsoft, Google and Amazon still report large cloud-service backlogs, totaling about $1.45 trillion. If those orders are realized as planned, they should largely cover the liabilities tied to those lease structures.
CICC says the AI debt cycle is still in its early-to-middle stage
Taking all of the company-level data together, the report concludes that the five cloud firms as a group are still in the early-to-middle stage of the AI debt cycle and have not yet shown a classic financing imbalance.
The central risk, in CICC’s view, is not “debt that is too large” but “a financing structure that is changing.” As AI capex continues to expand, financing is gradually moving from internal funding toward external borrowing, or in Minsky terms, from hedge finance toward speculative finance. The report says that some degree of “financing migration” has begun, but that does not mean a Minsky moment has arrived.
By company, Microsoft and Meta still fit the profile of classic hedge finance. Google and Amazon are showing early signs of shifting from hedge finance toward speculative finance. Oracle, because of continued pressure on free cash flow and negative net cash, appears more financially fragile than the others.
US household and corporate leverage remains relatively low
CICC then widens the lens to the macro level, arguing that AI company financing conditions alone do not determine whether debt risk could become systemic. Broader leverage in the US economy matters as well.
For households, the note says US household debt was about 66% of GDP in the first quarter of 2026, well below the nearly 100% peak seen before the 2008 subprime crisis. The household debt service ratio was about 8%, broadly stable and also far below the roughly 12% high before that crisis. CICC says that points to a meaningfully improved household balance sheet after a long period of repair.
For corporations, US non-financial corporate debt is about 70% of GDP, below the average level of the past decade. The corporate debt service ratio is about 37%, near historical lows. The note says AI investment has pushed up borrowing by some large technology companies, but the broader US corporate sector has not moved into a generalized high-leverage state and overall debt burdens remain manageable.
Part of the leverage has shifted from the private sector to the government
The report says one reason private-sector debt burdens remain relatively low is that leverage has shifted toward the government sector. Since the pandemic, the US government has run large fiscal expansion programs and significantly wider deficits.
In the note’s framing, government deficits correspond to private-sector net assets. A larger deficit leaves the private sector with more net assets, meaning the government has absorbed more of the leverage burden while private-sector balance sheets have improved.
Bank capital is strong and broker-dealer leverage remains subdued
CICC also argues that the resilience of the financial system lowers the chance that current AI debt turns into systemic risk. Since 2008, under tighter financial regulation, capital adequacy in the US banking system has stayed at relatively high levels. Common equity tier 1, or CET1, ratios for both global systemically important banks and other commercial banks remain in historically elevated ranges, which the report says shows a solid capacity to absorb risk.
The asset-to-equity ratio of broker-dealers also remains below levels seen during the subprime crisis period. That suggests leverage inside the financial system is not especially severe.
AI funding relies more on bond markets than on bank credit creation
Another distinction in the report is the way this investment cycle is financed. CICC says current AI capex is being funded more through bond markets, unlike the property bubble period, which depended far more on bank credit expansion.
Bond financing, the note says, is essentially a reallocation of existing funds among investors. It does not directly create money and does not produce the credit multiplier effect associated with the banking system. As a result, the main risk sits with bond investors rather than on bank balance sheets. In that sense, AI debt risk looks more like capital-market credit risk than bank asset-quality deterioration, and spillover into the banking system should be relatively limited.
Putting those factors together, CICC says the current US AI investment wave still looks more like a high-capex cycle than a systemic debt event. Debt deserves attention, but it does not yet amount to systemic risk. The bigger issue is whether future returns on AI investment can continue to support growth in operating cash flow. If capex keeps exceeding cash generation for too long, financing structures could evolve from hedge finance to speculative finance or even Ponzi finance, bringing Minsky-style instability closer.
CICC says the recent AI asset pullback looks like repricing and rotation
The report also uses this framework to interpret recent market moves in the US. Since July, AI-related assets have been under pressure, which has fueled debate over whether the AI bubble is over. CICC says the move looks more like a repricing of the investment thesis combined with sector rotation than a broad collapse in risk appetite.
First, the market had a technical need for adjustment. Over the past year, AI became one of the most crowded trades in global markets. In the second quarter especially, upstream parts of the AI chain, including chips and memory, posted large gains and attracted both leveraged capital and trend-following money. CICC says that once new macro variables emerge, such as higher expectations for Federal Reserve rate hikes or worsening geopolitical risk in the Middle East, those crowded assets become more vulnerable to profit-taking.
Investors are focusing more on returns than on capex growth alone
The second factor is uneven profit distribution across the AI chain. The note says upstream segments such as chips and memory have captured excess profits because of tight supply, while downstream cloud providers, corporate customers and consumers have borne much of the cost. CICC argues that this allocation of profits is difficult to sustain over the long run.
As AI infrastructure investment moves deeper into the cycle, the market is paying closer attention to capital returns rather than capital input alone. The note points to Google’s latest earnings as a case in point. Revenue and profit both beat market expectations, but the stock still pulled back after management raised future capex guidance. CICC reads that as a sign that investors are no longer rewarding unlimited expansion by default and are placing more weight on efficiency.
Improving US fundamentals are helping funds rotate into value and defensive sectors
The third factor is that stronger US macro fundamentals are helping drive fund rotation and broader market participation. CICC says fixed-asset investment and consumption have stayed firm this year, the labor market has shown signs of improvement, and manufacturing PMI has continued to recover. Markets have started to price in the possibility of renewed Federal Reserve tightening, which has pushed both nominal and real Treasury yields higher.
The note also says new Federal Reserve Chair Warsh has emphasized reform and signaled tighter balance-sheet control, adding to concerns about future liquidity conditions. In a repricing of interest rates, technology stocks often face valuation pressure, while value and defensive groups such as financials, energy and utilities tend to attract flows. That is why CICC sees the recent US market move as a classic sector rotation rather than a broad-based withdrawal of risk appetite.
The AI investment thesis is moving from scale to efficiency
The final section of the note argues that the AI wave may not be over, but the logic of investing in it is changing. The market is moving from a first phase centered on scale to a second phase centered on efficiency and returns.
For cloud providers, that means stricter capital discipline and a need to prove that AI spending can generate sustainable returns. For chip and memory companies, excess profits are cyclical and should normalize as capacity expands. For software companies, large-model developers and AI application firms, competition is likely to intensify, and the key differentiators will be cost control, product efficiency and commercialization rather than model size alone.
CICC says the AI industry chain is now going through a revaluation process in which the driver shifts from capital to efficiency. That process may come with valuation adjustments and profit redistribution, but it also points to a more mature stage in the AI investment cycle. In the report’s view, tighter capital constraints, healthier earnings models and fuller market competition would help improve capital allocation efficiency and support more sustainable growth across the AI industry.