The article "AI Shadow Credit Part I: The Hidden Bombs in the AI Boom" exposed a critical risk in AI infrastructure: the 2000 internet bubble was fueled by equity and collapsed linearly, whereas the current AI buildout is debt-driven, threatening a bond market crisis similar to 2008's subprime debacle. The process of acquiring land, borrowing for construction, signing take-or-pay contracts, securing mortgage financing, and issuing ABS for data centers mirrors real estate development. However, AI infrastructure is even more dangerous due to the rapid depreciation of semiconductors.

Currently, the capital expenditure growth of the five major cloud vendors has outpaced cash flow growth, with Google and Amazon's free cash flow turning negative. If growth slows and the second derivative declines, they will enter a phase of relying on refinancing to survive. The article also highlighted "shadow loans" hidden off the balance sheet, making debt levels appear safer while actually increasing risk.

Last week, new evidence emerged: Nvidia partnered with Apollo, BlackRock, Goldman Sachs, and KKR to leverage over $500 billion in third-party capital for AI computing infrastructure through a financial platform. Nvidia's announcement of financing support for AI infrastructure was initially seen as a positive, but its stock fell about 3%. This market reaction is worth studying, as it suggests investors are questioning whether the financial leverage on AI capital expenditure has become too high. Apollo, Blackstone, and KKR are prominent players in the "private credit" market, which experienced significant crisis events in 2025. They offer flexible, fast lending without the strict regulations of traditional banks, but with low transparency, high leverage, and no mandatory stress tests, making the hidden risks unknown. Previously, "shadow lending" was limited to the five major cloud vendors, but with Nvidia's orchestration, it is now spreading to smaller third-party cloud providers, accelerating the financialization of computing power and raising a critical question: how much of AI's revenue growth is driven by genuine demand versus capital expenditure?

Identifying the Real AI Demand

The fact that Google and Amazon's free cash flow has turned negative is a landmark event. Initially, AI infrastructure was supported by core business cash flows like advertising, cloud computing, software revenue, and merchandise commissions. However, starting this year, growth has become increasingly dependent on financing. This is not just a leverage risk but also questions the authenticity of revenue, which in turn exacerbates leverage risk.

Take OpenAI as an example. If it relied solely on equity financing, it could barely afford training costs. In reality, OpenAI and Anthropic don't pay cloud vendors in cash; they "book" the costs, similar to airlines leasing from Boeing or Airbus. Repayment takes various forms: part is from the cloud vendors' investment funds, meaning the investment isn't truly cash but computing power converted into equity; part is in the form of computing credits, as the cloud vendors are heavy AI users who offset their payments with "computing vouchers"; the remainder is based on the expected future revenue growth of OpenAI and Anthropic, which is the ARR everyone currently tracks.

This is like a village shop run by your father-in-law, who didn't ask for a bride price when you got married, borrowed money from you to buy stock, and you could only buy from him to support his business. Initially, it was daily necessities, but now you're buying expired items—this is "financing creating demand." On the surface, Microsoft increases capital expenditure to buy GPUs and build data centers, generating AI service revenue. AI companies then purchase cloud services, creating new revenue. From major model providers to the five cloud vendors, data centers to third-party computing rentals, and Nvidia to upstream suppliers, orders seem prosperous. In reality, a significant portion of this revenue is interlinked—you owe me, I owe him, he owes you—forcing circular service usage, making AI revenue lack "independence" and not fully organic.

When investing in the AI supply chain, one must look beyond orders. The key is how much of AI demand originates from end users versus intra-industry spending. The latter, while not "false demand," is at least discounted demand. Genuine independent demand should be defined as pure cash flow revenue excluding cloud vendors' self-built AI, strategic investments, computing pre-payments, and cloud credit "computing vouchers." This represents the true willingness to pay for AI services over the long term. Independent demand is hard to quantify, but a proxy is the "second derivative of capital expenditure" from the previous article, a mathematical approach to bypass financial accounting. The first derivative of capital expenditure is its growth rate, while the second derivative reflects whether demand growth is accelerating. Even if capital expenditure converts efficiently into demand, it can only drive the first derivative. For the second derivative to rise continuously, new external demand is needed, such as the demand for AI programming and agents this year, which sustained the second derivative's upward momentum in the first half. This is why the previous article argued that if the second derivative falls, it indicates that new revenue cannot match new capital expenditure, forcing reliance on refinancing.

As AI transitions into a credit cycle, it's crucial to navigate the AI supply chain's debt relationships. The ultimate debtors, if traced to the bottom, are not a single AI company but various participants in a long financial chain. On the surface, AI model providers and cloud vendors are the debtors, leading many to believe AI debt risk is minimal. They have long-term contracts promising annual GPU purchases for the next decade. From an accounting perspective, this may not immediately appear as debt, but it's essentially backed by future cash flows, corresponding to AI cloud computing and major model vendors' ARR (with overlapping parts). However, increasingly, the actual borrowers are off-balance-sheet data center SPVs. More entities are using project financing by creating special SPVs to purchase and build AI assets, taking on debt, with revenue from long-term leases from OpenAI and Anthropic. Legally, the debt lies with the SPV, invisible on the balance sheets of the five cloud vendors and Nvidia, but it's effectively their debt and that of AI major model providers, which is why AI infrastructure has become akin to real estate. In real estate, the actual operating project companies' assets and liabilities are often not reflected in listed company reports. Early AI infrastructure debt belonged to banks, such as loans for building data centers. Many consider AI financing risk-free based on current bank financial health. However, banks typically don't hold all risk long-term; they package it into ABS and sell it, as in 2008. Traditional bank safety doesn't mean society has no financial risk.

The real ultimate creditors are institutional investors. ABS is sold to pension funds and insurance companies with high demand for investment-grade bonds offering long-term cash flows, making AI infrastructure bonds attractive. There are also private credit funds behind SPVs, investing in private loans for higher returns than public bonds. Ultimately, the true bearers are ordinary investors. Understanding the AI supply chain's debt relationships leads to identifying the weakest links, which are debt risk observation points.

Assessing the Most Vulnerable Links

This issue has two analytical angles: the timing of pressure emergence and the severity of eventual losses.

Angle 1: Who Feels Cash Flow Pressure First?

The first link to feel pressure: AI application companies and small AI startups. This is the most fragile layer in the ecosystem. Computing power is expensive, competition is fierce, and many AI startups have costs exceeding revenue. They rely on continuous funding to burn cash despite fast growth. If AI demand slows, clients cut AI budgets, or API price wars increase customer acquisition costs, these companies will die first.

The second link, crisis amplifier: data center operators. Data centers have extremely heavy assets and high debt ratios, still expanding through financing, with revenue dependent on a few large clients. GPU depreciation is rapid. If large clients reduce purchases, vacancy rates rise, and new financing dries up, they will quickly fall into debt crisis—this is the link closest to real estate developers.

The third link, financing hub: the five major cloud vendors and Nvidia. Currently, they are the main debtors, acting as the AI debt's "load-bearing wall." They are generally not considered risky due to strong balance sheets, diverse clients, and AI being just a part of their cloud business. Stored GPUs can be repurposed for general cloud computing if AI demand falls. In a future crisis, they are more likely to experience "declining investment returns" rather than "debt crisis," but a 50% valuation drop for shareholders is likely.

The final link: major model providers like OpenAI and Anthropic. While a crisis may start with slowing revenue growth for them, their financing is primarily equity without rigid maturing debt. If not publicly listed, they are actually the safest. However, differences between OpenAI and Anthropic are greater than between them and other links. Anthropic derives about 80% of revenue from enterprise clients with high stickiness and stable income. Its unit economics have moved out of loss territory, and its cash burn rate is converging to a manageable level. In contrast, OpenAI is more fragile and volatile, with about 60% of revenue from disloyal individual consumers. Its cash burn rate is extremely high, remaining around 57% until 2027, with no clear path to positive cash flow. Anthropic's debt is protected, such as in the Apollo and Blackstone TPU computing financing project worth about $35 billion, where its bonds leveraged the credit ratings of its investment-grade supporters. OpenAI, however, has no real co-signer; Microsoft removed all structural support in April 2026. Microsoft has internally reviewed OpenAI's books for years and may have recognized the credit risk. Another major supporter, SoftBank, raised a $10 billion margin loan using OpenAI equity as collateral. When lenders hesitated, SoftBank even had to provide personal guarantees. While OpenAI's unique position means its own debt risk is low, it is the most likely "trigger" for a crisis in the entire AI ecosystem.

Angle 2: Who Suffers Most if a Crisis Occurs?

If the worst happens, like the 2008 global debt crisis, the safest in the supply chain—meaning no debt issues, though shareholders face a minimum 50% stock decline—is Nvidia. Next are OpenAI and Anthropic for reasons explained earlier. Then come cloud giants except Oracle, as the impact would be on profit margins rather than debt crises. Oracle stands out with a debt service ratio (debt principal present value/operating cash flow) of only 48%. If risk levels rise and financing rates increase, it could immediately face a debt crisis.

The above links mainly involve equity impairment risk. Beyond Oracle, the links with bankruptcy risk include: first, AI startups relying on expansion financing; if financing slows, cash flow dries up immediately. While financial impact is limited, their large number could significantly affect the real economy. Second, AI data centers with numerous financing projects; whether from project delays during booms or declining lease prices during downturns, this could trigger a crisis. Data centers are typical heavy asset businesses with known risks, so their financing is evaluated strictly. The riskiest links are not these but third-party GPU rentals and new computing cloud companies. Despite flashy tech packaging, their core business model is buying large numbers of GPUs to rent out. They currently profit from GPU price spreads and computing inflation premiums. With extremely high leverage, if demand becomes oversupplied, they immediately become insolvent.

The most typical example is CoreWeave, described as the "Evergrande of the AI cloud computing industry." The company leverages AI-optimized technology stacks and efficient capital operations to become one of the few professional cloud vendors capable of large-scale, rapid delivery of latest AI infrastructure, benefiting from structural supply-demand imbalances. But its most innovative aspect is its financing structure: using clients' long-term contracts as underlying assets for debt financing, obtaining funds at costs far below its own credit rating, creating a "orders-financing-expansion-more orders" growth flywheel. This essentially mimics real estate's rapid expansion model. The company has extremely high debt, with a debt-to-equity ratio of 13.8, sustained massive capital expenditure, persistent negative net profit, and an adjusted operating profit margin guidance of 8%, far below its long-term target of 25%–30%. Delivering hundreds of billions to trillions of dollars in orders in a short time makes its supply chain highly vulnerable. Since all its contracts are mortgaged, if revenue growth falls below expectations or GPU depreciation, asset impairment pressure will quickly crush its balance sheet. High-leverage, fast-expanding third-party computing rentals like CoreWeave are the most likely links to explode in the future.

Predicting Potential Risk Events

After assessing risks across links, more practical predictions can be made. If the crisis escalates, several risk events may occur sequentially.

The first potential risk event: OpenAI's IPO. Currently, OpenAI and Anthropic's payments rely not on revenue but on order financing (i.e., credit). So financing ability is not just about raising capital but also core to operating costs, which is Sam Altman's forte. For favorable financing conditions to continue, each new financing round's valuation must be higher than the previous. Thus, OpenAI's valuation increase isn't just asset appreciation; its IPO pricing is critical. If priced too low or if the stock opens high and drops, it will harm future financing ability. Additionally, private credit information can be kept confidential, but an IPO will accurately disclose every financial detail, making it a potential risk event, possibly occurring late this year or early next year.

The second potential risk event: borrowers cutting computing commitments to preserve cash. If major model companies' ARR can't sustain growth, order obligations may shrink, constituting contract breaches. This would be a landmark event, similar to the first "unfinished buildings" after a real estate bubble, triggering government vigilance and account freezes.

The third potential risk event: default events first hit the star new cloud computing companies of recent years, like CoreWeave, Lambda, and Crusoe. As analyzed earlier, when GPU purchase and rental spreads invert, they are effectively insolvent. The first to fall could trigger an "Evergrande moment," with CoreWeave's collapse akin to Country Garden, and Oracle, seen as relatively safe, like Vanke. The key is that when one event occurs, others become inevitable; only time remains.

The fourth potential risk event: freezing of the entire AI credit system. Remaining performance obligations would be repriced as counterparty risk, GPU-backed notes wouldn't roll over, and upstream supply chain companies (including Chinese computing power chain firms) would face massive asset impairments. By then, debt risk is fully exposed, primarily impacting stock markets. Equity values would suffer devastating blows. Until impairments complete, all company valuations would be driven to historic lows, including the five (possibly four remaining) cloud vendors, entering a prolonged "earnings killing" cycle.

But AI differs from real estate crucially: technology continues to advance, actual penetration rises, and demand grows rapidly. The infrastructure was just too ahead. Ultimately, participants with the strongest capital, most cautious AI infrastructure investments, and smallest risk exposures will acquire shelved data center projects at low prices and guarantee necessary lease contracts. This could be Apple, Nvidia, or Warren Buffett. By then, the debt crisis-induced recession will have lasted years, with the Fed lowering rates to under 3%. The de-leveraged AI supply chain, starting a new round of large-scale infrastructure at lower costs, can then enter the next boom cycle.