Summary of Key Points
This article highlights the debt risks underlying the current AI infrastructure development: Unlike the internet bubble in 2000, which was fueled by equity financing, the AI industry is driven by debt, similar to the real estate sector (acquiring land, borrowing money, and issuing ABS). This approach is compounded by the rapid depreciation of semiconductors. NVIDIA, in collaboration with institutions, has secured $500 billion in third-party capital for AI infrastructure, but its stock price has fallen, reflecting market concerns about high leverage. AI revenue involves a significant amount of internal transactions (mutual credit agreements and using computing power to offset costs), which may overestimate actual demand. The debt is hidden off-balance sheet (e.g., through SPV companies), ultimately borne by ordinary investors. The most vulnerable part of the chain is third-party computing power leasing services (such as CoreWeave), which could face risks such as failed IPOs, order breaches, and credit freezes in the future.
Why has AI infrastructure become a "debt version of real estate?"
In 2000, the internet bubble relied on equity financing, and its collapse only affected the stock market. Today, AI infrastructure relies entirely on borrowing: data centers acquire land, build facilities, sign long-term contracts with fixed payments, and use mortgage financing to issue ABS (packaging loans into financial products for sale), just like real estate developers' cycles of "acquiring land → borrowing money → selling properties → acquiring more land." However, the risk is greater because semiconductors like GPUs depreciate within 3-5 years, much faster than buildings.
Recently, NVIDIA, along with private equity firms like Apollo and Blackstone, raised $50 billion in funding, which should have been positive (allowing them to sell more GPUs), but its stock price dropped by 3% due to market fears of excessive leverage. These private lending institutions operate without the strict regulations of banks; their lending processes are fast but opaque, and they do not conduct stress tests (e.g., assessing whether borrowers can repay in a downturn). Previously, only the five major cloud providers used off-balance sheet debt; now NVIDIA has drawn in smaller cloud companies, spreading this shadow lending phenomenon.
How much of AI revenue is actually internal circulation?
Much of AI revenue comes from internal transactions within the industry chain. For example, OpenAI uses Microsoft's computing power without paying cash, recording the transaction as a credit; Microsoft invests in OpenAI not with cash but by offsetting its investment with computing power. Large model companies also use "computing power vouchers" to settle payments with cloud providers. It’s like buying goods from your father-in-law’s small shop—you can only buy what he sells, often at inflated prices (a form of "financing creating demand").
To assess real demand, we need to exclude these internal transactions and look at the "second derivative of capital expenditure" (in simple terms, the rate of change in growth). If the growth rate of capital expenditure is slowing down, it indicates that new revenue is not enough to support further investment, and more borrowing is needed to sustain operations.
Who will ultimately bear the burden of AI debt?
On the surface, it seems that large model companies (like OpenAI) and cloud providers (such as Microsoft and Amazon) are in debt, but the actual debt is hidden off-balance sheet. They establish special shell companies (SPVs) to build data centers and borrow money, which are not reflected in their parent company’s financial statements. Banks package these loans into ABS and sell them to pension funds and insurance companies (which seek stable long-term returns), with private equity funds also purchasing a portion. Ultimately, this money comes from ordinary investors (such as through pension and insurance products), meaning the public is bearing the cost.
This is similar to the 2008 subprime mortgage crisis: banks sold mortgages off-balance sheet, appearing safe on their books, but the social risk remained unchanged.
Which parts of the industry chain are most vulnerable?
- AI startups: High costs for computing power and fierce competition mean they rely on continuous financing. Any decline in demand will lead to their demise.
- Data center operators: With heavy assets and high debt, their revenue depends on a few large clients. A decrease in orders or increased vacancy rates can result in defaults.
- Cloud providers and NVIDIA: They are the backbone of the industry, with strong balance sheets, but AI investment returns may decline, potentially causing their stock prices to fall by 50%.
- Large model companies: OpenAI is more at risk than Anthropic—60% of its revenue comes from non-frequent users, and it has a high cash consumption rate (157% of revenue is spent on expenses); it will struggle to profit before 2027. Anthropic, on the other hand, gets 80% of its revenue from corporate clients and is already profitable.
Who is in the worst position?
- Third-party computing power leasing companies (e.g., CoreWeave): Like the real estate sector’s Evergrande, they have a high leverage ratio (13.8:1), relying on a cycle of "orders → financing → expansion" with all contracts mortgaged. If GPU rental prices fall below depreciation rates, they will immediately become insolvent.
- Oracle: With a debt service ratio of 48%, even a slight increase in interest rates could lead to defaults.
- AI startups: Their cash flows dry up once financing stops.
Four potential future risks:
1. Problems with OpenAI’s IPO: If its financial details are not disclosed during the IPO, or if the stock price drops after listing, it will affect its ability to raise funds (since it relies on credit agreements).
2. Large model companies reducing computing power orders: If future revenue commitments (ARR) fail to grow, they may default and reduce orders, similar to unfinished real estate projects.
3. Collaps of third-party computing power companies: Companies like CoreWeave could fail first, causing panic in the market.
4. AI credit freezes: If loans cannot be extended, upstream companies (e.g., A-share-based computing power providers) will experience asset impairment, leading to a stock market crash.
However, AI is different from real estate: technology is still evolving, and actual demand is growing, though the infrastructure development has surpassed current needs. After a crisis, financially strong companies (like Apple, NVIDIA, and Buffett) may acquire failing projects at low prices and wait for interest rates to decline before resuming operations.
This article serves as a reminder that behind AI's prosperity lie hidden debt risks. We should not focus only on order growth but also on leverage levels and actual demand. Just as no one expected the subprime mortgage crisis in 2008, the shadow lending associated with AI could become the next major financial issue.