Summary of Key Points
NVIDIA, in collaboration with top financial institutions such as Blackstone and Goldman Sachs, has established a $500 billion computing power financing platform. Essentially, this platform allows AI laboratories to “mortgage” GPUs by paying for them over time using future revenue from their computing services. Capital providers are willing to invest but are cautious about risks, so NVIDIA uses its own credit as a guarantee (even promising to cover 25% of the GPU’s residual value). Additionally, an “isolated layer” in the form of a Special Purpose Vehicle (SPV) is used to transfer these risks off the capital providers’ books. This four-party structure addresses the issues faced by AI laboratories, which lack the funds to purchase GPUs, capital providers who are hesitant to invest, and banks that are reluctant to lend. However, the greatest risk lies with NVIDIA, which has already drawn concerns about becoming an “unregulated bank.”
The Four-Party Structure: Essentially, a Mortgage for AI Laboratories to Buy GPUs
You can think of this structure as an “AI version of a mortgage loan”:
- Supplier (NVIDIA): Not only does it sell GPUs but also acts as the guarantor. If you don’t have enough money to buy a GPU, NVIDIA will find capital providers to lend you the funds. If you fail to repay or if the GPU’s value declines, NVIDIA will cover the difference (up to 25% of the residual value). The goal is to turn GPUs into “financiable assets” by securing large orders, with Jensen Huang stating his ambition to integrate them into a $22 trillion private equity market.
- Capital Providers (Blackstone, Goldman Sachs, etc.): They provide the funds but are risk-averse. Therefore, the money is not given directly to the AI laboratories; instead, it goes through an SPV, which keeps the risks off their books and allows them to earn stable interest (similar to lending money for property purchases).
- SPV (Special Purpose Vehicle): This is a shell company that takes on all the liabilities. It buys the GPUs and rents them to the AI laboratories, ensuring that both the capital providers’ and the laboratories’ financial records remain clean (for example, Anthropic can borrow $35 billion with no debt shown on its own books).
- End Users (OpenAI, Anthropic): They use the computing power and pay a monthly rent, similar to paying for a mortgage. This approach allows them to secure scarce computing resources in advance, as GPUs are in high demand.
These four parties complement each other: AI laboratories lack the funds, so capital providers step in; capital providers are risk-averse, so NVIDIA provides a guarantee; capital providers want stable returns, so AI laboratories sign long-term contracts (e.g., OpenAI’s 5-year lease for 10 GW of computing power).
Why Not Buy Directly? Simple Solutions Fail
Some may ask: Can’t AI laboratories just buy GPUs directly or borrow from banks? The reasons are as follows:
- Direct Purchase: AI laboratories often incur annual losses in the tens of billions, and a 1 GW data center costs $50 billion—more than they can afford. Even if they could, GPUs evolve rapidly (new models emerge annually), and their value might decrease, turning them into “inventory holders” (Anthropic’s president stated they don’t want to acquire more computing power than they need).
- Bank Loans: AI laboratories lack stable cash flows and collateral (GPUs depreciate quickly), and their business models are not yet proven, making it difficult for banks to approve loans.
- Direct Capital Lending: Capital providers’ funds come from other sources (pensions, etc.), and they are reluctant to invest in companies with uncertain futures. They prefer “guaranteed debts” rather than high-risk investments.
- NVIDIA’s Financing Role: If NVIDIA were to fund the purchases outright, it would act like a lending company, and its financial statements could be affected negatively (e.g., an increase in its CDS price indicates market concerns about its risk).
Thus, this four-party structure is not just a showcase of technical sophistication; it combines various solutions to overcome mutual barriers:
- NVIDIA covers the gaps where others fail.
- It distributes risks among all parties involved.
The Biggest Risk: Is NVIDIA Becoming an “Unregulated Bank”?
The success of this structure hinges on NVIDIA, as it acts as the guarantor and bears all the ultimate risks:
- Hidden Debts: NVIDIA promises to cover 25% of the residual value for each project, with a total limit of $125 billion (equivalent to the assets of a mid-sized bank). Although it claims GPUs will not depreciate (given their current scarcity), it still insures against this risk.
- Market Reactions: On the day the platform was announced, NVIDIA’s CDS price rose by 5.3 basis points, and its stock price fell by 2.86%. This suggests market concerns about its ability to handle these risks. Later, NVIDIA reduced its guarantee for OpenAI from $250 billion to below $120 billion to alleviate further doubts.
- Critics’ Concerns: Critics like Big Shorter view this structure as even more risky than the Enron scandal, suggesting that NVIDIA is using complex mechanisms to hide its debt (although NVIDIA denies this).
NVIDIA faces a dilemma: without providing guarantees, it cannot sell its chips; with guarantees, it risks becoming an unregulated bank (since banks have capital adequacy requirements that NVIDIA does not meet). It is the most critical component of this entire system, and its failure could lead to the collapse of the entire $500 billion deal.
Domestic Computing Power Leasing: Can This Structure Be Adopted in China?
China’s computing power leasing market is expected to exceed $260 billion this year, but the approach differs from international practices:
- Who Bears the Risks?: In China, state-owned enterprises (SOEs) fund the construction of computing centers and then rent them out. Their credit is used to secure loans and cover depreciation costs, with meager returns in the form of rental income (e.g., a Jinan Urban Investment project requires a 10-year payback period).
- Residual Value Management: There are no standardized guidelines for GPU residual value assessment, resulting in losses for some financial leasing companies (up to $300 million due to depreciation).
- Construction vs. Leasing: Many projects involve building computing centers first and then finding customers, leading to low utilization rates (sometimes less than 50%), resulting in unnecessary depreciation.
Compared to the international four-party structure, China’s approach focuses on risk absorption by state-owned enterprises, while foreign models aim to leverage capital more effectively. China has not addressed the issue of residual value management, and there is a lack of guarantors to cover potential losses. Additionally, the order of construction and leasing processes differs between the two countries.
This structure is not a one-size-fits-all solution, but it clearly defines who bears the risks. If China wants to adopt it, it must first address three key issues: who will bear the costs, how residual values will be calculated, and where the orders will come from.
Final Thoughts
Here are three key considerations for evaluation:
1. From the Guarantor’s Perspective: Who are you willing to guarantee for, and do others trust your ability to fulfill your commitments? (NVIDIA’s credit is a crucial factor.)
2. From the Capital Provider’s Perspective: Would you invest in GPUs without a chip manufacturer’s guarantee? (Creditworthiness is essential.)
3. From the User’s Perspective: Are you willing to sign long-term contracts (e.g., 3–5 years) that exchange future benefits for immediate costs, relying on your ability to survive in the long term?
The four-party structure is not mysterious; it simply involves mutual borrowing of credit and risk sharing. However, all the risks ultimately fall on the guarantor. Whether you choose to take on that role depends on individual judgment.