Hook
A freshly funded AI chip giant with a $2.5 trillion market cap carries nearly $30 billion in off-balance-sheet commitments. The narrative is seductive: another tech darling hiding debt, another Enron-in-waiting. But the math doesn't add up — not because the number is small, but because the category is wrong. These are not liabilities; they are pre-emptive purchases of future capacity, a bet on the durability of AI demand.
Context
Nvidia's dominance in AI accelerators is undisputed: ~80% market share, gross margins above 70%, and a product cycle (Blackwell → Rubin) that keeps competitors at bay. The concern raised by analysts and journalists is that the company's off-balance-sheet obligations — primarily long-term purchase commitments with TSMC, SK Hynix, and cloud GPU providers — have swelled to nearly $30 billion. Under US GAAP, these are not recorded as liabilities on the balance sheet; they appear as footnote disclosures in the 10-K under "contractual obligations" or "purchase obligations." The term "off-balance-sheet liability" is technically imprecise, but it resonates as a red flag in a market still scarred by the 2008 financial crisis and the collapse of Terra/Luna.
Based on my risk audit experience, I've seen similar structures in the DeFi lending space — where "off-balance-sheet" often concealed toxic risk through maturity mismatches. Nvidia's case is structurally different, but the pattern of narrative-driven fear is identical.
Core
Let me deconstruct the $30 billion figure into its four components.
1. Irrevocable Purchase Orders with TSMC (CoWoS and advanced nodes) Nvidia secures wafer capacity through long-term agreements that obligate it to pay for a minimum volume of wafers and CoWoS packaging slots. These are not leases — they are product purchase commitments. The accounting treatment is straightforward: no asset or liability is recognized on the balance sheet until the supplier delivers the goods. The cash outflow is contingent on future production. As of Nvidia's FY2024 10-K, the total purchase obligations were approximately $18 billion, with the majority due within 12 months. The remaining $12 billion comes from HBM guarantees and cloud service provider commitments.
2. HBM Prepayment Agreements with SK Hynix and Samsung High-bandwidth memory is the bottleneck in AI chip performance. Nvidia has signed multi-year prepayment contracts to secure HBM3E and future HBM4 supply. These are structured as product prepayments, not loans. The risk is not default but underutilization: if AI demand softens, Nvidia may be forced to take delivery of memory it cannot sell, leading to inventory write-downs.
3. Lease Obligations for DGX Cloud Infrastructure Nvidia's cloud service (DGX Cloud) involves renting data center space and power. These are operating leases, and under ASC 842, the right-of-use asset and lease liability are recorded on the balance sheet — so they are not truly off-balance-sheet. However, some analysts exclude them from the headline number, inflating the perceived hidden liability.
4. Guarantees to GPU Cloud Providers Nvidia provides certain guarantees to firms like CoreWeave, promising to buy back GPUs if the cloud provider cannot deploy them. These are contingent liabilities that may never crystallize, but they are included in the $30 billion estimate.
Precision is the only antidote to chaos. The accounting reality is that the vast majority of Nvidia's off-balance-sheet obligations are purchase commitments, not debt. They do not bear interest, do not require collateral, and do not represent a solvency risk. The real risk is operational: if AI demand growth slows from 200% to 50% year-over-year, Nvidia will have excess capacity that it must still pay for. This would compress gross margins from 75% to perhaps 60% — still enviable, but a sharp decline from current levels.
Contrarian
The bulls argue that these commitments are a sign of strength, not weakness. They are correct in one dimension: securing capacity in a supply-constrained market is a competitive moat. TSMC's CoWoS capacity is limited, and Nvidia has locked up ~60% of it through 2025. Any competitor trying to scale must now compete for the remaining 40% at higher prices. Nvidia's off-balance-sheet commitments are effectively a deposit on future monopoly rents.
But the contrarian angle misses the structural fragility. The commitments are asymmetric: they lock in costs, but not revenues. Nvidia's customers — hyperscalers like Microsoft, Meta, and Amazon — are also building their own AI chips. If Google's TPU or Amazon's Trainium achieve sufficient performance, these customers may reduce their Nvidia orders, leaving Nvidia holding the bag of prepaid wafers. The off-balance-sheet commitments are a bet that the hyperscalers will remain dependent on Nvidia's CUDA ecosystem for the next 3–5 years.
Logic survives the crash; emotion dissolves. The market's anxiety is not about the $30 billion itself — it's about the implied assumption that AI demand will grow exponentially forever. That assumption is untested. The last time a tech company made such a large bet on future demand was the fiber-optic buildout of the late 1990s, followed by the telecom bust. The difference is that Nvidia is generating $28 billion in free cash flow annually, meaning it can absorb a moderate downturn. But a severe one — a 50% drop in demand — would expose the $30 billion as a sunk cost.
Takeaway
The question is not whether Nvidia's off-balance-sheet commitments are a liability. They are not, in any accounting sense. The question is whether they represent a rational hedge or a leveraged bet on a self-fulfilling prophecy. If AI is the next internet, then Nvidia is buying the right to print money. If AI is a bubble, then these commitments accelerate the timeline of the crash. The market is pricing in the former. The cold dissector waits for the data to confirm.
