The timestamp is 14:00 UTC. The ledger shows a $1.2 billion debt issuance by CoreWeave, secured against NVIDIA H100 GPUs. The same GPUs that NVIDIA sold to CoreWeave. The same GPUs that CoreWeave then leases back to AI startups. The same AI startups that burn cash to buy compute from CoreWeave, which then repays its debt to NVIDIA-affiliated funds. The circle is closed. But the point of entry is missing.
Ed Zitron, CEO of EZ Primary Research, told CNBC that NVIDIA is not just a supplier. It is a customer, a financier, and a credit enhancer. He called it 'lending its credit.' I call it a recursive loop that has no fundamental exit. The ledger does not lie, only the storytellers do. And the story of AI compute demand is a story told by the same people who issue the loans.
Context: The Three Hats of NVIDIA
NVIDIA designs and manufactures GPUs. That is hat one. It then sells these GPUs to infrastructure providers like CoreWeave, Lambda, and Crusoe Cloud. These providers build data centers and lease compute capacity to AI companies. That is hat two: NVIDIA is an indirect lessor of compute. But the critical third hat is the one that few auditors examine. NVIDIA, through its venture arm and partnerships with financial institutions, helps these infrastructure providers secure debt financing. The debt is often collateralized by the very GPUs that NVIDIA sold. The lenders look at the long-term contracts between infrastructure providers and AI companies as proof of cash flow. But those contracts are with a handful of cash-burning firms: OpenAI, Anthropic, Inflection, and a few others.
Zitron’s concern is that the demand within this system is highly concentrated. He is right. I have spent the last six years auditing on-chain lending protocols and crypto mining balance sheets. The pattern is identical. The same metrics that preceded the 2022 crypto lending collapse—concentration of borrower, illiquid collateral, and recursive financing—are now visible in the AI compute supply chain.
Core: The On-Chain Evidence Chain (Translated to Off-Chain But Analogous)
Let me walk through the data, not the headlines. First, the debt structure. CoreWeave has raised over $6 billion in debt financing since 2023. The majority of this debt is secured by NVIDIA GPUs. The valuation of those GPUs is based on their resale value and their earning potential. The resale value is determined by NVIDIA’s pricing power. The earning potential is determined by the lease rates that AI companies pay. Both are endogenous to the same narrative.
Second, the concentration of demand. A breakdown of CoreWeave’s customer base shows that its top three customers account for over 70% of its revenue. Those customers are OpenAI, Anthropic, and a third unnamed entity. OpenAI and Anthropic have not proven profitability. They are burning cash at a combined rate of over $5 billion per year. Their revenue is a fraction of that. The risk is not whether AI will transform the world. The risk is whether the current capital structure can survive a period of slower-than-expected adoption.
Third, the financing mechanism. NVIDIA does not lend directly. It provides purchase commitments. It invests in the equity of infrastructure providers. It arranges debt facilities through its network of banks. This is what Zitron calls 'lending its credit.' In crypto terms, it is a form of rehypothecation. NVIDIA’s balance sheet is used as a backstop to issue debt that is then used to buy NVIDIA’s own products. The total amount of this indirect leverage is not publicly disclosed. But based on the debt issuance of the top five GPU infrastructure providers, the figure is likely between $15 billion and $20 billion. That is a significant fraction of NVIDIA’s annual revenue.
Precision is the only hedge against chaos. Let me be precise: if the AI demand growth rate slows from 100% year-over-year to 30% year-over-year, the lease rates on those GPUs will drop. The collateral value of the GPUs will drop. The debt covenants will be triggered. The lenders will demand repayment. The infrastructure providers will have to sell GPUs into a market where NVIDIA is the only buyer. NVIDIA will be forced to either absorb the supply or let the system collapse. The history of crypto mining shows that mining hardware manufacturers do not step in to save leveraged miners. They step in to buy back hardware at fire-sale prices. The same pattern is forming.
Contrarian: The Correlation Fallacy of Demand
The common narrative is that AI compute demand is infinite. The data says otherwise. The total addressable market for AI training and inference is large, but the current demand is artificially inflated by a few players who are spending money they do not have. This is not a sustainable demand curve. It is a demand curve that is financed by the same entity that supplies the product. This is a textbook case of circular causation.
Zitron’s argument is that the concentration of demand among a few unprofitable companies is the weak link. I agree, but I would go further. The weak link is the assumption that the debt can be repaid without a continuous inflow of new capital. In crypto, we called this the 'ponzinomics' model. The investors who bought the debt of CoreWeave are effectively betting that OpenAI will raise more money from SoftBank or Microsoft. That is a bet on the ability of a few private companies to continue raising capital at increasingly large valuations. History repeats, but the code changes the rhythm. The code here is the debt structure. The rhythm is the quarterly refinancing cycle.
A counterargument is that NVIDIA’s GPUs have intrinsic value outside of AI. They are used for graphics, scientific computing, and cryptocurrency mining. The resale value provides a floor. But the price of a GPU is not fixed. The price of an H100 has already dropped from $40,000 to $30,000 in the secondary market as supply has increased. If the debt collateral is marked to market, the lenders will take losses. The question is not whether the system will fail. The question is how much of the failure will be absorbed by NVIDIA’s balance sheet versus the lenders.
Takeaway: The Signal to Watch
Next week, CoreWeave is expected to announce its first-ever quarterly earnings as a public company. The key metric is not revenue. It is the average utilization rate of its GPU fleet and the average lease rate. If utilization falls below 70% or lease rates decline by more than 10%, the debt covenants will be in danger. I will be watching the 10-Q filing for any mention of 'collateral valuation adjustments' or 'impairment charges.' The ledger does not lie. The auditors will eventually have to write down the assets.
The takeaway for crypto investors is this: the same dynamics that destroyed the 2022 lending market are now playing out in the AI compute supply chain. The collateral is physical, but the financing is digital. The leverage is hidden. The demand is concentrated. The exit liquidity is not a feature. It is a risk. Do not mistake NVIDIA’s credit for your own safety.
I follow the bytes, not the headlines. The bytes here show a recursive loop. The loop will break. The only question is when.