The Story
Nvidia's arrangements to underwrite the AI cloud providers that buy its chips are being tested for the first time, as shares in CoreWeave, Nebius and IREN each fell at least 30 per cent in October.
Nvidia disclosed roughly $36 billion of off-balance-sheet AI cloud agreements and backstops in its July quarterly filing. It has invested more than $7 billion directly in AI cloud providers and guaranteed up to $105 billion tied to OpenAI's proposed 10 gigawatt facility in Ohio, a guarantee that terminates if OpenAI provides collateral or secures a sufficiently strong credit rating. Total exposure has been reported at between roughly $164 billion and $172 billion depending on what is counted.
The CoreWeave arrangement is the most direct. Nvidia, which holds an equity stake, has committed to purchase up to $6.3 billion of CoreWeave compute through April 2032 if the company cannot sell it to anyone else. With Sharon AI and other partners, the deals involve revenue-sharing structures.
CoreWeave's five-year credit default swap spread has moved past 11 per cent to 1,115 basis points, up 171 per cent this year and 36 per cent in the past month. Its unsecured notes due 2031 and 2032 traded below 81 cents on the dollar in mid-October. The company's market value has fallen to around $31 billion from a peak near $90 billion, with debt above $29 billion.
Nebius fell even after announcing a $3.5 billion financing facility from Citi, and has reportedly declined to participate in the guarantee programme. Jensen Huang said this week that CoreWeave does not need Nvidia's backstop and that its cost of capital is decreasing.
Nvidia shares recorded a third consecutive weekly decline, leaving the month down 7.6 per cent. Most of the credit backstops have been put in place over the past six to twelve months.
Why It Matters
Follow the CoreWeave commitment through to its conclusion and the accounting gets strange.
Nvidia sells CoreWeave the GPUs. CoreWeave borrows to pay for them, racks them, and sells the compute. If CoreWeave cannot find a buyer for that compute, Nvidia has agreed to buy it, up to $6.3 billion, through April 2032. So Nvidia books revenue on the chips, and then stands behind the revenue the chips are supposed to generate. Should the guarantee be drawn, Nvidia pays for output produced by hardware it has already been paid for.
The more consequential effect is on who is actually being lent to. When a bank finances CoreWeave's GPU purchases against a contract that Nvidia has underwritten, it is partly assessing Nvidia's credit rather than CoreWeave's. That is how a company with over $29 billion of debt and an interest bill exceeding its adjusted operating income raised money at all. It is also why a widening spread on CoreWeave now reads directly as a question about Nvidia.
Huang's position is that none of this will be needed, and that CoreWeave's cost of capital is falling. The credit market disagrees in a measurable way: a five-year CDS at 1,115 basis points is the price of insuring against exactly the event he says will not happen, and it has risen 171 per cent this year. One of those two assessments is wrong, and the CDS is the one with money behind it.
The detail that complicates the simple bearish reading is Nebius. It reportedly declined the guarantee programme, secured $3.5 billion from Citi, and fell anyway. If the companies outside the backstop are being marked down alongside the ones inside it, then what is being repriced is not Nvidia's underwriting. It is the assumption that renting out GPUs is a good business.
The Strategic Read
The structure has a precedent and it is not a comfortable one.
In 1999 and 2000, Lucent, Nortel and Motorola lent money to telecom carriers so those carriers could buy their equipment. The sales were real, the revenue was booked, and the loans were on the balance sheet. When the carriers could not pay, the vendors discovered they had financed their own order books. Lucent's vendor financing portfolio reached around $7 billion before it was written down.
The differences matter and should be stated. Nvidia's commitments are largely guarantees rather than loans, and a guarantee only costs money if it is drawn. The underlying demand for compute is real in a way that dark fibre in 2001 was not, and CoreWeave has contracted backlog with creditworthy counterparties. Nvidia also has the strongest balance sheet in the industry and can absorb a great deal.
What is similar is the circularity, and the fact that it is concentrated in the final stretch of a boom. These backstops were mostly written in the past six to twelve months, which is to say after the easy financing had gone and the harder financing needed help.
For India the relevance is immediate. CoreWeave has a 240 MW buildout here, and domestic data centre projects are being financed on the same logic of long contracts against depreciating hardware. If the cost of that capital is being repriced globally, the projects announced at one rate get built at another, or do not get built.
For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.