The Guarantor of Last Resort

Nvidia’s own filings show it has become investor, guarantor, supplier and backup customer in the same deals. What happens when the guarantor says no?

Article related image
iStock.com
Author
Srinath Sridharan

Dr. Srinath Sridharan is a Corporate Advisor & Independent Director on Corporate Boards. He is the author of ‘Family and Dhanda’.

Author
Arjun Raghavendra M.

Arjun Raghavendra M., a former IRS officer, practices law in Supreme Court

November 16, 2026 at 8:14 AM IST

Set aside why Jensen Huang left and read Nvidia’s own disclosures. Its August quarterly commentary lists $366 billion of supplier commitments, $56 billion to secure land, power and capacity for customers who commit to its chips, and up to $108.5 billion of credit guarantees. Of the guarantees, $105 billion backs a 4.25-gigawatt campus in Ohio that will exclusively host Nvidia hardware under 20-year leases to OpenAI, with an option to backstop a further 3.8 gigawatts. Each hardware generation at that site, the filing says, could mean $150–200 billion in Nvidia revenue.

AI labs will supply about a quarter of Nvidia’s business next year, and its equity portfolio, the same filing says, is focused on AI model makers. For smaller cloud operators, Nvidia commits to pay for a minimum share of a data centre’s capacity, the floor their lenders need, and takes a share of the revenue above it. Add the $1.5 billion it is putting into the Ohio developer and the circuit closes. One company is investor, guarantor, exclusive supplier and customer of last resort. Money leaves as credit support and returns as chip sales. Colette Kress, then finance chief and now interim chief executive, said it herself: “We recognise the scale of this support, and we know some will call this circular financing. We see it differently.”

She also, by the Wall Street Journal’s account, fought parts of it. The original Ohio proposal was a $250 billion guarantee; Kress questioned how such commitments were valued, pushed back on a programme that Huang’s son-in-law, Nico Caprez, was reportedly driving, and cut the backstop by roughly half. Now she runs the company, and on Tuesday she presents her first results as its chief. The question markets must price is not the scandal but the succession: what happens to an industry built on a supplier’s balance sheet when the supplier is run by the person who doubted it?

Nvidia can afford every guarantee it has written. Even after issuing $25 billion of notes last quarter, it holds $57 billion in cash and securities, earns gross margins around 75% and generates over $20 billion of free cash flow a quarter, against a cap that phases in from fiscal 2029 and declines as OpenAI pays its leases. The chips, Nvidia says, are fungible and can be redeployed if a partner falters. True of the hardware; not of the debt that bought it. That is why the Lucent comparison is wrong, and the danger worse. In 1999–2001, Lucent and Nortel lent telecoms the money to buy their equipment and booked the sales; Lucent never recovered and Nortel failed, because the vendor could not survive its customers’ default. Nvidia can. The vendor will survive withdrawing its credit. The customers it created will not, nor will the lenders, landlords and utilities who backed them on the strength of Nvidia’s name. In 2008 the guarantor of last resort was AIG, which wrote protection on mortgage securities until the protection was worth more than the company. Vendor financing is a confession dressed as a strategy. It says the end customer’s demand is not yet real, so the vendor will manufacture it with credit.

The chain of consequence is short. A finance-minded chief executive declines to extend new guarantees. The neoclouds, operators whose assets are debt-financed hardware and whose creditworthiness rests on their supplier’s floor, cannot refinance. A quarter of next year’s business, by Nvidia’s own count, comes from the AI labs those operators were built to serve; without the floor, the “avalanche” of committed supply Huang reportedly feared arrives by another route. Everything after that, from data-centre debt to index funds to half-built substations, is the system’s story. Nvidia’s part is the first link, and the first link is now one person’s decision.

The hardest consequence is governance, noticed last because it falls first. The lesson of AIG was that a guarantee is a liability whether or not it sits on the balance sheet. Nvidia’s disclosures are explicit about the first-order kind: the cap, the phases and the table are all in the filing. No filing can disclose the second-order liability: an industry whose marginal demand exists because one supplier underwrote it. When that underwriting is withdrawn, the write-down is not Nvidia’s. It is everyone else’s.

There is one consolation, and it is for users, not producers. A compute glut means cheap inference first: ruinous for businesses priced on selling AI at today’s rates, a gift to every business that merely uses it. The fibre-optic glut of 2002 built the cheap internet of 2005. Consolidation and dearer compute may follow, but the first act of a sustenance crisis is a solvency crisis for sellers and a windfall for buyers.

For India the exposure is indirect and the opportunity direct. Much of the capital that left Indian equities this year would seek a home if the AI trade reversed, and India is the obvious one, if its house is in order to receive it. That is a Budget question. Tuesday’s is simpler: does the guarantor still guarantee? Markets reward the first to measure. They punish the last to notice.