Nvidia is underwriting its own demand, and that should worry you
A reported $100 billion credit guarantee for OpenAI data centres turns the largest chip vendor in the world into the lender of last resort for its biggest customer. We have seen this shape before.
Nvidia is reportedly in talks to guarantee roughly $100 billion of credit for OpenAI’s data centre build-out. Separately, it discussed putting about $3 billion into SB Energy, a power developer. Neither of those is a chip sale. Both are Nvidia paying to make sure chip sales happen.
That is the story worth sitting with. The most profitable semiconductor company in history has decided that the binding constraint on its revenue is no longer its ability to fabricate accelerators. It is its customers’ ability to finance them, and the grid’s ability to power them. So it is buying its way past both.
The mechanism has a name
When a supplier lends a customer the money to buy the supplier’s product, that is vendor financing. It is not inherently fraudulent or even unusual — GE Capital did it for turbines, Boeing does versions of it for aircraft, and every equipment leasing desk on earth exists because of it. It is a legitimate answer to a real problem: the buyer’s cash flows arrive years after the capital expenditure does.
What makes vendor financing dangerous is what it does to the seller’s reported numbers. Revenue recognised today is funded by a liability the seller guaranteed. If the customer’s business works, the loan performs and everybody is a genius. If it does not, the seller has booked the revenue and owns the credit loss — the same dollar of optimism counted twice, in opposite directions.
Telecoms ran this exact play into the ground between 1998 and 2001. Lucent, Nortel and Motorola lent aggressively to competitive carriers who used the money to buy switching gear. The gear revenue was real. The carriers were not. When the buyers failed, the vendors discovered they had been booking their own subsidies as growth.
The $3 trillion nobody has to show you
A Wall Street Journal analysis put the number in perspective: nine large technology companies are now carrying roughly $3 trillion in AI-related commitments — data centre leases, chip purchase agreements, financing arrangements — that do not appear as conventional balance-sheet debt.
Read that carefully, because the word doing the work is commitments. These are contractual obligations to spend. They are disclosed, mostly in footnotes, mostly as operating leases and purchase obligations. They are not hidden in a criminal sense. They are simply not where a reflexive glance at a debt-to-equity ratio would find them.
- Data centre leases — long-dated, often 15 years, frequently on facilities purpose-built for a single tenant. Hard to re-let if the tenant walks.
- Chip purchase agreements — take-or-pay commitments for silicon that depreciates on a two-to-three year curve whether it is utilised or not.
- Financing arrangements — the category Nvidia’s reported guarantee lands in, where the obligation is contingent and therefore easiest to under-weight.
None of that is a prediction of collapse. Plenty of it will be repaid out of genuinely enormous AI revenue. The point is narrower and harder to dismiss: the sector’s leverage is real, it is large, and it is arranged so that the standard metrics understate it.
The tell to watch is power, not chips
The Guardian reported that Microsoft holds roughly 2.2 million AI chips globally against about $280 billion in infrastructure spending since 2022, and Microsoft itself has been unusually direct that the bottleneck is power availability and completed shell capacity — not chip supply.
That reframes everything. If accelerators were scarce, you would expect vendors to ration and prices to rise. Instead the largest buyer says it has chips it cannot energise. Nvidia’s move into a power developer’s cap table is the rational response to that, and it is also an admission: shipping more silicon into a grid-constrained market does not convert to revenue.
A chip you cannot plug in is inventory. A chip you financed for a customer who cannot plug it in is a receivable.
The distinction the next four quarters will litigate
What would change the read
Three things, in order of how much they would matter:
- Whether the guarantee is confirmed and quantified in filings rather than press reports. Contingent obligations of this size eventually surface in a 10-Q, and the language will be more informative than the leak.
- Whether OpenAI’s revenue growth outruns its committed spend. Vendor financing is only a problem if the borrower cannot service it.
- Whether grid interconnection queues shorten. This is the unglamorous variable that decides whether $3 trillion of commitments become assets or write-downs, and it moves on utility timescales — years, not quarters.
The bull case and the bear case agree on the facts here. They disagree only about whether the demand Nvidia is underwriting was going to show up anyway.