On August 10, Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build compute-financing platforms — a structure intended to mobilize more than $500 billion in third-party capital for AI infrastructure. Nvidia may backstop up to roughly a quarter of it, about $125 billion.
That commitment would not appear as debt on Nvidia’s balance sheet.
Read that sentence again, because it is the entire story. The largest seller of AI chips is now standing behind the borrowing of the people who buy AI chips, in a form that does not register as borrowing.
Why anyone needed to do this
The scale first, because the scale is the reason.
The five dominant American hyperscalers — Amazon, Alphabet, Meta, Microsoft, Oracle — are expected to spend roughly $800 billion on capital expenditure in 2026, and close to $1.1 trillion in 2027. Global data-center investment is on pace for about $2.9 trillion through 2028.
Big Tech’s own operating cash flow, enormous as it is, covers only about $1.4 trillion of that. The remaining $1.5 trillion has to come from credit markets.
It is already arriving. Nearly $500 billion in AI-related debt has been issued so far this year. In 2025 the five hyperscalers issued a record $121 billion — more than four times their average annual borrowing between 2020 and 2024. Morgan Stanley expects private credit to supply the largest share of what remains, roughly $800 billion, with another $200 billion in corporate bonds and $150 billion in securitized products.
This is the part that gets lost in the argument about whether AI is a bubble. Whatever you believe about the technology, the buildout has quietly stopped being financed out of profits and started being financed out of debt. Those are different regimes with different failure modes.
Vendor financing has a history
What Nvidia is doing has a name and a track record.
In the late 1990s, Lucent Technologies and Nortel Networks sold telecom equipment into a fiber-optic buildout that looked, at the time, like an obvious and permanent expansion of demand. To move product faster, both companies began lending money to their own customers — small carriers who could not otherwise afford the equipment.
For two years it worked beautifully. Revenue climbed. The customers bought more. The lending was booked as an asset.
Then the carriers failed, because the demand they had been built to serve did not arrive on schedule. Lucent and Nortel were left holding both the receivables and the reputational wreckage, and neither company ever recovered its former position. The lesson was not that fiber was worthless — we use that fiber today. The lesson was that when a seller finances its own buyers, reported demand and real demand stop being the same number, and nobody can tell how far apart they’ve drifted until the financing stops.
Nvidia’s structure is more sophisticated than Lucent’s. Third-party capital sits between Nvidia and the borrower. The backstop is partial. The counterparties are the largest asset managers on earth rather than startups.
But the underlying shape is the same: the company that benefits most from chips being bought is now helping to ensure chips get bought.
Minsky’s ladder
Hyman Minsky argued that financing arrangements fall into three types, and that economies tend to migrate down the list during good times without noticing.
Hedge finance: cash flow covers interest and principal. The borrower is fine regardless of what markets do.
Speculative finance: cash flow covers interest but not principal. The borrower must refinance, which means they are fine only as long as credit stays available.
Ponzi finance: cash flow covers neither. The position survives only if the underlying asset keeps appreciating.
Minsky’s point was that stability itself causes the migration. When nothing has gone wrong for a while, hedge financing looks needlessly conservative, and the returns go to whoever moves one rung down. The system becomes fragile not through recklessness but through the ordinary competitive logic of a long calm.
The AI buildout began as hedge finance — Microsoft and Google paying for data centers out of enormous operating profits. The $1.5 trillion gap is the sound of it moving to the second rung. Whether it reaches the third depends entirely on a question nobody can currently answer.
The honest counterargument
I want to be careful here, because the doom version of this story is lazy and the numbers don’t fully support it.
The revenue is real and it is growing fast. AWS is running at roughly $150 billion annualized, up 28% year over year. Google Cloud is near $80 billion, up 63%. Azure’s AI business is at a $37 billion run rate, up 123%. Those are not vapor. Q2 earnings were extraordinary — 86% of S&P 500 companies beat consensus, and blended earnings growth came in at 47.4% against a 23.3% expectation.
The bear case rests on a gap, not an absence. Sequoia’s David Cahn has estimated that roughly $600 billion in annual revenue is needed to justify the current spend, and that the gap is widening rather than closing. But “widening gap” and “no revenue” are very different claims, and only the first one is true.
There is also a real possibility that this ends the way the fiber buildout ended: the capital gets destroyed, the investors lose, and the infrastructure turns out to be enormously valuable to whoever buys it in the wreckage. That is not a happy outcome for the lenders. It is a fine outcome for the technology.
What to actually watch
Not the chip sales. Sales are the number most contaminated by the financing.
Watch whether private credit keeps showing up. The $800 billion Morgan Stanley expects from private credit is the load-bearing assumption in the entire structure, and private credit is the least transparent, least regulated, most recently expanded part of the financial system. A Treasury draft report has already drawn the dot-com comparison and flagged systemic risk if the cycle reverses.
Watch how much of the obligation sits off balance sheets. Nvidia’s $125 billion backstop is the visible instance. It is unlikely to be the only one.
And watch the spread between announced capex and cash-funded capex. That gap is the cleanest available measure of how far down Minsky’s ladder the industry has traveled.
The chips are real. The data centers are real. The revenue is real and growing.
The question is who is holding the paper, and whether they know what they’re holding.
This article first appeared in The Interregnum at nishanv.substack.com.
Original article
Comments