Nvidia just posted a quarter that would have looked absurd for a chip company a few years ago. Revenue hit $96.2 billion, up 106% on the year, according to its own results, with data centre alone contributing $89 billion, gross margin holding at 75%, and guidance for this quarter above $108 billion. “AI has reached its inflection point,” Jensen Huang told investors, adding that “compute is revenue.” On the surface the story is simple: demand is enormous, Nvidia owns the picks and shovels, so Nvidia prints money.
That part is true. But the more interesting story sits underneath, because this quarter is the clearest evidence yet that Nvidia is becoming something stranger than a chipmaker. It is turning into a piece of the financial system standing behind its own customers.
First, how rich is Nvidia?
You cannot understand the financing story without the size of the engine funding it. Data centre is now roughly 93% of revenue, split between hyperscalers at about $48.7 billion and a bucket Nvidia calls AI clouds, industrial and enterprise, or ACIE, at about $40.3 billion, the latter up a remarkable 138% year on year. Operating income ran past $63 billion, and Nvidia still handed shareholders around $25.8 billion in buybacks and dividends, slightly more than the roughly $21 billion of free cash flow it generated.
This explains the asymmetry underneath everything else. Microsoft, Amazon, Google and Meta can self-fund data centres, raise cheap debt and sign decade-long power deals on their own credit. A smaller AI startup or neocloud cannot. It may have real demand, but an AI factory swallows billions in land, gigawatts of power, buildings, cooling and tens of thousands of GPUs long before the revenue lands. Nvidia says so itself, noting that many AI clouds and model builders cannot line up the long-dated infrastructure deals and investment-grade financing that hyperscalers take for granted. The bottleneck has moved. Nvidia no longer just has to make chips fast enough; it has to worry whether the people who want them can afford to plug them in.
What “financing the boom” actually looks like
Nvidia’s answer is to spread its balance sheet across the ecosystem in layers. The lightest is giving big, investment-grade customers longer to pay, which is why accounts receivable ballooned from $38.5 billion in January to $63.1 billion by the end of Q2. It gets more structural from there. Nvidia holds around $36 billion of commitments to buy cloud capacity from AI-cloud partners, giving them enough revenue visibility to raise their own financing; the CoreWeave deal, where it agreed to buy up to $6.3 billion of unsold capacity, is the template. It backstops land, power and data-centre obligations, and in the biggest case has signed guarantees capped at as much as $105 billion behind an OpenAI-anchored buildout in Ohio.
The boldest move is newest. In August, Nvidia announced a platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500 billion of third-party capital for AI infrastructure. Nvidia is not putting up that money; the financiers underwrite it while Nvidia supplies technology and, in some deals, credit support such as residual-value guarantees of up to 25% of an opportunity. Above all this sits an equity portfolio worth roughly $99 billion, spanning stakes in the very companies that buy its chips, the layer that draws the “circular financing” charge Huang keeps rejecting by arguing the underlying demand is genuine and the outside capital never becomes Nvidia’s revenue.
Morgan Stanley’s “$200 billion” needs reading carefully
This is where headlines go wrong. Morgan Stanley opened credit coverage on Nvidia for the first time in late August with a Neutral rating, arguing conventional leverage badly understates the picture because so much ecosystem support now lives in contingent, contractual and off-balance-sheet forms. Its eye-catching figure, credit exposure nearing $200 billion by end-2028, has about $170 billion of it contingent, off-balance-sheet exposure, and it is an analyst model built on assumptions about programmes still taking shape. It is not $200 billion of debt Nvidia has disclosed.
The same discipline applies to every scary number here. The $105 billion of OpenAI-related guarantees phase in over years as nine data centres come online, shrink as leases are paid, and carry an OpenAI commitment to reimburse Nvidia. S&P, which reaffirmed Nvidia at AA with a stable outlook, does not treat it as $105 billion of debt either; it models the slice of exposure it thinks Nvidia realistically carries, an adjustment starting around $4.2 billion in FY2029 and building toward the high thirties of billions by FY2032 before receding. The honest read sits between two lazy extremes: neither nothing because it is off the balance sheet, nor a liability already taken on. It is real contingent exposure whose cost depends on timing, asset values and how customers perform.
The bull case: a flywheel and a moat
Play the tape forward optimistically and this is one of the smartest moves a dominant company has ever made. A dollar of guarantee or residual-value support can unlock many dollars of hardware demand Nvidia would otherwise never book, which is leverage in the economic sense without showing up as debt. If the infrastructure it helps finance goes on to generate cash, Nvidia can capture recurring, usage-linked revenue on top of the original sale, turning a one-off transaction into something closer to an annuity. And by wiring financing, land, power, cloud distribution and capital around its own platform, it forces rivals to compete not just on silicon but against an entire ecosystem. The presence of independent underwriters like Apollo and BlackRock, who make money when borrowers repay rather than when Nvidia sells chips, even adds a layer of external discipline a self-funded version would lack.
The bear case: wrong-way risk
Now play it the other way, because the danger is not the size of any single obligation. It is correlation. Each exposure looks survivable alone, but they share one driver: whether AI actually monetises as expected. If end-user AI spending slows, utilisation falls, AI-cloud revenue drops, weaker neoclouds strain, GPU resale values soften, and the guarantees and residual-value backstops all turn relevant at once, exactly when Nvidia’s own sales and equity portfolio would also be under pressure. Analysts call insurance that pays out when you can least afford it wrong-way risk. Reuters Breakingviews framed the same dynamic as a potential “bad-news spiral,” and the Wall Street Journal has taken to calling Nvidia a “banker to the AI boom,” on the view that engineering deployed to keep customers buying can turn painful if the cycle reverses. Nvidia’s own history is a warning: it took a roughly $4.5 billion charge on unsold H20 chips and the purchase commitments behind them when export rules gutted that product’s demand. A commitment is not debt, but if the forecast behind it is wrong, the cost becomes real anyway.
It’s not Enron, and it’s not a normal chipmaker
Two things are true at once. There is no evidence in the filings that Nvidia is manufacturing revenue by booking future orders as current sales; on the contrary, it took in $15.6 billion of customer advances in the first half, meaning some customers pay ahead. And there is no basis for stacking every lease, guarantee, commitment and investment into one number and calling the company insolvent. Nvidia is very far from distress, with tens of billions in liquidity and a business still growing north of 100%. But waving all of this away misses a genuine change. Nvidia is now, simultaneously, the supplier, investor, creditor, capacity buyer, guarantor and potential revenue-sharing partner of the same ecosystem, which is a different economic animal from a company that only sells accelerators.
The upshot for anyone watching the stock is that the question has quietly changed. It is no longer just how many GPUs Nvidia sells next quarter. It is how much of the demand in front of it can finance itself out of real cash flow, and how much needs Nvidia to keep lending its balance sheet to make the deal happen. The numbers this quarter do not show a system breaking. They show a company that has waded far enough into finance that you can no longer analyse it as an ordinary chipmaker.
FAQs
Is Nvidia in financial trouble?
No. S&P rates it AA with a stable outlook, leverage is roughly 0.4x, and the core business generated tens of billions in cash last quarter. The concern is the direction of travel, not current solvency.
What is “balance-sheet-as-a-service”?
It is Morgan Stanley’s term for Nvidia using its financial strength to help finance the AI ecosystem, through guarantees, credit support, cloud commitments and residual-value backstops, rather than only selling hardware.
Is the $200 billion, $105 billion or $500 billion actual Nvidia debt?
No. The $200 billion is an analyst estimate of contingent exposure by 2028, the $105 billion is a phased guarantee cap that shrinks over time, and the $500 billion is third-party capital Nvidia helps arrange but does not fund itself.
Is this “circular financing”?
Critics argue Nvidia invests in customers who then buy its chips. Huang counters that the demand behind it is genuine and that the outside money never counts as Nvidia’s revenue. The honest answer depends on whether end-user AI demand holds up.
Why does Nvidia finance its own customers?
Because the bottleneck has shifted from making chips to whether smaller AI clouds and startups can raise the capital, land and power to deploy them. Nvidia uses its balance sheet to remove that constraint and unlock more hardware sales.