Nvidia's 'Circular Financing': The Revenue Is Real, the Risk Is Nvidia Becoming the AI Buildout's Underwriter

Jensen Huang spent part of this week at Goldman Sachs' Communacopia conference dismissing the accusation that Nvidia's AI investing is a feedback loop — "circular financing," in the industry's phrase. He laughed it off and called it how business has always been done: when one company contracts another, each looks at how much the other buys back, and if a relationship looks off-balance, someone pushes the other to spend more. It only attracts scrutiny, he said, because the dollar figures got so big.

The critique deserves a closer read than a shrug, because it points at something real. It just names the wrong thing.

Put plainly, the accusation runs like this. Nvidia finances and takes equity in the AI labs that buy its chips. OpenAI takes Nvidia money, spends it on Nvidia hardware, and Nvidia books the sale as revenue. Sceptics — some Goldman Sachs analysts and investor Michael Burry among them — have compared the setup to Cisco financing the cash-strapped telecom ICG Communications in the late 1990s: vendor loans that looked like demand until the customer collapsed. The version Nvidia tells in response is partly fair, partly a dodge. Separating the two is where the investing insight sits.

The $500 billion is really three loops

The headline figure — $500 billion, sometimes $750 billion — is not one number. It is three different capital structures running through the same node, and they carry very different risks.

The first is Nvidia's own equity. Nvidia has said it planned to invest up to $100 billion in OpenAI, and OpenAI's own February funding announcement recorded $30 billion from Nvidia at a $730 billion pre-money valuation. Nvidia's chief financial officer, Colette Kress, has put the company's total investments and commitments to AI companies at roughly $50 billion.

The second is balance-sheet backstopping. Nvidia agreed to provide a guarantee of up to $105 billion to help finance the Ohio data center OpenAI will lease, alongside a $1.5 billion investment in the project's developer, SoftBank's SB Energy.

The third structure is what dominates the totals — and it is not Nvidia's money at all. In August, Nvidia signed financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR designed to mobilize over $500 billion of institutional capital to buy Nvidia GPUs and data-center infrastructure, which AI companies then lease. Most of that capital comes from insurance and pension pools seeking long-duration assets. Nvidia has said it may back some of those vehicles with residual-value guarantees of up to 25% — effectively insurance that the chips behind the loans stay worth something.

The decomposition matters because it dissolves the strongest nervous version of the bear case. Nvidia is generating around $127 billion of free cash flow over the last twelve months against a roughly $50 billion figure for its own investments and commitments. Put differently: even if you zeroed out Nvidia's entire "circular" stack, it would not carry a quarter where total revenue roughly doubled to $96.2 billion and data-center revenue jumped 117% to $89 billion. That growth is being driven by real, third-party, final demand for compute — not by Nvidia lending its own money to customers and calling the loan proceeds revenue. The label "fake revenue" does not survive contact with the balance sheet.

Nvidia is becoming the underwriter, not the supplier

Which brings up the part Huang did not really address, and the part his own defense inadvertently concedes.

His strongest argument is structural, and it has real merit. Today's frontier labs are, as he says, the first generation of companies that need tens of billions of dollars before they turn a profit. They are not investment grade and lack the track record to borrow cheaply, so Nvidia fills a gap traditional lenders will not. And the hardware is arguably not a sunk cost: if one customer stumbles, the chips can be redeployed to another workload.

But that redeployability argument cuts both ways. The chip inventory is not Nvidia's exposure alone; it is the collateral behind a growing wall of third-party debt whose buyers — pension and insurance pools — assumed a low-risk, long-duration asset. Analysts have noted the residual-value guarantee functions as a quiet price cut, Nvidia using its own profits to reduce customers' cost of capital. When lots of companies wobble at the same time, "redeployable" compute stops being Nvidia's escape hatch and becomes everyone's negative collateral at once, and the guarantee is a liability Nvidia would owe precisely when the AI companies cannot pay.

That is the shift hiding inside the "circular" scare: Nvidia is sliding from chip vendor — sell the hardware, take the cash, the customer owns the downside — to underwriter of the whole buildout, holding equity in unprofitable labs and backstopping the debt that buys its own product. The market seems to sense it even if the language is wrong. Nvidia's stock was up only about 12% for the year even as revenue more than doubled, and the July announcement of a fresh round of deals worth over $750 billion triggered a 5% single-day drop that cost it the title of world's most valuable company. A few data points do not settle a debate, but they are consistent with a market that has stopped paying for the growth and started pricing the tail risk.

Where the judgment lands

The most useful way to read Huang's dismissal is as a defense of the wrong charge. The revenue behind the doubling is real and is not being manufactured by Nvidia's own loans. But the critique that the AI buildout is increasingly financed on a circular — and, per Goldman Sachs, a quarter of all US investment-grade issuance now AI-related — basis is closer to true than he lets on, and third-party debt needs to keep finding buyers for the mechanism to hold.

So the operator's question is not "is Nvidia's growth fake?" It is: how much of Nvidia's edge is being spent underwriting the demand for its own chips, and what happens to that equity and that 25% guarantee if the institutional capital door closes at the same time the lab tenants hit a wall? That is the live risk the "circular financing" argument gestures at but fails to name — and it is the one worth watching, because it is where the model stops looking like a great vendor story and starts looking like a leveraged one.