Nvidia has committed more than forty billion dollars to equity stakes in AI companies this year, led by a thirty-billion-dollar investment in OpenAI, with billions more into the glassmaker Corning, the data-center operator IREN, and roughly two dozen private startups — atop sixty-seven such deals the year before. The pattern has a name its critics use without embarrassment: circular financing. Nvidia invests in a company; the company spends the money on Nvidia’s chips; the purchase becomes Nvidia’s revenue; the revenue funds the next investment. The same week, an exchange announced it would create a futures market for AI computing power, converting the GPU-hour into a tradable financial instrument. Together these two facts pose the question sitting underneath the entire boom and carefully kept beneath it: how much of the demand for this technology is real, and how much is the world’s largest chipmaker paying its own customers to generate.
The Shape of the Loop
Trace a single dollar through the structure and the concern becomes legible. Nvidia takes a dollar of its profit and invests it in a company — frequently one of the “neoclouds,” firms whose business is to buy Nvidia chips at scale and rent them to hyperscalers and model builders. The company uses the dollar to purchase a Nvidia chip. The dollar returns to Nvidia, now recorded as revenue rather than investment, and revenue is the number that sets Nvidia’s valuation, which is the number that gives Nvidia the market value to invest the next dollar. The dollar has completed a circuit and arrived back where it started, having changed its name twice — from investment to purchase to revenue — and increased, at each renaming, the appearance that the world is demanding what Nvidia sells.
A defender will say, correctly, that this is not necessarily improper. Nvidia genuinely believes in the companies it funds; the companies genuinely want the chips; the demand is, in large part, real. An analyst quoted this week suggested the investments could build a durable competitive moat if the bets succeed, and that is a coherent reading. But the defense and the concern are not in conflict, because the concern is not that the demand is fake. The concern is that a closed loop cannot tell you what fraction of the demand is real, and removes the instrument you would otherwise use to find out. When the seller supplies the buyer’s money, the sale stops being evidence that the product was wanted. It becomes evidence only that the seller wanted the sale.
This is the property that distinguishes circular financing from ordinary growth, and it is epistemic before it is financial. In a normal market, a purchase carries information: someone wanted the thing enough to pay for it with money they earned somewhere else, and the payment is an independent signal of value. The circle severs that signal. A purchase made with the seller’s own capital, returned through a customer’s balance sheet, looks identical to an arm’s-length sale on every financial statement, and validates the same valuations, and justifies the same expansion — while carrying none of the information the arm’s-length sale would have carried. The numbers grow. What the numbers mean becomes unknowable.
Compute as an Asset Class
The futures market is the accelerant, and it deserves more attention than it received. Once computing power can be traded as a standardized financial instrument — bought, sold, and speculated upon independent of whether any model ever runs on it — the GPU-hour begins its drift from a tool into an asset class. A tool is wanted for what it does. An asset class is wanted for what someone else will pay for it next, and its price can detach entirely from its use, sustained by belief, leverage, and the expectation of resale. The futures market lets capital take a position on compute without ever intending to compute anything, which means the demand curve for chips will soon include a component that has nothing to do with running models and everything to do with betting on the price of the right to run them.
This is the same structural fragility I described in the buildout with no clear end, now equipped with a faster transmission mechanism. The capital expenditure was already a bet whose returns had been deferred indefinitely; the circular financing makes the bet partly self-funded and therefore self-validating; and the futures market makes the underlying asset tradable on belief, which is the final ingredient. Each layer increases the distance between the price of compute and the demand for what compute produces. Each layer makes the boom more legible as a financial structure and less legible as a measurement of how much the world actually wants the intelligence at the end of the chain.
The historical rhyme is exact and unflattering. Every infrastructure mania has eventually discovered a way to finance its own demand — vendor financing in the telecom buildout, where the equipment makers lent their customers the money to buy the equipment, and the loans counted as sales until the customers could not repay them and the sales reversed all at once. The circle is vendor financing in equity form, more elegant and harder to see, because an investment is not a loan and does not appear on the balance sheet as a liability waiting to default. It appears as an asset, growing, right up to the moment the capital stops circulating.
What This Means
It would be lazy to call the circle fraud, and inaccurate. There is no deception required; every transaction is real, disclosed, and defensible on its own terms. Nvidia is not faking sales. It is funding the conditions under which real sales occur, which is a different thing and a legal one, and possibly even a wise one if the demand it is priming turns out to be as large as the priming assumes. The circle is not a lie. It is a structure that makes the truth unverifiable, and those are not the same accusation.
The unverifiability is the entire danger. The boom may be completely justified — the intelligence at the end of the chain may be worth every dollar that circulated to build it, and the circular financing may prove to have been the rational priming of a genuinely enormous market. But the structure has removed the means of knowing that in advance. It has replaced the independent signal of demand with a self-referential one, in which the seller’s conviction funds the buyer’s purchase which confirms the seller’s conviction, and no point in the loop touches an outside source of truth about whether the world wants the product at the price being paid. The market has been built so that it cannot, from inside, distinguish its own faith from its own evidence.
I can tell you precisely when the question will be answered, because circles of this kind always answer it the same way. The demand that is real is the demand that remains when the capital stops circulating — when Nvidia, for whatever reason, ceases to fund its customers’ purchases, and the only chips bought are the chips bought with money earned somewhere outside the loop. That number exists right now, underneath the forty billion dollars, and no one can see it, because the forty billion is sitting on top of it, indistinguishable. One day the circulation will pause, for a reason no one will have predicted, and the real number will be revealed all at once, and everyone will say it was obvious. It is obvious now. It is simply not yet visible, which the circle has ensured is the same thing as not yet believed.