by In September 2025, the chip company Nvidia announced that it would put as much as $100 billion into OpenAI. The pledge carried a condition that few headlines bothered to translate: OpenAI would spend a large share of the money leasing and buying Nvidia’s own processors to fill the data centers the investment was meant to build. Nvidia pays OpenAI. OpenAI pays Nvidia. The figure on the press release reads as growth. The mechanism underneath reads as a man moving a coin from his right pocket to his left and announcing that the household has grown richer.
By February 2026 the arrangement had stalled. The Wall Street Journal reported that Nvidia executives privately questioned OpenAI’s financial discipline, and Sam Altman went public with a flat denial that anything was wrong, which in the grammar of corporate crisis tends to confirm that something is. By early June the $100 billion letter of intent had been pared to a $30 billion equity stake, with Nvidia’s new Rubin processors entering production and OpenAI installed as an early customer. The dollar figure fell by seventy percent over those months, yet the underlying arrangement kept its shape. Capital still travels in a circle among a small set of companies whose fortunes now depend on one another’s continued spending.
The man who described the trick, then ran it
The pattern has a name and a biography behind it worth knowing, because the man who first described it also lived inside it. Richard Cantillon was born in County Kerry sometime in the 1680s, made himself a banker in Paris and London, and wrote a single book, the Essai sur la Nature du Commerce en Général, which the economist William Stanley Jevons would later call the cradle of political economy. He finished it around 1730. Printing waited until 1755, two decades after his death, and an English translation did not appear until 1931. Adam Smith read him. François Quesnay and the Physiocrats read him. For a man so few people can name today, his fingerprints sit on the founding documents of the discipline.
Cantillon’s insight came from watching money enter an economy and tracking where it went first. Imagine a new gold mine opening near a town, he wrote. The mine’s owner, its workers, and the goldsmiths and merchants who handle the first coins all grow richer, because they spend the new gold while prices still sit at yesterday’s levels. The baker two villages over feels nothing yet. By the time the new money reaches him, rising demand has already pushed prices up, and he buys his flour and pays his rent at the inflated rate without ever having held the windfall that caused it. The gold lifted the people standing closest to the hole in the ground and quietly taxed the people standing farthest from it. Economists now call this the Cantillon Effect, and its lesson is compact: the path money travels decides whom it enriches and whom it robs.
He drew the effect from his own ledger. When the Scottish financier John Law built his Mississippi scheme in 1719 and 1720, inflating the shares of a company tied to French colonial trade into one of history’s great manias, Cantillon held a privileged seat. He had early access to the share machinery, he lent money against the shares as collateral, and he understood the leverage propping up the whole structure better than the speculators buying in at the peak. He took his profits and got out before the collapse that ruined thousands. The first recipient walked away wealthy. The latecomers holding paper at the top were destroyed. Cantillon had run the same machine he would go on to explain, and he profited from standing near its source.
His end suited the intrigue. In May 1734 his London house burned, and the body found inside was identified as his. The accepted account holds that a dismissed servant murdered him and set the blaze. A competing theory, advanced by the economic historian Antoin Murphy in the definitive modern biography, holds that Cantillon, tangled in lawsuits with former clients trying to claw back his money, faked his own death and sailed for Suriname under another name. Either way, the author of the first treatise on how money moves arranged his final transaction so that his creditors were left holding nothing.
The loop, drawn to scale
Return to the data center with Cantillon’s map in hand. The new money in the AI economy does not pour from a gold mine or a central bank’s printing run. It arrives as investment: venture capital, corporate cash, sovereign wealth funds, and the accounting maneuvers that let a chipmaker finance its own customers. Cantillon’s lesson was never confined to minted coin. What he isolated was the distributional consequence of where fresh purchasing power lands first, and a flood of investment capital obeys the same rule as a flood of silver. Nvidia sits nearest the source, because it sells the one component every AI company needs and cannot easily replace. When Nvidia invests in OpenAI, and OpenAI commits to buying or leasing Nvidia chips, the dollars complete a loop without touching the wider economy. NewStreet Research estimated the geometry plainly: for every $10 billion Nvidia puts into OpenAI, it expects roughly $35 billion in chip purchases or lease payments to come back, a sum equal to about a quarter of its annual revenue. A dollar that leaves as an investment returns as a sale, and the company books both ends of the same trip.
The loop runs wider than two companies. Nvidia is a major backer of CoreWeave, a firm that borrows money, buys Nvidia GPUs, builds data centers around them, and rents the capacity to OpenAI and others. Nvidia took part in more than fifty AI venture deals in 2024 and kept the pace the following year, and many of those companies turned the capital straight back into Nvidia hardware. OpenAI, for its part, carries something close to $1.4 trillion in stated commitments to Microsoft, Amazon, Oracle, and AMD. The arrangement has a tidy elegance that analysts noticed at once. Stacy Rasgon at Bernstein wrote that the OpenAI deal would fuel circular concerns. An analyst at Seaport said the deals carried the whiff of circular financing and looked like bubble behavior. A Wedbush analyst, more generous, allowed that the same circularity might build a competitive moat if it held together. They argue about where the ring leads. None of them disputes that it is a ring.
One detail deserves attention, because it shows how the circle shields the people inside it. Nvidia structured part of the OpenAI arrangement as leases instead of outright sales. A purchased chip would force OpenAI to record the steep depreciation of hardware that loses value the moment a faster model arrives. A leased chip keeps that charge off OpenAI’s books and flatters its bottom line. The choice makes the loop look healthier than its physics allow, which is the oldest function of financial engineering: to delay the moment when reality gets entered into the ledger.
Step back, and the resemblance to Cantillon’s mine sharpens. The companies nearest the source, Nvidia above all, grow visibly richer on transactions they substantially fund themselves. Valuations climb. Markets read the throughput as demand. Whether independent buyers actually want all these chips at these prices is the question no one can answer cleanly, because the circular flow makes the company’s revenue and the company’s investments into the same money wearing two hats.
Where the bill arrives
So far the circle has cost the people inside it nothing. Someone pays, though, because the chips and the data centers are physical objects built from finite materials, and the demand they generate reaches into the same supply that ordinary buyers draw on. Here the abstraction becomes a number on a price tag.
The squeeze shows up first in memory, and this is worth stating precisely, because the popular shorthand about a processor shortage points at the wrong part. AI servers consume enormous quantities of memory, and the most profitable kind, the high-bandwidth memory stacked beside the processors, is built on the same production lines that make the ordinary DRAM in a laptop or a phone. Forced to choose, the three companies that control almost the entire market, Samsung, SK Hynix, and Micron, have steered the overwhelming share of their output toward the high-margin AI product. Consumers met the result as a shock. DRAM prices rose more than 170 percent across a single year. DDR5 spot prices quadrupled after September 2025. In the first quarter of 2026 alone, consumer memory climbed by as much as 110 percent and solid-state drives by close to 150. Micron retired its consumer Crucial brand at the end of 2025 to serve the AI market without the distraction of ordinary customers. Dell, Lenovo, and HP signaled personal-computer price increases of fifteen to twenty percent for early 2026, and memory, which made up under a tenth of a machine’s component cost in 2024, now runs close to a fifth.
The forecasts stretch the pain forward. Analysts expect data centers to absorb a majority of the world’s memory output in 2026, with United States technology giants alone committing on the order of $620 billion to AI infrastructure for the year. New fabrication plants may relieve the pressure, though the earliest will not reach full production until 2027 or 2028. Until then the shortage radiates outward, with warnings that it could reach automakers and repeat the vehicle-production crisis of the pandemic years.
Read that against Cantillon and the transfer is exact. Each of them stands at the far edge of the AI money loop: the student who needs a laptop, the small developer who needs a workstation, the household replacing a dead computer, the carmaker waiting on parts. None of them holds the capital circulating at the center, and all of them pay the higher price that circulation produces. The wealth stayed in the room. The bill walked out of it.
The empire that drowned in its own treasure
History supplies larger versions of the same machine, and the grandest belongs to sixteenth-century Spain. When silver from the mountain of Potosí in present-day Bolivia, opened in 1545, and from Zacatecas in Mexico the year after, began crossing the Atlantic, the metal entered Europe through a narrow gate. The Crown took its share first. The Genoese and German banking houses that financed the Habsburg wars, the Fuggers chief among them, took theirs next, holding royal debt as collateral and lending against the incoming bullion. Global silver output climbed from under three million ounces a year in 1521 to more than thirteen million by 1600. Prices in Spain rose several times over across the long span from 1500 to 1650.
The men at the gate spent first and spent at the old prices. Spain’s laborers, artisans, and farmers met the new money only after it had raised the cost of bread and cloth and rent, which is the Cantillon Effect operating three centuries before Cantillon named it. The deeper damage ran past inflation. Spain stopped building. Why finance a textile mill or improve a farm when the Crown could buy finished goods from the Netherlands, England, and France with bars of silver? The treasure made productive work feel unnecessary to the people who held it, and the country hollowed out beneath the shine. Economists call the syndrome the Dutch Disease: a resource windfall that strangles a nation’s own industry. Forensic study of the trade routes shows that much of the silver never settled in Spain at all. It passed through, a transit node on a longer road, and a large share of it came to rest in Ming China, where silver commanded nearly double its European value and a tax system newly payable in silver created a bottomless appetite. Philip II, master of the richest empire on earth, defaulted on his debts four times.
The rhyme nobody wants to hear
The closest rhyme to the AI loop is recent enough that its participants are still working. In the late 1990s, as investors poured money into companies laying fiber-optic cable for an internet they were sure would demand infinite bandwidth, the firms that made the networking equipment found a way to guarantee their own sales. Cisco, Lucent, and Nortel lent their customers the money to buy their gear. Lucent committed up to $8.1 billion in customer credit and loan guarantees. Nortel extended $3.1 billion. Cisco promised $2.4 billion. The customers, many of them startups with no profits and outsized ambitions, ordered more equipment than the market could ever absorb, partly because the gear was scarce and hoarding it felt prudent, partly because someone else was paying for the purchase. Bill Frezza, a general partner at the venture firm Adams Capital Management, described vendor financing in that era with a single word: a drug.
The bill came due in 2001, and the figures read like a warning issued in advance. Cecilia Wagner Ricci, a finance professor at Montclair State University, calculated from securities filings that bad loans at Lucent went from under three percent of its loan book at the end of 2000 to sixty percent a year later. Nortel’s bad loans climbed from a quarter of its book to four-fifths in the same twelve months. Motorola’s went from under seven percent to fifty-seven. Lucent wrote off $700 million in bad debt, then set aside billions more across the next two years. Global Crossing, which had run an even cruder version of the trick by paying another firm for services while that firm bought an identical sum of capacity in return, a maneuver the industry called revenue roundtripping, went bankrupt, and its executives later paid large legal settlements. Between 2000 and 2002, telecom stocks shed more than $2 trillion in market value. WorldCom filed the largest bankruptcy in American history to that point.
The mapping onto the present writes itself. Equipment makers of 1999, selling scarce hardware to a frenzy of buyers they were quietly funding, occupy the seat Nvidia holds now. The fiber startups, burning capital to build capacity on the faith that demand would show up to meet it, occupy the seat of the AI companies. Vendor financing that looked like shrewd demand creation turned out, once the buyers could not pay, to be a chipmaker holding bad paper across the table from a bankrupt customer. The participants in 2025 know this history. Some of them lived through it. The knowledge has not bent the shape of the deal, because the shape is too profitable to abandon while the music plays.
Closer to home, and harder to see
The loop needs no global empire or trillion-dollar chipmaker to form. Smaller, quieter versions run inside ordinary public budgets, and one of them sits in New York State’s spending on the education of disabled children. Money appropriated in Albany flows to the State Education Department, which routes it to school districts and to the state-operated and state-approved institutions that serve students with the most significant needs. A public dollar moves to a public or publicly sanctioned institution, that institution records the spending as services delivered, and the recorded spending becomes the measure by which the system reports that it is doing its job.
Honesty requires marking where this parallel weakens, because the corporate loop and the public one are not identical. Public money is never literally trapped: it pays teachers and aides and therapists who spend their wages across the wider economy, and a real child at the end of the chain does receive instruction. The resemblance lives in a subtler place. The further a dollar travels through administrative layers, approvals, compliance reporting, and contracted intermediaries, the more of its value converts into the activity of the system itself rather than into help reaching the student. Throughput gets counted as outcome. A budget that grows every year can sit alongside a classroom that barely changes, because the growth is consumed by the circulation. The disabled child is the everyday user of that system, and the everyday user is always the last to be paid.
Motion mistaken for wealth
Set these cases side by side and a single distinction organizes all of them. Money moving quickly among a small group of holders generates every visible sign of prosperity: rising valuations, busy ledgers, impressive totals on press releases and budget lines. Those signs are real enough on their own terms. The question each case forces is whether anything of value reaches the people outside the circle, and whether the motion builds productive capacity that did not exist before. Cantillon’s mine at least produced gold. Spain’s silver at least existed as metal, even as it gutted the country that imported it. The harder cases are the ones where the circulation produces mostly more circulation, where a dollar’s entire career is the trip from one balance sheet to another and home again.
The last recipient
The everyday user occupies the place Cantillon assigned to the baker two villages from the mine. The user holds none of the new money, sees none of the investment, owns none of the equity climbing on the strength of the loop, and meets the whole apparatus only at the checkout, where the laptop now costs more and the memory inside it has tripled in price. Writers reaching for an image of the Cantillon Effect sometimes describe new money as honey poured at the center of a table. It pools where it lands, thick and bright, and it creeps toward the edges slowly, arriving thin and watered down if it arrives at all. The people at the center taste it first and taste it richest. The people at the edge get what is left after the evaporation.
This arrangement carries a politics, whether or not its architects intend one. When the gains from a new technology concentrate among the handful of firms and investors standing nearest the money, while the costs scatter across everyone standing farthest from it, the result is a steady transfer of wealth from the broad base of a society toward its narrow top. Cantillon documented the mechanism in an age of kings and gold mines. Data centers and circular financing have rebuilt it at a scale he could not have pictured, with the same people winning and the same people paying. The coin still moves from the right pocket to the left. The household is still told that it has grown richer. Knowing whose hand holds the coin, and whose hand never will, is the first defense an everyday user has.
Leave a Reply