
Nvidia sells the chips that run the artificial intelligence boom. It also stands behind much of the money its customers use to buy them. It invests in those customers, backs their loans and promises to buy the computing power they cannot rent out. Much of that money comes back to Nvidia as sales. Jensen Huang ran that system himself, and on Nov. 12 Nvidia’s board forced him out. The promises he signed did not leave with him.
Colette Kress, the chief financial officer now serving as interim chief executive, gave her first report on Nov. 17. Nvidia’s sales for the three months ending in October topped $100 billion for the first time, more than double a year earlier, and its profit more than doubled too. When the company last reported, in August, she wrote, “We continue to make strategic commitments across our supply, infrastructure, and partner ecosystems.” On the Nov. 17 call, she told analysts it was “business as usual at Nvidia.” Those commitments are now hers to carry.
The spy story put an affair at the center of Huang’s fall. Nvidia shareholders have a different worry: who keeps the money moving now? Before he fell, analysts cited by Bloomberg columnist Matt Levine said: “Nvidia is almost completely dependent on him.”
How the money goes around
CoreWeave is the plainest case. CoreWeave rents out computing power from data centers filled with Nvidia chips. Nvidia sells it those chips. Nvidia also owns its stock and is its second-largest shareholder. And under an agreement CoreWeave disclosed in September 2025, Nvidia will buy any computing power CoreWeave cannot sell to anyone else, through April 2032, in a deal first valued at $6.3 billion. Nvidia is CoreWeave’s supplier, its owner and its customer of last resort.
OpenAI is the biggest case. Nvidia agreed in February to invest $30 billion in OpenAI. In August, it agreed to guarantee up to $105 billion for a cluster of data centers in Pike County, Ohio, built by a SoftBank company and leased to OpenAI for up to 20 years. Nvidia is the only chip supplier for the first phase. OpenAI could not raise money on that scale in its own name. Nvidia’s name stands behind it.
Anthropic, another big AI company and another big Nvidia buyer, got the same treatment. In November 2025, Nvidia committed up to $10 billion to Anthropic, and Anthropic promised to spend $30 billion on computing power from Microsoft’s cloud, built on Nvidia hardware. In September, Reuters reported that Nvidia was in talks to put up to $10 billion more toward Anthropic’s planned public stock sale.
Then there are the smaller cloud companies. By the end of October, Nvidia had promised to buy up to $32 billion of computing power from them. The promises take effect once those companies borrow the money to buy Nvidia’s equipment and get it running. Nvidia books the sale of that equipment as revenue up front, then stands ready to buy the computing power that equipment produces.
Nvidia bankrolls the startups too. Through Nov. 10, it had taken part in a record 129 venture capital deals in 2026, deals that made up more than a quarter of the value of all such deals worldwide, PitchBook counted. Some of those startups burn through huge sums each month and have yet to bring in much money, The Wall Street Journal reported.
Selling to itself
Nvidia says plainly why it does this. Many of its customers, its Nov. 17 report says, are “growing faster than their balance sheets and long-term credit profiles can support.” In plain words, they cannot pay for their expansion on their own.
Nvidia supports its buyers in different ways. Sometimes it buys their stock. Sometimes it guarantees their leases so lenders will provide the money. Sometimes it promises to buy the computing power they cannot sell, which gives them a contract to show their lenders. And sometimes it simply gives them longer to pay. Each way, the chip sale goes on Nvidia’s books like any other, as revenue. For Nvidia, the sale counts the day the chips ship. For the buyer, the hard part is still ahead: it has to sell the computing power those chips produce, at a price that brings back a profit. What closes Nvidia’s sale is the buyer’s promise to do that someday and earn enough to pay back its investors, its lenders, and its guarantor. Some of the money behind that promise is Nvidia’s own, coming home. More of it is borrowed on Nvidia’s word.
Marx explained what credit does in a boom. The credit system, he wrote in “Capital,” Volume III, Chapter 27, “accelerates the material development of the productive forces,” and at the same time accelerates “the violent eruptions” of capitalism’s contradiction: crises. That happens because so much of the capital “is employed by people who do not own it.” Credit lets companies build past the limit their own capital would set. The data centers get built, and the chips get made. The claims on future profit pile up faster than the profit itself.
The technology under those claims is real. A chip in an Ohio data center is a real machine. A guarantee that its owner will pay the rent for 20 years is a bet on profits not yet made.
The guarantees are credit: Nvidia lending its name so that others will lend the money. The promises to buy are contracts to pay later, which the cloud companies can show their own lenders. The stock is something else: a claim. Nvidia’s shares in CoreWeave, the shares Anthropic plans to sell to the public and Nvidia’s own stock are what Marx called fictitious capital, titles to a slice of future profit that are bought and sold as if they were capital themselves. Their prices move on what buyers expect that future profit to be, so they can climb far above anything actually earned, or fall far below it.
In the quarter ending in October, Nvidia marked down its own stakes in private companies by $1.8 billion. Its profit more than doubled in that quarter, and its stock still lost nearly a quarter of its value in November, CNBC reported.
The holders of those claims reach all the way to the White House. Donald Trump’s accounts have traded tens of millions of dollars of Nvidia stock since he returned there in January 2025, CNBC found.
Nvidia makes enormous profits today because it holds a monopoly on the chips AI runs on and can charge accordingly. The value comes from workers: those who make the chips in Taiwan, test and package them in Malaysia, and build and run the data centers. The monopoly lets Nvidia take far more than its share of that value. Those superprofits pay for Nvidia’s investments in its customers and stand behind its guarantees and promises to buy. The monopoly sets the price, the price brings in the cash, and the cash, and the credit it backs, keep the customers buying at that price. Lenin described finance capital as the merging of the banks with industrial monopoly. At Nvidia, the industrial monopoly does the banking itself.
The buyers it finances are the buyers it depends on. Three customers that buy directly from Nvidia paid about two of every five dollars it took in from February through July.
It has bet on the boom from the supply end too. By the end of October, Nvidia had promised to buy $390 billion of memory and factory output from its suppliers in the years ahead, more than three times the figure six months earlier. Much of the latest increase comes from soaring memory prices, The New York Times reported. That is production ordered for sales that have not happened yet, to customers who are partly spending Nvidia’s own money. Huang knew it. In August, he told Nvidia’s finance team that the company was facing a glut of the chip wafers and memory it had committed to buy, as the big cloud companies planned to use other chips, the Journal reported.
The bill comes later
In its July filing, Nvidia valued the guarantees it had already made on data center sites as “not significant.” The Ohio guarantees it signed in August could cost it up to $105 billion. They take effect in stages as the data centers open, the first not expected before 2028. Counting all its guarantees, Nvidia could owe up to $118 billion, its Nov. 17 report says. It has paid nothing on any of them so far, but it carries the risk from the day it signs.
AIG looked the same way before 2008. The insurance giant sold guarantees on bonds built from home mortgages. With AIG standing behind those bonds, the banks kept lending, and AIG counted the guarantees as nearly free money. When the mortgage market broke, the guarantees came due all at once. The United States government committed about $180 billion to keep AIG from collapsing and taking the banks down with it.
Henry Paulson, the Treasury secretary who helped run that rescue, warned on Nov. 17 that markets are underestimating the risks of AI. He said he had never seen “so much money chasing one opportunity,” with so little in earnings behind it.
Nvidia has already flinched once. In July, it was discussing a guarantee of about $250 billion for the Ohio data centers, the Journal reported. By August, with investors worried about how much risk it was taking on, it had cut that to less than half. Kress had grown skeptical of the size of the deal, the Journal reported.
Wall Street lends too
When Nvidia pulls back, someone else has to carry the risk, and Wall Street is already in the deals. Huang’s deals made billions of dollars in fees for banks, asset managers and advisers, the Times reported. The biggest tech monopolies have borrowed hundreds of billions of dollars in 2026 to build their own data centers.
Wall Street wants to be paid for the risk. In August, Nvidia and Blackstone, Apollo, KKR and other firms announced a $500 billion target for loans to Nvidia’s customers, with the chips themselves as security. The firms had not raised a dime when they announced it, the Times reported. By October, Reuters reported, lenders were demanding stronger guarantees from Nvidia than it had offered. By mid-November, none of the loans had closed, and at least one of the firms was talking about pulling out, the Times reported.
Banks and investment firms also raise money for the data centers run by companies in Nvidia’s partner network. Cantor Fitzgerald, the Wall Street firm now run by Commerce Secretary Howard Lutnick’s sons, has raised money for one of them: Bitdeer, a Singapore company building a data center in Malaysia filled with Nvidia’s most advanced servers, the Times reported on Nov. 15. Lutnick’s Commerce Department writes the rules those data centers operate under, and he pushed to keep those rules loose, the Times reported. On Nov. 16, the administration announced tighter rules anyway, written at the direction of the Treasury Department and the National Security Council.
The man and the system
The checkbook was closing before Huang was pushed out. In October, at Nvidia’s conference in Berlin, he told the cloud companies that buy its chips they would have to collect payment up front from their own customers to get Nvidia’s chips, the Journal reported.
Huang made the deals. The board removed him. It cannot remove the deals. The CoreWeave agreement runs to 2032. The Ohio lease runs as long as 20 years. Every investment, guarantee and promise to buy rests on profits the AI boom has yet to produce.
Those profits depend on data centers that still have to be built and powered, and the communities around them are fighting back. In 2026, voters have sided against data center development in 46 of 47 local ballot measures, Ballotpedia counted. Near some data centers, wholesale electricity prices have more than tripled in five years, a Bloomberg analysis found.
The chips are real, and so are the data centers. The profit that is supposed to pay for them has not been made. Every monopoly in the boom is building at once, each racing for its share. The more capacity they build, the harder it becomes to sell the computing power at a price that brings in the profit their investment counts on. Marx called that contradiction overproduction. Huang saw an early sign of it in August. He warned of a glut in the wafers and memory Nvidia had ordered.
That is how capitalism builds a new industry: on credit, ahead of the profit that has to pay for it, until the bills come due. The spy story will get the blame for the pullback. It started in the boom itself, where Nvidia’s own money and credit help pay for the chips that show up as Nvidia’s sales.
Join the Struggle-La Lucha Telegram channel