Say I give my nephew $100 to start a lawn-mowing business. He uses the $100 to rent my lawnmower for the summer. If I then tell people my mower rental business did $100 in sales, I am not lying - he really did pay me. But nobody outside the family has spent a cent. The money went from my pocket to his and back to mine.

Whether that was a good idea depends entirely on one thing: do the neighbours pay him to cut their lawns? If they do, I made a smart investment. My nephew has a real business, my $100 is worth more than $100, and my rental sales were the start of something. If they don't, I just paid myself $100 and called it revenue.

That is circular financing. A supplier funds a customer, the customer buys from the supplier, and the supplier's sales go up. It shows up whenever a company with a lot of cash sells to customers who don't have enough of it - which is why it tends to appear at the top of big investment booms, when buyers want more equipment than they can pay for.

The term got its modern fame in 2025, when Nvidia, AMD, Oracle and others signed a run of deals with OpenAI and other AI companies in which money and orders seemed to flow in a circle. Bloomberg published a widely shared diagram of the money flows in October 2025, and the phrase "circular deals" has followed the AI trade ever since.

Every circular deal splits the supplier's reported sales into two kinds of money. Only one of them is new.

This is why circular deals make demand look stronger than it is. The supplier's sales go up. The customer's spending goes up. Analysts watching either company see a boom. But the same dollars are being counted on both sides, and the only number that settles anything - what the end users actually pay - shows up years later.

"Circular financing" gets used as an insult, but it covers three different arrangements. Two of them are normal business practices that can be abused. The third is almost always a sign of manipulation.

The accounting rules draw a line here too. Under U.S. revenue rules, money a company pays to a customer generally comes off the revenue it earns from that customer - unless the company gets a separate good or service back in return. If it pays more than that thing is worth, the extra comes off revenue. If it can't tell what the thing is worth, the whole payment comes off. The rule is called consideration payable to a customer, and its effect is that a company can't pay a customer and still book the full sale as revenue.

The simplest test for any of the three: would this money have come to the supplier anyway, if the supplier hadn't provided it? If yes, the deal is just financing. If no, the sale exists because of the loop.

Every time this topic comes up, someone mentions Lucent, and for good reason. In the late 1990s, a wave of new phone and internet companies wanted to build networks, and equipment makers like Lucent, Nortel and Cisco were happy to help them pay for it. By September 30, 2000, Lucent had committed to lend its customers up to about $6.7 billion and to guarantee another $1.4 billion or so of their debt. Only about $2.1 billion of that had actually been lent or guaranteed at that point - the rest was promises waiting to be called on.

The customers were mostly young companies whose own customers hadn't shown up yet. When the money ran out, they went under - Winstar Communications filed for bankruptcy in April 2001 and sued Lucent. Lucent's provision for uncollectibles and customer financings - its charge for money it no longer expected to get back - went from $66 million in fiscal 1999 to $505 million in 2000 to $2.25 billion in 2001. About 60% of that last charge came from just three customer financings, including Winstar and the Australian carrier One.Tel.

One thing worth being precise about: the vendor financing itself was disclosed. When the SEC went after Lucent in 2004, the case was about something else - side agreements, credits and other incentives that let it improperly record about $1.15 billion of revenue in fiscal 2000, including a $125 million year-end sale to Winstar. Lucent paid a $25 million penalty, which the SEC said was for its lack of cooperation with the investigation. The case wasn't about the loans themselves, which Lucent had disclosed. The loans were just the reason the sales evaporated.

Elsewhere in the same era, the loops turned into outright fraud:

The SEC's description of the AOL deals is about as clean a definition of round-tripping as you will find: the company "effectively funded its own online advertising revenue." And according to the SEC's later case against Qwest executives, the company internally likened its dependence on one-time deals to an "addiction," and its one-time capacity and equipment sales to "heroin."

Twenty-five years later, the same shape showed up in artificial intelligence. A handful of chip and cloud companies had enormous amounts of cash. A handful of AI companies needed enormous amounts of computing power and were losing money building it. The deals that connected them made each side a supplier, customer, and investor of the other at the same time.

Follow the money around one loop: Nvidia invests in OpenAI, OpenAI pays Oracle for computing power, and Oracle buys Nvidia chips to provide it. That three-way circle is the picture most people have in mind when they hear the phrase.

Look at the CoreWeave row on its own and you can see why people get nervous. Nvidia is CoreWeave's shareholder, its supplier, and - through that capacity agreement - its customer of last resort. Three roles, one company.

The deals also changed shape as the criticism grew. The open-ended $100 billion pledge from 2025 never materialized. What Nvidia actually signed up for was a fixed equity stake and a lease guarantee. That matters, because a guarantee is the closer parallel to Lucent: it doesn't cost anything until the customer can't pay, and then it costs a lot.

The case for is that this is how big buildouts have always been paid for. The customer has demand it can't fund fast enough, the supplier has cash, and an investment gets the equipment built years sooner than banks would. The supplier also takes on risk it doesn't have to - a stake can go to zero. AMD's warrant only pays OpenAI in full if OpenAI actually deploys AMD chips and AMD's share price climbs - the last batch of shares needs a $600 stock price - which ties both companies to the same outcome. And Nvidia, in a memo to analysts reported by Barron's in November 2025, argued directly that "unlike Lucent, NVIDIA does not rely on vendor financing arrangements to grow revenue."

The case against is that the end customers are the weak link. OpenAI and the other AI developers at the center of these deals were spending far more than they earned, and their ability to keep buying depended on raising more money - some of it from their suppliers. If the people paying for AI services don't eventually cover the cost of all this hardware, orders get cancelled, stakes get written down, and guarantees get called. That is exactly the sequence that broke the telecom equipment makers. Critics also point out that when the same few companies are each other's investors and customers, a problem at one of them doesn't stay at one of them.

You don't need inside information to see a loop forming. Most of it is in the annual report if you know where to look.