When the Seller Bankrolls the Buyer: How Circular Financing Actually Works

Last week Meta raised $37.5 billion through two stock sales, as its AI spending climbs. In the fine print, Nvidia, AMD and Broadcom, which all sell Meta chips, had indicated they might together buy up to $5 billion of the common shares on offer. Not a binding promise, the prospectus said.

So why would a company's suppliers buy its stock? Online, the answer usually comes as a diagram: money goes from the chipmaker to Meta, Meta spends it on chips, the chipmaker books the sale, and round it goes. Circular financing: the sale looks real, but the chipmaker is mostly paying itself.

The diagram leaves things out. Nvidia has since said it will put in $3 billion and AMD $1 billion; Broadcom hasn't given a figure. They would buy on the same terms as other investors. Nothing we found ties their money to chip orders, though Nvidia and AMD announced their investments alongside plans to sell Meta more. And $5 billion is small next to the $135 billion to $145 billion Meta expects to spend on capital projects this year.

Most of all, the diagram makes it look new. Sellers helping buyers pay is one of the oldest moves in business. What's new is how often the seller has little it can take back if the customer stumbles, and how big the bets have become: one chipmaker has agreed to guarantee up to $105 billion of a single customer's lease and power payments.

The Money Map

Here's what the diagram's loop actually does to a seller, in round numbers. Say a chipmaker puts $10 billion into a customer for a stake. The customer spends it all on chips that cost the chipmaker, say, $3 billion to make. The cash comes straight back, and the chipmaker reports $10 billion of sales and $7 billion of profit on the sale. If the customer thrives, the chipmaker wins twice. If the customer fails, the stake is worthless and the chipmaker is out $3 billion of chips, with $10 billion of sales still on its books.

It wasn't paying itself. It was betting on its customer, and the sales figure doesn't show the bet.

Five ways a seller can help a buyer pay

In practice, those three basic moves come in five common forms, and the last one runs the bet the other way.

Read down the table and the seller's protection gets thinner.

It's older than it looks

In 1841 the Philadelphia & Reading Railroad needed locomotives and coal cars to start running, and it was deep in debt, much of it already in default. A machine shop in Lowell, Massachusetts, Locks and Canals, agreed to build them anyway.

It didn't want to lend the railroad money. It wanted a stake in the railroad, according to Michael McMillen, a finance lawyer who has researched the deal. It settled for payment by the ton: paid only as fast as the railroad hauled coal, with the equipment held in trust as security. The railroad fell behind on its payments for years.

Swap coal for computing and you have the shape of many of today's chip deals: a seller paid as its customer grows. Today's seller often gets what Locks and Canals wanted, a stake. What it gives up is what Locks and Canals settled for: the locomotives held in trust.

Industry after industry has run some version since. GM set up its own lender in 1919 because, as GMAC's successor Ally tells it, banks didn't usually make car loans: help that makes a sale possible. In its 2025 fiscal year, Deere paid its own finance arm $389.9 million so customers could borrow at reduced or zero rates: a discount by another name. And at the end of 2025, Boeing had $15.2 billion of financing commitments for planes, most of it to customers it considers below investment grade, the riskier end of the credit scale, though it didn't expect to fund "a significant portion": a promise, not a payment.

Most of these sellers lent, or promised to lend, against something they could take back: a car, a tractor, a jet. Today's sellers more often take stakes, which leave nothing to repossess, or give guarantees, some backed by chips that lose value fast. One has turned the bet around altogether: AMD.

Watch the machine work: AMD, OpenAI and Meta

In October 2025, AMD struck a deal with OpenAI covering chips for up to 6 gigawatts of computing. A gigawatt measures how much power the computers draw, about what a large nuclear reactor produces. It isn't a price. Deliveries were due to begin in the second half of this year.

AMD's filing sets out how it works:

1. AMD hands over a warrant. That's a right to buy up to 160 million AMD shares, about a tenth of the company, for one cent each.

2. It unlocks in pieces. The first piece unlocks when the first gigawatt is delivered. The rest unlock as OpenAI buys more, up to the full 6.

3. AMD's stock has to hit targets, too. The targets climb with each piece, to $600 a share for the last: more than three and a half times AMD's price just before the deal was announced. Other conditions apply.

4. OpenAI collects, maybe. Once a piece unlocks, OpenAI can buy those shares for a penny each, any time until October 2030.

As of the end of June, none of the shares had unlocked, AMD's quarterly filing shows.

If OpenAI bought the full 6 gigawatts and every target were met, it could buy 160 million shares, worth about $96 billion at $600 each by our arithmetic, for $1.6 million. In effect it's a volume discount, paid in AMD's own shares.

The targets protect AMD's existing shareholders: they give up the full tenth only if the stock reaches $600, which would leave them with a smaller slice of a far more valuable company. The stock has already done most of that climbing, closing at about $560 on October 27. If it holds there, what's mainly left is the deliveries. AMD said it expected the deal to bring in "tens of billions of dollars" of revenue and to be "highly accretive" to its adjusted earnings per share: in other words, to lift its profit per share, on a measure that leaves out some costs.

Notice which way the bet runs. In this deal, AMD isn't buying into its customer. OpenAI is, potentially, buying into AMD.

In February, AMD signed a matching deal with Meta: up to 6 gigawatts, and a warrant for 160 million shares at a penny each that unlocks as Meta's purchases grow. Meta committed to buy the first gigawatt. By our arithmetic, the two warrants together cover about a fifth of the shares AMD had when the first was signed. Then last week AMD said it would also put $1 billion into Meta's stock sale.

So AMD has offered Meta its own shares for buying its chips, and says it will pay cash for Meta's. That's the loop the diagrams draw. But Meta isn't a buyer that can't get financing: investors have just agreed to buy $37.5 billion of its stock. So the first of the three reasons doesn't fit. The warrant fits the second: a volume discount, paid in AMD shares. The $1 billion stake fits the third, a small share in the growth of one of the world's most valuable companies.

Follow the money

Headline numbers can shrink. Nvidia's September 2025 plan to invest "up to $100 billion" in OpenAI was a letter of intent, not a contract. By March its chief executive, Jensen Huang, was calling it "probably not in the cards." Nvidia agreed instead to invest $30 billion in a funding round OpenAI closed this spring with $122 billion of commitments.

Amazon put $50 billion into the same round, all of it paid in by the end of September, its filings show. Alongside it, OpenAI expanded its cloud contract with Amazon by $100 billion over eight years. Here the loop is plain to see: cash one way, a bigger promise to buy the other. But the promise isn't cash, and nothing public requires OpenAI to spend Amazon's $50 billion with Amazon.

Then there are guarantees, which move no money unless something goes wrong. In August, Nvidia agreed to guarantee up to $105 billion of OpenAI's lease and power payments on a planned data-center campus in Ohio. It pays only if OpenAI doesn't.

With Anthropic, Broadcom wears four hats at once.

• Supplier: its AI chips end up in Anthropic's computers.
• Landlord: an Anthropic subsidiary leases some of that hardware from Broadcom under a five-year contract worth $125.2 billion, not counting interest, according to Anthropic's IPO filing.
• Backstop: Broadcom is handing part of the landlord role to outside investors, but not all of the risk. Last week three banks and Blackstone lined up $63 billion of committed financing for a new company that will buy Broadcom's AI chips and lease them to Anthropic. If that lease goes bad and the chips are sold, Broadcom promises the sale will cover at least 72.5 cents of every dollar still owed. If the chips fetch less, Broadcom pays the difference, up to about $43 billion. It calls that a low-probability event.
• Possible lender: under certain conditions, Anthropic can borrow up to $42 billion from Broadcom, or a financial partner Anthropic approves, in loans that can be turned into Anthropic shares. That's a loan that can become a stake.

What the help costs

Every kind of help has a cost, and it shows up in different places: in Deere's margin on the tractor, in the $50 billion Amazon has riding on OpenAI, in the slice of AMD its shareholders would give up. Guarantees cost no cash until they're called on, and they often sit on the books at little or no value until a payment looks likely, notes the Bank for International Settlements, the central banks' bank. That makes them the easiest help to miss.

When the machine breaks

If the AI boom has a precedent, it may be the late-1990s telecom build-out rather than the 2008 financial crisis, researchers at LSEG, which owns the London Stock Exchange, suggested last month.

Back then, new phone and internet companies needed networks before they had customers, and equipment makers lent them the money to buy their gear. Lucent alone had promised $6.7 billion of loans and $1.4 billion of guarantees by September 2000. One customer, Winstar, had a $2 billion credit line from Lucent and filed for bankruptcy in April 2001. In the year to September 2001, Lucent set aside $2.2 billion for bad debts and customer financing, about $1.3 billion of it for just three projects, Winstar's among them.

Nortel's customer financing, counting money promised but not yet lent, shrank from $5.2 billion to $2.1 billion during 2001. Some customers, it said, "failed to meet the borrowing conditions" or "reduced their commitments to purchase our products and services."

That sentence is the trap. The same weakness that stopped customers borrowing stopped them buying, so the equipment makers lost sales and took losses on money already lent, at the same time and from the same customers.

Those were loans, which get paid before shareholders when a company fails, and Lucent still had to set aside billions. A stake gets paid last.

Today's guarantees sharpen the trap. Bankers call it wrong-way risk: what you stand to lose on a customer is biggest exactly when that customer is weakest. If OpenAI ever missed its payments once the Ohio leases begin, Nvidia could be on the hook for them, up to $105 billion in all. At the same time, the $30 billion stake it agreed to buy in OpenAI would be losing value and, if the telecom pattern held, OpenAI's demand for Nvidia's chips would be shrinking. Three exposures, one customer. Broadcom's backstop has the same shape: it would pay most when AI chips are worth least, which could also be when it's selling fewer of them.

What's different this time, and what isn't

Only the first of these differences makes it safer.

The customers are bigger. Winstar was a young carrier. Meta is a giant that earns its own cash. OpenAI lined up $122 billion this spring, though money raised isn't money earned.

What a lender can take back wears out faster. GATX and Air Lease, which lease out railcars and jets, expect their equipment to last 25 years or more. Alphabet, Meta and Amazon assume five to six years for their servers, and last year Amazon cut some of its to five, citing the pace of change in AI. The leases behind Broadcom's deals with Anthropic run about five years, roughly as long as those servers are expected to last. That's the risk Broadcom's guarantee partly covers: it turns on what the chips fetch when they're sold.

It's a web, not a circle, in LSEG's phrase. Between 2021 and 2025, more than half (55.2%) of the money invested in the AI companies tracked by the BIS came from other AI companies, the central banks' bank found in October. In almost half of those deals by value (46.4%), the investor was also a supplier or a customer. And the terms, the BIS noted in June, are often poorly disclosed.

What hasn't changed. In the end, every loan, lease and stake here is settled by the customer's own customers, paying for ads, subscriptions and cloud contracts. Even Meta, which earns its money from advertisers, is expected to spend faster than that money comes in. Last week S&P Global kept Meta's AA- credit rating, near the top of its scale, but moved its outlook to negative, a warning that a downgrade is more likely. By its own measure, S&P expects Meta's spending to outrun the cash its business brings in by about $39 billion this year and $71 billion next, Reuters reported.

The part nobody explains

When a seller owns part of its customer, one fundraising round can lift the seller's earnings twice: as sales when the money is spent, and as a gain on its stake when the round is priced higher, long before the customer's own sales have proved that price.

Amazon has put about $18 billion, by our count, into Anthropic, which has committed to spend more than $100 billion on Amazon's cloud over the next decade. When Anthropic raises money from other investors at a higher price, Amazon marks up the value of its stake. In the second quarter of this year, Amazon reported $53.4 billion of pre-tax income from outside its main business, "primarily from our investments in Anthropic." Most of that, by our reading of its filing, was a change in value on paper, not cash, and in a single quarter it came to far more than Amazon had put in.

That's standard accounting. But it means a supplier's earnings can look strongest at exactly the moment it is most tied to one customer.

It's the telecom trap, seen on the way up. The sales and the gains rise together, and would fall together.

Why it matters

For investors in the sellers, sales to a customer the seller has helped finance aren't the same as sales to one paying its own way. Only the first leaves the seller holding a loan, a stake or a guarantee that rides on that same customer.

For lenders, a supplier's guarantee is only as strong as the supplier, and here the supplier's health can ride on the same customer.

For everyone else, it changes how to read the biggest numbers in the AI boom. Announced investment, committed financing, purchase commitments and actual sales are four different things, and the same dollar can show up in more than one.

The next time a deal like this makes the news

Ask:

1. Who actually handed over cash, and when? Or is it a letter of intent, a commitment, an "up to"?
2. What did the seller get back? A loan, a stake, rights to shares, or a promise to cover someone else's loss?
3. Does the customer have to buy from the seller, or is it just expected to?
4. If the customer struggles, what can the seller take back, and will it still be worth anything?
5. Who's paying from outside the loop, and is it enough?
6. Is the seller booking gains on its stake as well as sales?

A locomotive builder in 1841 and a chipmaker in 2026 are making the same wager: that the customer they help today will grow into one that pays. The seller can hold the goods, a loan, a promise or a stake. The further down that list it goes, the less it can take back.

Whatever the seller holds, the buyer still needs customers of its own.