Straight talk about circular deals in AI

The two firms at the center of the AI boom, Nvidia and OpenAI, are also at the center of a confusing mesh of recent “circular deals” with their suppliers and customers.[1] According to Bloomberg:[2]

The companies, which ignited an AI investment frenzy three years ago, have been instrumental in keeping it going by inking large and sometimes overlapping partnerships with cloud providers, AI developers and other startups in the sector. In the process, they’re now seen as playing a key role in ratcheting up the risks of a possible AI bubble by inflating the market and binding the fates of numerous companies together.

The New York Times headline asks:[3]

               Is A.I. Investment Getting Too Circular?

I don’t agree that these transactions are “ratcheting up the risks of a possible AI bubble.” Yes, previous bubbles involved dubious accounting and sketchy deals, but there’s nothing intrinsically alarming about firms partnering with each other. Cross-corporate holdings and deals might be confusing, but they’re neither necessary nor sufficient ingredients for a stock market bubble.

I would find “circular financing” concerning if it resulted in any of the following:

• Selling equity to outside shareholders.
• High (and perhaps hidden) leverage.
• Distorted information about fundamentals such as earnings and revenue.

For most of the recently announced circular deals, I’m not seeing these features. I’m especially interested in issuance; in a bubble, I’d expect to see AI firms selling equity, and most of these deals don’t involve equity issued to external shareholders. For me, net issuance is one of the Four Horsemen of the Bubble Apocalypse. No issuance, no bubble. 

Let’s take a specific example, the AMD/OpenAI deal. On the surface, it does involve issuance of equity, as OpenAI has received warrants that under certain conditions will turn into shares of AMD. Here’s how the indispensable Matt Levine describes it:[4]

This deal between OpenAI and AMD was obviously going to create a lot of stock-market value: The announcement of the deal would predictably increase the market value of AMD … Why not use that value to subsidize the deal? Schematically, OpenAI could buy AMD stock to predictably profit from the stock-price bump it created. Just going out and doing that in the market would be awkward — it might look like insider trading — but buying the stock from AMD is fine.

The essence of the deal is that OpenAI is “buying the stock from AMD.” That’s the opposite of the pattern that I’d expect in a stock market bubble, which is firms selling shares to gullible retail investors. If and when OpenAI receives its AMD shares, OpenAI may decide to just hang on to them, or it may decide to sell its AMD shares to external shareholders. We don’t know.

If firm A gives firm B some of its shares, but B holds the shares forever, that’s not a sign that A is overpriced. On the other hand, if B immediately sells its shares in A, then that indicates that B thinks A is overpriced. Firms are the smart money, and when you see them selling shares to retail investors, that’s a powerful sign that the shares are overpriced. But if a deal is truly circular, that is, involves no cash flowing in or out of A+B, I don’t think you can infer anything about the valuation of either A or B.

Let’s compare the AMD/OpenAI deal with a similar deal in 1917 between DuPont and General Motors (GM). Just as AMD is a supplier to OpenAI, DuPont was a supplier to GM. In 1917, DuPont bought shares in GM, becoming the largest shareholder. DuPont remained a GM shareholder for 40 years, until the Supreme Court forced DuPont to divest on antitrust grounds in 1957. DuPont’s investment in GM was wildly profitable, with GM eventually becoming America’s most valuable listed firm.

If OpenAI receives AMD shares, will it be like DuPont and hold on to these shares for 40 years? I have no idea. We’ll have to wait and see.

The long history of corporate alliances

I’m not sure why anyone is concerned about “binding fates of numerous companies together” as mentioned by Bloomberg. If company A is a customer of company B, and perhaps also company B is a customer of company A, then obviously their fates are bound together. Given this connection, it might be a good idea for A and B to form an alliance, or it might not, but the binding is already there.

For example, in the industrial revolution, steel companies sold steel to railroads while also being a major customer of railroads (to transport coal, iron, and steel). You could have the steel company and the railroad as completely separate firms, you could merge them into one company, or you could have some intermediate arrangement involving cross-corporate holdings, but they’re going to be intertwined. Even the Soviet method of industrialization involved a feedback loop of increasing steel production to build railroads to transport more raw materials to the steel mills.

In many countries, the process of industrialization involved various forms of corporate alliances, sometimes with cross-corporate equity holdings. These arrangements had costs and benefits, but whatever you think about them, they didn’t generally cause stock market bubbles. As the DuPont/GM example shows, many alliances were long-lasting and mutually beneficial.

“Fiscal incest” and “a de facto Ponzi scheme”

There are several historical episodes where seemingly circular financing contributed to stock market bubbles. Let’s discuss these episodes in chronological order.

First, the U.S. stock market bubble that peaked in 1929 involved a type of circular financing. Galbraith (2009) describes the pyramidal structure of investment trusts, which today we would call leveraged closed-end funds. Fund A would own shares in fund B which would then own shares in fund C. Galbraith uses the memorable term “fiscal incest:”

There was a rush to sponsor investment trusts which would sponsor investment trusts, which would, in turn, sponsor investment trusts. The miracle of leverage, moreover, made this a relatively costless operation to the ultimate man behind all of the trusts. Having launched one trust and retained a share of the common stock, the capital gains from leverage made it relatively easy to swing a second and larger one which enhanced the gains and made possible a third and still bigger trust … This fiscal incest was the instrument through which control was maintained and leverage enjoyed.

This process involves issuing equity, so at the final step there is a gullible individual investor who is buying an overpriced leveraged closed-end fund. As I discussed here, these trusts arose during a general wave of equity issuance occurring in 1929; this wave was indicative of a stock market bubble.

The “fiscal incest” of 1929 bears little resemblance to the recently announced AI deals. It more closely resembles the leveraged single-stock ETFs of today, indicative of a market swarming with risk-loving retail investors, as I’ve previously discussed. I agree that these ETFs are a sign of market froth, but they have nothing to do with “circular deals.”

Second, the Japanese stock market bubble peaking around 1989 also involved intercorporate equity holdings. As background, Japan has a long tradition of cross-holdings of corporate equity, starting with the zaibatsu in the 19^th century and continuing with the keiretsu after World War II. For example, Toyota is a customer of DENSO, and for many decades both companies were major shareholders of each other. As part of Japan’s recent pro-shareholder reforms, cross-holdings have become less common among Japanese firms.

Chancellor (2000) argues that cross-holdings contributed to the Japanese stock market bubble of the late 1980s. That may be, but I don’t think we can say that cross-holdings caused the bubble. After all, Japan had extensive cross-holdings for more than a century before 1989.

One story about the Japanese bubble involves land prices interacting with equity prices. Banks owned corporate equity and made property loans, while corporations owned land. When the bank owned shares in company XYZ, and the price of XYZ rose, the bank now had more collateral and could make more property loans, which sent land prices higher, which increased the value of XYZ’s land holdings, which sent the share price of XYZ higher.

To me, this feedback loop is more reminiscent of the theory behind today’s Bitcoin treasury firms: they issue shares, use the proceeds to buy Bitcoin, push up the price of Bitcoin, and as a result their shares rise in value. Whether this Bitcoin treasury story is true or false, it’s not relevant for AI.

Last, let’s consider the U.S. tech-stock bubble peaking in 2000. I’ve previously described Shiller’s idea of a “naturally occurring Ponzi scheme” where rising prices attract buyers. Here, Graham (2010) describes a similar feedback loop:

By 1998, Yahoo was the beneficiary of a de facto Ponzi scheme. Investors were excited about the Internet. One reason they were excited was Yahoo's revenue growth. So they invested in new Internet startups. The startups then used the money to buy ads on Yahoo to get traffic. Which caused yet more revenue growth for Yahoo, and further convinced investors the Internet was worth investing in. When I realized this one day, sitting in my cubicle, I jumped up like Archimedes in his bathtub, except instead of "Eureka!" I was shouting "Sell!"

In this example, we don’t have any cross-corporate ownership, but we do have a sort of circularity, similar to unwary travelers who are lost in a snowstorm and start following their own footprints. At first glance, Graham’s “de facto Ponzi scheme” seems fairly relevant to today’s AI deals. But is it really?

Graham’s scenario is not actually circular and self-contained. There are outside investors who are pumping money into the system, giving cash to startups which then give cash to Yahoo. A Ponzi scheme always needs outside investors pumping money into the system. That’s what we are lacking today: outside investors. No outside investors, no Ponzi.

A different feature of the tech-stock bubble was vendor financing. An example is Lucent, the telecom giant that was one of the darlings of the 1990s and at one point America’s most widely held stock. Lucent lent money to its customers in order to boost sales.

Here’s Lowenstein (2005):

The company had to increase its stock price, or employees would depart en masse for Silicon Valley. It had to win friends on Wall Street, or it wouldn’t be able to use its stock to acquire other similarly overpriced firms, as its competitors were doing …

To further inflate sales, Lucent committed a grotesque $8 billion to customer financing. At some point, Lucent wasn’t selling equipment any more; it was giving stuff away and labeling it a sale. 

This revenue-goosing was accompanied by massive issuance by Lucent and similar firms, primarily via stock-financed mergers and stock-based compensation. As mentioned in Endlich (2004), the perception was that “Lucent’s stock is as good as currency.” Lucent spent this currency with abandon. In contrast, we don’t see similar issuance by the Magnificent Seven today; they are mostly repurchasing equity, not issuing equity. Maybe their stock today is also “as good as currency,” but they don’t seem eager to spend this currency.

Snake vs. anglerfish

A stock market bubble is not about the circular flow of money in a closed system. A stock market bubble is about the flow of money from optimistic/gullible investors into the stock market. Ouroboros, the snake eating its own tail, is not a good model for a stock market bubble. A better model is an anglerfish: using a bioluminescent appendage, the anglerfish lures unwary prey and then gobbles it down.

You can’t have an anglerfish without prey. You can’t have a casino without customers. And you can’t have a bubble without inflows.

Today we have an AI boom, but not yet much of an AI bubble. A crypto bubble? Sure. A quantum bubble? Yep. But an AI bubble? Not yet.

Here are the signs and portents that I’ll be looking for:

• A massive wave of IPOs from AI firms.
• AI firms go on an acquisition binge, using their own shares as currency.
• Megacap tech firms stop repurchasing shares.

These elements were all present in 1999/2000 and to a lesser extent in 2021. But right now, I’m not seeing it.

Yes, it’s possible that the world is spending too much on AI, but that might happen with or without “circular deals.” Yes, it’s possible that megacap tech firms are overpriced, but they sure aren’t acting like it, since they’re not selling equity.

Endnotes

[1] References to this and other companies should not be interpreted as recommendations to buy or sell specific securities. Acadian and/or the author of this post may hold positions in one or more securities associated with these companies.

[2] “OpenAI, Nvidia Fuel $1 Trillion AI Market With Web of Circular Deals,” Bloomberg, October 7, 2025.

[3] The New York Times, October 7, 2025.

[4] “OpenAI Is Good at Deals,” Bloomberg, October 6, 2025.

References

Chancellor, Edward. Devil take the hindmost: A history of financial speculation. Penguin, 2000.

Endlich, Lisa. Optical Illusions: Lucent and the crash of telecom. Simon and Schuster, 2004.

Galbraith, John Kenneth. The Great Crash 1929. Houghton Mifflin Harcourt, 2009.

Graham, Paul. What Happened To Yahoo. 2010

Lowenstein, Roger. “How Lucent Lost It.” The MIT Technology Review, February 1, 2005.