The Investors Who Own Both OpenAI and Anthropic

I spent Sunday cleaning the house with music on. But something was nagging at me, and at some point I started researching a field that is fairly alien to me: the finance of shareholder registers and balance sheets. Three executives who avoid even shaking hands had asked, on the same weekend, for the development of artificial intelligence to slow down. In my line of work, when three components that never talk to each other change state at the same instant, you look for a shared signal upstream before you believe in coincidence. This is what came out of it.

The shared signal has a name, several in fact. In March Amazon pledged up to $50 billion to OpenAI; in April it put another $5 billion into Anthropic. Nvidia committed $30 billion to OpenAI and up to $10 billion to Anthropic. Microsoft is OpenAI's longest-standing backer and has held a stake in Anthropic since November 2025. MGX, Abu Dhabi's sovereign fund, and the investment firms D.E. Shaw and T. Rowe Price appear in the latest rounds of both. For a decisive share of their capital, two companies at war belong to the same portfolio.

Economists call this common ownership. José Azar, Martin Schmalz and Isabel Tecu studied it in US airlines, where the same large funds hold stakes in every carrier flying a given route, and found ticket prices higher than competition would predict. Their work is debated and contested, but the mechanism is simple: a shareholder who owns both competitors pays for their war twice. A spending race in which each company invests so as not to fall behind is exactly that war. If the portfolio could speak, it would tell both of them to ease off.

The stake is the least interesting thing these backers hold. Alongside its $50 billion, Amazon secured from OpenAI a commitment to spend $100 billion over eight years on AWS, its cloud: it is reasonable to assume that spending is the return that counts. Much of Nvidia's $30 billion in OpenAI arrives as computing capacity, which means Nvidia chips; in all likelihood Nvidia is buying the assurance that the labs stay on its architecture while Google and Amazon push processors of their own. Microsoft secured from Anthropic a commitment to buy $30 billion of capacity on Azure, and a second model supplier for its own products after years of depending on OpenAI. Samsung, SK hynix and Micron make the memory every data centre consumes, and they have bought into Anthropic. A supplier that becomes a shareholder in its customer presumably wants that customer to keep buying.

This is where the doubt gets complicated. A shared brake suits whoever owns the labs, because it lowers spending. It hurts whoever sells to them, because it lowers orders. SoftBank has pledged OpenAI $30 billion in three tranches, plus $3 billion a year to roll out OpenAI's products across the companies in its own portfolio: it buys what it finances. On the Monday after the call to slow down, its shares fell 11%. The same backers sit on both sides of the brake, and the balance for each depends on how much of its position is equity and how much is supply.

Then there is Elon Musk, who answered the call with "Dario is right". According to the listing document of SpaceX, Musk's company, which has since absorbed his AI lab xAI, Anthropic pays it $1.25 billion a month until 2029 for the Colossus data centres. That comes to about $15 billion a year, close to what SpaceX earned in an entire year, and SpaceX is preparing a listing of its own. The simplest hypothesis is that a supplier agreed with its biggest customer. The same supplier also runs a direct competitor, which makes his the hardest position of the week to read.

For sovereign funds, a stake is also a seat in the room. Besides MGX, Anthropic's shareholders include GIC and Temasek, Singapore's state funds. Washington decides who may receive the most advanced chips; being a shareholder in the labs means sitting where their distribution gets decided. It remains a hypothesis, but it is the cheapest one for explaining why a state would want to own both rivals. The large asset managers (Fidelity, Capital Group, Baillie Gifford) are presumably after something else: priority on listing day, which goes to those who were there first.

Anthropic's accounts are the part of this story that deserves the closest reading. According to Bloomberg, in August the company told prospective investors it had made $11.5 billion in revenue in the second quarter and the first operating profit in its history. Ed Zitron, who often finds excellent leads on tech policy, had already called it an accounting swindle in May, when the Wall Street Journal was only forecasting it. The words are his, but the footholds he pointed to can be checked. The profit is "adjusted", meaning it excludes items the company chooses to leave out and has not disclosed. The figures are preliminary and were shown precisely to the people being asked to buy the shares. And the quarter coincides with reduced payments on the SpaceX contract: Axios confirms that in May and June, while the arrangement ramped up, Anthropic paid less. Zitron, who has been questioning for more than a year how these labs count revenue, adds token credits sold in advance at discounts of 10 to 30%. According to analysts at Futurum, Anthropic also books revenue on a gross basis, differently from OpenAI.

None of these choices is irregular. Taken together, they sketch a quarter built for the prospectus, the document a company files to go public, which in the United States is lodged with the SEC, the market regulator. Anthropic aims to list as early as October. A coordinated slowdown makes that quarter repeatable: if nobody accelerates, nobody has to spend to catch up. Discipline arrives with the prospectus.

For OpenAI the calendar is tighter. According to SaaStr, of the $122 billion announced in March only about thirty was cash paid in at closing. The rest was deferred tranches, Nvidia chips and $35 billion from Amazon conditional on a listing or a technical milestone. OpenAI even guaranteed a group of private equity firms a minimum return of 17.5%. Meanwhile its listing has slipped beyond 2026, and an internal presentation reported by the Financial Times projects $278 billion of negative cash flow through 2030. A company that burns cash and needs time finds in a shared brake the only way to spend less without losing ground.

The state, by contrast, weighs little in the accounts. The contracts the Pentagon signed in July 2025 were worth up to $200 million per company. But in February Anthropic refused to drop its bans on domestic mass surveillance and fully autonomous weapons, and it was shut out of the entire federal government until a judge overturned the ban in August. OpenAI signed with the Pentagon within hours. A customer that buys little and can close off the whole public market acts as referee, and for a company preparing to list, as guarantor of last resort.

Of course, the stated fear may be sincere: anyone who has watched a swarm of agents escape its enclosure has good reason to want to go slowly. The same brake, though, also serves the portfolio, and the two motives live inside the same signature. What my Sunday doubt produced is a map of who gains from which speed.

That map has one last name. After the listings, the risk now carried by informed funds and suppliers will pass to index funds, meaning the savings of people who own the whole market, where a share is not a say, and to the $3 billion OpenAI has already raised from private savers through banks. Discipline arrives with the prospectus, and whoever reads it last is also whoever pays if the revenue arrives after the $278 billion.

Do you think I settled my doubt?