AI Circular Deals: Who Pays Whom, and What Shrank Once the Contracts Were Signed

Nvidia announced it would invest up to $100 billion in OpenAI. The finalised investment was $30 billion. That single gap is the most useful thing to understand about AI circular deals, because it shows the difference between a number that moves markets and a number that appears in a contract.

What a circular deal actually is

A circular deal is an arrangement where the money moves in a loop between parties who are simultaneously investor, supplier and customer to each other. A chipmaker invests in a model developer, the model developer buys chips, and the purchase is booked as the chipmaker's revenue.

Nothing about that structure is unusual on its own. Strategic investment alongside a commercial relationship is ordinary corporate behaviour. It becomes worth examining when the loop is large enough that one party's revenue growth depends materially on capital it supplied itself. Scale changes the question. A $50 million strategic investment alongside a supply deal is housekeeping. A commitment measured in tens of billions, recorded as revenue by the investor, is a different object and deserves a different level of scrutiny.

The critique is about revenue quality, not legality. If a supplier funds its own demand, reported growth can look stronger than underlying customer demand actually is. Nothing in accounting standards requires a company to separate the two, and none of the parties here has volunteered to.

The loops worth knowing

Nvidia and OpenAI

In September 2025 the two companies announced a letter of intent covering at least 10 gigawatts of Nvidia systems, with Nvidia intending to invest up to $100 billion in OpenAI as the systems were deployed. The announcement moved markets immediately.

By February 2026 the finalised figure was $30 billion, taken as part of a larger OpenAI funding round. In March, Jensen Huang said publicly that the $100 billion outcome was probably not in the cards, citing OpenAI's move toward a public listing.

Microsoft, Nvidia and Anthropic

In November 2025 Nvidia and Microsoft announced investments of up to $10 billion and up to $5 billion respectively into Anthropic. The same agreement included Anthropic committing to purchase $30 billion of Microsoft cloud capacity.

Satya Nadella described the arrangement as the companies increasingly becoming customers of each other. That is an unusually plain description of the structure, and it is worth taking at face value rather than treating as spin.

The cloud commitments

OpenAI has announced multi-year purchase commitments across several cloud providers, some of whom are also investors. These commitments are the largest numbers in the entire AI economy and they run for a decade or more.

A ten-year purchase commitment is not a ten-year certainty. Read the term length before reading the headline value, because a $250 billion commitment across ten years is a very different object from $250 billion of orders this year. Divide by the term and compare the annual figure to the buyer's current revenue. That single arithmetic step reframes most of these announcements.

The same logic applies to chip supply arrangements. A commitment to deploy a stated volume of hardware over several years is a plan, and plans of that size are routinely renegotiated when demand or pricing moves.

Announced against signed, in one table

Most coverage reports the announcement and never revisits it. Bloomberg has maintained a running graphic of these arrangements since early 2026. This is the follow-up.

Every row uses the parties' own announcements and subsequent public commentary. Where a figure is reported rather than confirmed by the company, the wording says so.

What the filings said

Nvidia's quarterly filing addressed the OpenAI arrangement directly in its risk disclosures, stating there was "no assurance" that the investments would be completed on the expected terms. The same language covered its planned Anthropic and Intel investments.

That is the company telling you, in the document that carries legal weight, not to treat the press release as settled. It is the most reliable sentence in this entire story and almost nobody read it.

The CFO said the same thing out loud. Speaking at a technology conference in December 2025, Colette Kress confirmed the companies were still working toward a definitive agreement, and that Nvidia's disclosed chip bookings did not include any orders tied to the arrangement. The market had already priced the announcement as though they did.

The case for the defence, stated properly

The circularity critique is often presented as self-evidently damning. It is not, and the counter-argument is serious.

Supply lock-in is rational when supply is scarce

Advanced accelerators have been hard to obtain for three years. In that market, buyers do not simply place orders. They pair long-term purchase commitments with financing to secure a place in the queue.

Asset managers have described the arrangements as lining up suppliers, builders and customers to meet demand that genuinely exists. On that reading the loop is coordination, not inflation.

The demand underneath is real

Both major labs are growing revenue at rates with no software precedent. Whatever you think of the financing structures, the underlying product demand is not imaginary, which is examined in more detail in the piece on where AI revenue and return actually land.

Where the defence gets thin

The strongest version of the critique is not that the deals are fake. It is that they make revenue harder to interpret. When an investor is also a customer, an outside analyst cannot easily separate demand from subsidy, and the companies do not disclose enough to do it for them.

What would actually make this dangerous

The vendor financing precedent

In the telecom build-out of the late 1990s, equipment suppliers lent customers the money to buy equipment. Revenue looked strong until the customers could not pay, and the losses landed back on the suppliers.

The parallel is imperfect. Today's investors are equity holders in profitable businesses rather than lenders to unprofitable start-ups. The mechanism to watch is the same one, which is whether the buyer can pay without the seller's help.

Concentration

A small number of parties appear on multiple sides of multiple deals. That concentrates counterparty risk in a way no single company's disclosures will show you.

An investor reading one annual report sees one company's exposure. Nobody publishes the consolidated view, and no regulator currently requires it. That gap is the strongest technical argument the sceptics have, and it is separate from any claim about whether demand is real.

What would change my reading. Three things. Announced deals repeatedly shrinking or lapsing on execution. Purchase commitments being renegotiated downward. Financing shifting from equity toward debt secured against the assets being purchased. None of those is a forecast. They are observable events, and you can check for them each quarter.

Why this matters more in 2026 than it did in 2025

In 2025 these arrangements were announcements. In 2026 they are contracts with delivery schedules attached, and the difference changes what you can learn from them.

Three things shifted. Several of the largest frameworks reached the point where a definitive agreement was either signed or quietly reduced, which gave the market its first real test of announcement quality. Both major labs moved toward public listings, which forces disclosure that private companies never had to provide. And the funding mix started moving, with equity raises and structured debt appearing alongside operating cash flow.

That last shift is the one to track. When a build is funded from operations, the funding source tells you nothing. When it is funded from raised capital, the terms of that capital become a live constraint on the build itself.

There is also a disclosure effect worth naming. A private company can describe an arrangement in a press release with no obligation to revisit it. A company preparing to list has to describe the same arrangement in a filing, under a different standard, and those two descriptions have not always matched.

How to read the next AI mega-deal announcement

The next one is weeks away. Run these five checks before forming a view.

What this changes if you are just buying software

Very little in the short term, and something specific in the medium term.

The short-term answer is that your token prices and licence costs are set by competition, not by these structures. The medium-term answer is that if capital markets tighten, the subsidy inside cheap inference tightens with them.

So the practical move is to know what you would do if the compute inside your stack cost 40% more. Not because it will, but because a vendor who cannot answer that question has not modelled their own business. The wider spending picture behind that question is covered in the breakdown of the $725 billion 2026 build.

Frequently asked questions

Where to start this week

Two concrete actions.

First, take the last AI partnership announcement you saw and run the five checks in the table above. Look specifically for "up to", for the term length, and for whether the investor is also the supplier.

Second, ask your two largest AI vendors one question in writing. What happens to our price if your inference costs rise 40%? The answer, or the absence of one, tells you how much of your cost base is currently being subsidised by capital markets.

Related reading