AI Circular Deals Mapped: 10 Arrows Between Nvidia, OpenAI, Microsoft and AMD

Microsoft booked $10.0 billion of revenue from OpenAI in the quarter to 30 September 2026, and it has funded $12.0 billion into OpenAI in total. That pair of numbers breaks the simple story about AI circular deals. We mapped 10 arrows between Nvidia, OpenAI, Microsoft and AMD. Only 3 carry a cash figure.

The short answer. AI circular deals are real, but most of the famous arrows are not cash. Microsoft's OpenAI revenue is genuine and exceeds what Microsoft put in. The circularity sits one level up, where OpenAI's suppliers supply much of its capital. Sort every arrow by instrument before you call it revenue.

Reader for this post: founders and directors, because the question is revenue quality, and that reader has to price vendors whose financial health rests on these loops.

The AI circular deals map: 10 arrows, sorted by what moved

The map sorts 10 arrows into 4 classes, and the class matters more than the amount. Most diagrams of this market draw every flow in the same ink, so a signed contract, a contingent guarantee and a cash wire look identical. Our earlier explainer of how circular deals work tracked what shrank after signing. This one asks a narrower question about each arrow: what is it today?

The scope is deliberate. It covers the four companies in the usual argument, plus Oracle and SB Energy because OpenAI's capacity runs through them. Anthropic, CoreWeave and the wider web of AI clouds would add arrows, not change the method. Save the figure and use it in your next argument. The source line is printed on it for that reason.

Three AI circular deals arrows with a cash figure

Only 3 arrows combine a dollar amount with money that has actually moved. These are the only ones you can use without a qualifier about future performance, and even they differ in quality.

Arrow 1: Nvidia's $30 billion in OpenAI. Bloomberg reported that Nvidia put $30 billion into OpenAI's $122 billion round, and this is the weakest of the 3 on sourcing. Nvidia's filings report private stakes only in aggregate, so the figure rests on press reporting. Our filing-by-filing trace of the Nvidia commitment covers how the $100 billion headline became this number, so we will not repeat it here.

Arrow 2: Microsoft's $12.0 billion. Microsoft has funded $12.0 billion of a $13.0 billion commitment and holds about 24% of OpenAI on an as-converted basis. That is the entire cash cost of the most valuable position on this map, and it has been spent over years. Its accounting result lives below operating income, as our reading of the equity-method line sets out.

Arrow 3: OpenAI's $10.0 billion to Microsoft. This is the best-documented arrow in the whole market, because accounting rules forced it into the open. OpenAI is an equity-method investee, which makes it a related party under ASC 850, so Microsoft has to disclose the revenue and the receivable. Microsoft was still owed $7.2 billion by OpenAI at quarter end, up from $6.0 billion three months earlier. Nvidia and AMD have no equivalent disclosure, and that absence is why this arrow gets more scrutiny than larger ones.

Microsoft's OpenAI revenue is real, and not self-funded

Microsoft's OpenAI revenue cannot be a simple round trip, because the amounts do not allow it. A round trip needs the money going out to be about as large as the money coming back. Here it is not close, and the working below uses only filed numbers plus one press figure, which is labelled.

Why the 80% is the number to watch

Annualise Microsoft's quarter and you get $40 billion against the roughly $50 billion annualised revenue the Financial Times reported for OpenAI. That is crude, because one is a quarter's accounting revenue and the other is an end-of-period run rate. The direction still holds. One supplier is taking something like four-fifths of what OpenAI earns, before Oracle, AWS, chips or payroll. So OpenAI's customers alone are not paying this bill, and the balance has to come from investors.

Part of this arrow is a royalty, not compute

The $10.0 billion is stated "inclusive of revenue-sharing payments", and Microsoft does not split it. That matters because a revenue share is paid out of OpenAI's sales, while Azure spending is a cost of producing them. Under the April 2026 amendment, OpenAI keeps paying through 2030 at the same percentage, subject to a total cap, while Microsoft no longer pays OpenAI at all. Neither the rate nor the cap appears in the reporting we opened. I would treat anyone quoting a precise Azure-only figure from this disclosure with suspicion.

Who funds OpenAI: the vendor financing test on the $122B round

The circularity sits in OpenAI's funding, not in Microsoft's revenue line. Classic vendor financing means a supplier lends a customer the money to buy its goods. OpenAI's round is looser than that, but the same question applies: how much of the capital that pays suppliers came from suppliers?

Three versions of the same ratio

Choose your definition and state it, because the number moves. Amazon's $50 billion and Nvidia's $30 billion make $80 billion of $122 billion, or 66%. Bloomberg reports that $35 billion of Amazon's money depends on an IPO or a technology milestone, and removing it leaves $45 billion of $87 billion, or 52%. Reuters describes SB Energy as SoftBank-owned, so adding SoftBank's $30 billion takes it to $110 billion, or 90%.

My position is that 52% is the honest headline, because uncalled money cannot pay a bill. It is also high enough to make the point. Over half of the firm capital in OpenAI's largest round came from two companies that sell it compute. No rule requires that money to return to them, and none of the sources we read shows a contractual tie. Still, the commercial relationships are not hidden.

Two backlog arrows: $550 billion that is not yet revenue

Arrows 4 and 5 are the largest on the map and the least proven, because a remaining performance obligation records a contract, not a delivery. Remaining performance obligation, or RPO, is revenue a company has contracted but not yet recognised. These two contracts total $550 billion, and our ledger of OpenAI's compute commitments compares the wider set with its revenue.

Arrow 4: Azure, and a 10-Q that does not say. OpenAI contracted to buy an incremental $250 billion of Azure services in October 2025. Microsoft's commercial RPO stood at $677 billion in September 2026, with about 30% due within 12 months. The 10-Q note does not give an OpenAI share. The last figure we could find is CNBC's recollection of about 45% of $625 billion in January, which works out to roughly $281 billion. Microsoft's CFO said on the July call that RPO grew 25% excluding OpenAI, so the rest of the book is growing as well.

Arrow 5: Oracle, where concentration is sharper. Bank of America analysts put OpenAI at more than 50% of Oracle's $638 billion RPO, per SiliconANGLE, which implies over $319 billion on our arithmetic. That is an analyst estimate. Oracle also said most of the increase came from contracts where customers prepaid for GPUs or supplied them. That detail is the part to keep, because it moves who funds the hardware. Our Oracle backlog analysis takes the credit side further.

Two contingent arrows: Nvidia's guarantee and AMD's warrant

Contingent arrows are not cash today, and both still shape behaviour. Each one gives a supplier's balance sheet or equity to OpenAI in exchange for a commercial position, without sending money at signing.

Arrow 6: Nvidia's $105 billion Ohio guarantee. Nvidia's guarantees are capped at $105 billion and cover about 4.25 GW of OpenAI leases at SB Energy's Ohio campus. They start phase by phase, with the first expected in Nvidia's fiscal 2029. They pay only on certain tenant defaults and end if OpenAI reaches a satisfactory credit rating. In return, Reuters reports, Nvidia is the exclusive chip provider on the site and invests $1.5 billion in SB Energy. Jensen Huang told reporters the deal "was not circular financing". On the narrow definition he is right, since no cash goes to OpenAI. I think this is still the tightest link on the map, because the supplier underwrites the customer's rent and gets the chip orders.

Arrow 7: AMD's warrant reached the income statement. AMD's warrant moved from footnote to expense in the September 2026 quarter. AMD booked a $1,097 million charge for the change in fair value of its warrant liability in other income, and excluded it from non-GAAP results. AMD's policy records grant-date fair value as a reduction to revenue, and later changes in other income. The 10-Q pages we could read did not state the revenue reduction. The cash flow line shows $1,179 million of non-cash warrant impact, $82 million above the fair value change. If all of that residual were the revenue reduction, it would be 0.6% of the quarter's $13,377 million revenue. That is our inference, not AMD's disclosure.

The shares remain contingent. AMD's EPS note excludes them until all conditions are met, and CFO Jean Hu said AMD is not expecting fourth-quarter share count to be affected. Our breakdown of the warrant's vesting terms covers the tranches and the dilution maths.

Three chip arrows with no dollar figure

Arrows 8, 9 and 10 are the sales the whole argument is about, and no filing we read puts a per-customer number on them. Everyone arguing about Nvidia's revenue quality is inferring this part.

What the filings do say

Nvidia's quarter to July 2026 shows $89.0 billion of Data Center revenue, one direct customer at 16% of total revenue and five at 70% of receivables. It lists AI model makers among both direct and indirect customers without quantifying them. AMD's CEO named Anthropic, Meta, Microsoft, OpenAI and Oracle as lead customers bringing Helios racks online, and AMD's data centre revenue reached $8.5 billion in the quarter. Microsoft's additions to property and equipment were $39.4 billion in the September quarter, so OpenAI's $10.0 billion equals about 25% of that outflow.

A ceiling test for Nvidia's revenue

You can bound the question without knowing the split. Nvidia's private equity holdings show $31.0 billion of net additions in six months, against $177.8 billion of revenue, or 17.4%. Suppose every dollar came back as Nvidia orders in the same period, which no filing claims and which is implausible. Over 82% of revenue would still have no link to those stakes. That is a useful marker, not a clean bound, since it leaves out public stakes and the timing is loose. Our note on how auditors treat these stakes explains why most of this never reduces revenue.

The loop almost nobody draws

The most literal circular language in these filings has nothing to do with OpenAI. Nvidia reports $36 billion of AI cloud agreements, where AI clouds buy its data centre products and Nvidia commits to buy cloud services from them. The clouds can stop providing that capacity and sell it to third parties at better rates, and Nvidia's commitment falls as they do. A separate $29 billion of cloud service agreements supports Nvidia's own research. AMD reports guarantees of up to $8.5 billion, with no counterparties named. If I were arguing this on X, I would post this paragraph instead of the OpenAI diagram.

Where this map is weakest

This map has three real weaknesses, and the third one cuts against its own framing.

The arrows are incomplete by design

Leaving out Amazon's own arrows, Anthropic, CoreWeave and Meta makes the map legible and also makes it partial. Amazon appears in the funding bar but not as an arrow, despite its compute contract with OpenAI. Reasonable people would draw this differently. The number 10 is our count, not a fact about the market, and adding Amazon would make it 12.

Several anchor figures are press figures

Nvidia's $30 billion, Amazon's split, the Oracle contract size, the BofA share and OpenAI's run rate all come from reporting. We read Bloomberg's article only up to its paywall, and we could read just the opening of the FT piece. Our 80% ratio inherits that weakness and compares unlike measures. Treat it as a direction, and check it against an audited figure if OpenAI's listing documents ever provide one.

The strongest case against calling this circular

Supply is scarce, and pairing capital with capacity is a normal way to secure it. Microsoft's case is the awkward one for critics: its returns from OpenAI's spending far exceed its cash in, and its backlog is growing without OpenAI. If OpenAI keeps raising and customers keep paying, every arrow here simply looks like an aggressive supply chain. The loop becomes a problem only if outside capital stops, and this map cannot tell you when or whether that happens.

Frequently asked questions

Where to start

Next time a circular deals chart lands in your feed, redraw one arrow before you share it. Pick the largest number on it and write down its class: cash, backlog, contingent or undisclosed. Most of the time the largest number will not be cash, and that one label changes the argument you are having.

If you depend on one of these vendors, set a single reminder for Microsoft's next 10-Q and read Note 1. The OpenAI revenue and receivable sit in one paragraph. If the receivable keeps growing faster than the revenue, that tells you more about OpenAI's cash position than any announcement this year. Nvidia's next filing, due later this month, is the other date to keep.

Related reading