In a circular AI deal, a chipmaker or cloud provider puts money into an AI company, and that company spends the money buying the investor's own chips or cloud capacity. Nvidia, OpenAI, Oracle, AMD and Microsoft have signed hundreds of billions of dollars in these loops since September 2025.
This post sticks to the plumbing — who paid whom, how much, and what actually happened to each deal. It isn't a call on any stock.
One thing to clear up first, because the headline numbers move. The $100 billion Nvidia–OpenAI figure quoted everywhere in late 2025 never landed as written, and the Nvidia stake in CoreWeave that a few aggregator sites still print as 91% is about 11.5% in the SEC filing. Every figure below is as of September 2026.
circular AI deals explained — money and chips moving in a loop between Nvidia, OpenAI, AMD and MicrosoftWhat are circular AI deals, exactly?
A circular AI deal is one where the supplier funds its own customer. The chip company or hyperscaler takes an equity stake, extends credit, or hands over stock warrants, and the AI lab commits to years of buying that same supplier's chips or computing power. Cash leaves by one door and comes back in as booked revenue.
Nothing about that is exotic on its own. Suppliers have financed customers for as long as suppliers have existed.
What makes analysts uneasy, as J.P. Morgan Asset Management lays out, is scale plus opacity. When enough of the demand is funded by the people booking the revenue, it gets hard to tell how much of the AI boom is customers spending their own money.
The comparison people reach for is the telecom bust. One retrospective of that era describes Nortel lending more than $7 billion to startup carriers, much of it unsecured and tied to future equipment orders, with Lucent committing $8.1 billion on similar terms — and Nortel eventually filing for bankruptcy when the loans soured. That account comes from a single secondary source, so treat the exact figures as an approximation rather than a filing.
OpenAI's CFO, Sarah Friar, has been quoted saying that "most of the money will go back to Nvidia." The primary interview transcript isn't publicly confirmed, so the quote is worth knowing and not worth leaning on.
Circular AI deals explained deal by deal: who paid whom
Four arrangements carry most of the weight, plus Nvidia's stake in CoreWeave sitting alongside them. Analysts have been quoted putting the full web of criss-crossing investment-and-purchase agreements above $800 billion, though that total is an aggregate estimate rather than a number pulled from any single filing.
| Deal | Announced | Size and structure | Where it stands (Sept 2026) |
|---|---|---|---|
| Nvidia → OpenAI | Sept 22, 2025 | Letter of intent for at least 10 GW of Nvidia systems, with up to $100B invested as each gigawatt came online | Never signed as written. Nvidia's filing reported Nov 19, 2025 said there was "no assurance" of a definitive agreement; a $30B stake in OpenAI's March 2026 round took its place |
| AMD → OpenAI | Oct 6, 2025 | Up to 6 GW of Instinct MI450 GPUs, starting with 1 GW in 2026, plus a warrant on up to 160 million AMD shares at $0.01 each (roughly 10%), vesting with deployment and with AMD's stock climbing as high as $600 for the last tranche | Warrant tranches still tied to deployment; the first gigawatt is due in 2026 |
| Oracle ↔ OpenAI | Signed Sept 2025, starts 2027 | $300B of cloud over five years, about 4.5 GW of power — roughly what 4 million US homes draw — averaging $60B a year from a company with around $10B of annual revenue at signing | Running; about half of Oracle's contracted backlog traces to OpenAI |
| Microsoft + Nvidia → Anthropic | Nov 2025 | Microsoft up to $10B and Nvidia up to $5B at a valuation near $350B, with Anthropic committing to buy $30B of Azure compute built partly on Nvidia chips | The newest of the group, and a three-way version of the same loop |
Oracle is where the loop shows up most plainly in an audited number. Its remaining performance obligations — contracted work it hasn't billed yet — hit $455 billion at the end of the quarter ending August 31, 2025, up 359% year over year, per Oracle's own results release.
By mid-2026 that backlog was reported in the $638–664 billion range. Roughly $300 billion of it has been attributed to OpenAI alone.
If we had to pick one line to watch, it's that last one. A backlog is a promise to pay, and this one leans on a single customer whose spending commitments have run well ahead of its revenue.
Is any of this illegal?
No. Investing in a customer is legal, and the large deals were announced publicly. The live question is disclosure. On May 7, 2026, SEC deputy chief accountant Sheri York said investors benefit when they can understand how these financing transactions connect, pushing for clearer reporting of the structures.
She has company. A Morgan Stanley report from April 2026 found that most of more than $1 trillion in hyperscaler investment commitments wasn't fully reflected on financial statements.
The Bank for International Settlements went further. Its annual economic report, released June 28, 2026, named an AI capex bust, opaque circular financing, and record sovereign debt as the three largest risks to global financial stability — and warned that a repricing could hit credit markets with force comparable to 2008.
The other side of the argument deserves air, and we think it's stronger than the bubble headlines allow. J.P. Morgan Asset Management points out that most hyperscaler AI spending comes from strong internal cash flow rather than piles of new debt, that AI-related deals now account for about one of every two dollars of US venture funding, and that North American data center vacancy hit a record low of 1.6% in the first half of 2025.
Empty capacity is not the problem right now. Demand is real. Whether it stays real at $60 billion a year is the unanswered part.
Our read: the structure isn't the scandal — the accounting is. A deal that shrinks from $100 billion to $30 billion in six months tells you these are intentions until they're contracts.
Related: This is part 2 of our look at the AI chip trade. Part 1 covers AMD's $1 trillion week and the bubble debate.
At a glance
- Who it's for — anyone who keeps seeing these deals in headlines and wants the actual structure, not a stock tip.
- The shape: supplier invests in customer, customer buys supplier's chips or cloud, revenue loops back.
- Numbers that matter as of September 2026: Nvidia's $100B intent became a $30B stake, Oracle–OpenAI runs $300B over five years, Oracle's backlog grew from $455B to the $638–664B range, and the whole web is estimated above $800B.
- The one caution: it's legal, but the SEC and the BIS have both said publicly that the disclosure around it is thin.