In September 2025, Nvidia announced a letter of intent to invest “up to $100 billion” in OpenAI, released progressively as OpenAI deployed each gigawatt of Nvidia systems. OpenAI is one of the largest buyers of Nvidia hardware on the planet, so the shape of the arrangement was hard to miss: money would leave Nvidia as an investment and come back as chip revenue. The label “circular deal” stuck almost immediately, and it has since become shorthand for a whole category of AI-era financing.
What happened next is just as instructive as the announcement. By early 2026 the Wall Street Journal reported that the $100 billion framework had stalled without a signed contract, and CNBC’s follow-up reporting described a relationship that had cooled but could not actually uncouple. What reportedly replaced the framework was a $30 billion direct equity stake, taken as part of OpenAI’s roughly $110 billion funding round in early 2026 — smaller, simpler, and no longer tied to deployment milestones, but still a chipmaker owning a slice of its own biggest customer.
This post explains the mechanic in plain terms, walks through the deals that are actually on the record as of August 2026, and covers why companies structure financing this way, why it makes investors nervous, what the 1990s telecom version of this movie looked like, and which disclosures let you check the facts yourself. It is an educational explainer, not advice.
What vendor financing means, in plain terms
Vendor financing is any arrangement where a supplier helps fund the customer who buys from it. The classic form is credit: the vendor lends the customer money, or ships product against an IOU, and books the sale as revenue while carrying the loan as an asset. The AI-era form is usually equity or something equity-adjacent: the vendor invests cash in the customer, grants it cloud credits, or guarantees its obligations, and the customer — flush with that capital — turns around and spends heavily on the vendor’s products.
The loop has three steps:
- Vendor puts capital into the customer (equity, credits, loans, or guarantees).
- Customer commits to buy the vendor’s chips, cloud capacity, or systems — often in the same announcement.
- Vendor books real cash revenue, part of which was funded by its own balance sheet.
Two things are worth stating plainly. First, none of this is hidden or illegal: these deals are announced in press releases and disclosed in filings, which is exactly how we know about them. Second, the revenue is real — actual cash changes hands for actual hardware that actually runs. The concern is not fabrication. It is that some fraction of the demand was financed by the seller, which makes the demand signal harder to read from outside.
The deals on the record, as of August 2026
Every row below was checked against primary announcements and reporting as of August 2026. This map changes fast — the Nvidia and OpenAI arrangement was restructured between its announcement and this writing — so treat dates as load-bearing.
| Arrangement | Structure | The circular part |
|---|---|---|
| Nvidia and OpenAI | “Up to $100 billion” letter of intent (Sept 2025); reportedly replaced by a $30 billion equity stake (early 2026) | OpenAI is among Nvidia’s largest chip customers |
| Microsoft and OpenAI | $13.8 billion invested since 2019; restructured Oct 2025 into a roughly 27% stake | OpenAI committed to purchase an incremental $250 billion of Azure services |
| Nvidia, Microsoft, and Anthropic | Nvidia up to $10 billion, Microsoft up to $5 billion (Nov 2025) | Anthropic committed to $30 billion of Azure compute, running on up to a gigawatt of Nvidia systems |
| Amazon and Anthropic | $8 billion through 2024, plus $5 billion and up to $20 billion more tied to milestones (Apr 2026) | Anthropic committed to over $100 billion of AWS spend across ten years, much of it on Amazon’s Trainium chips |
| AMD and OpenAI | Warrant for up to 160 million AMD shares at $0.01, vesting as OpenAI deploys 6 GW of AMD GPUs (Oct 2025) | The customer earns equity in the vendor for buying — the loop runs in reverse |
| Nvidia and CoreWeave | Equity stake plus a $6.3 billion backstop, disclosed in a securities filing in late 2025 | Nvidia agreed to buy CoreWeave’s unsold cloud capacity through 2032 — capacity that runs on Nvidia GPUs |
A few of these deserve a closer look. AMD’s press release spells out that the warrant tranches vest on both deployment volume and AMD share-price milestones — the customer’s payoff is literally indexed to the vendor’s stock. The three-way Microsoft, Nvidia, and Anthropic partnership compresses the whole pattern into one announcement: two suppliers invest up to $15 billion combined in a lab that simultaneously commits $30 billion to one supplier’s cloud running the other supplier’s chips. And Anthropic’s April 2026 agreement with Amazon pairs fresh investment with a ten-year, $100 billion-plus AWS commitment. Beyond the equity deals sits a lattice of enormous commercial commitments — OpenAI’s reported $300 billion, five-year Oracle contract being the largest — that are not vendor financing but bind the same handful of companies to each other’s fortunes.
Why vendors finance their customers
The frontier labs need staggering amounts of capital: training runs, inference fleets, and data-center commitments run far ahead of their revenue, real and fast-growing as that revenue is. The suppliers, meanwhile, are among the most cash-rich companies in history. The deals exist because each side has what the other lacks, and because the vendor gets several things a normal sale does not provide:
- Locked-in demand. Multi-year purchase commitments convert uncertain future orders into contracted backlog.
- Capacity and allocation alignment. The customer gets supply certainty for scarce chips; the vendor gets a customer with committed reasons not to defect to a rival’s silicon.
- Strategic position. An equity stake in a frontier lab is a claim on the upside of the very demand the vendor is enabling — and a seat near decisions about future architecture.
- Ecosystem defense. Financing that keeps a customer building on your platform is also financing that keeps them off someone else’s.
The AMD warrant shows the same logic from the customer’s side: a buyer big enough to move a vendor’s stock price can demand a share of that upside as the price of committing.
Why it spooks investors
Three concerns come up in nearly every skeptical analysis. The first is revenue quality. If a chipmaker invests $10 billion in a lab that then spends billions on its chips, the revenue is real cash — but the growth rate investors are paying a premium multiple for was partly funded by the company itself. The honest question is not whether the revenue exists but what fraction of it is self-financed, and that fraction is genuinely hard to compute from outside.
The second is demand-signal distortion. Order backlogs and capacity commitments are the market’s main telescope into how much AI demand actually exists. Vendor financing fogs the lens: a gigawatt commitment from a customer the vendor funds is a weaker signal than the same commitment from a customer spending only its own money. When the FTC studied the cloud-provider and AI-developer partnerships in a January 2025 staff report, it documented how the big cloud investments came bundled with equity rights, revenue sharing, and exclusivity provisions — and that a meaningful part of the invested value returns to the investors as compute spending.
The third is correlated downside. A vendor that finances its customers holds two exposures to the same event: if the customer stumbles, the vendor loses future orders and the value of its stake or receivable. The CoreWeave backstop is the starkest version — Nvidia has agreed to become the buyer of last resort for unsold capacity on its own GPUs. In a downturn, deals like that turn a demand problem into a balance-sheet problem at the exact moment both hit at once.
The precedent: Lucent, Nortel, and 1999
This has all happened before, with debt instead of equity. In the late 1990s, upstart carriers — the competitive local exchange carriers, or CLECs — raised on the order of $82 billion to build networks, and the equipment giants competed for their orders by lending them the money to buy. A detailed post-mortem in American Affairs notes that by 2000 Lucent had extended about $1.5 billion in vendor financing to its customers, with Nortel close behind at roughly $1.4 billion. Total customer-financing commitments ran far higher — Lucent’s reached around $8.1 billion and Nortel’s about $3.1 billion. The sales those loans enabled were booked as revenue and celebrated as growth.
Then the customers died. When the CLECs collapsed in 2001 and 2002, the receivables collapsed with them: Lucent took provisions for bad customer debt of roughly $2.2 billion in 2001 and another $1.3 billion in 2002, on top of an accounting scandal and a revenue restatement. The company that had been the most widely held stock in America nearly failed, survived only via merger, and Nortel eventually went bankrupt outright. The mechanism matters more than the drama: vendor financing had converted uncertain demand into reported revenue up front, and into credit losses later. Growth that the vendors appeared to be observing was, in part, growth they were funding.
How 2026 differs from 1999 — and how it rhymes
The differences are real. The biggest buyers of AI infrastructure today are the hyperscalers, which fund most of their capex from enormous operating cash flows rather than debt — the numbers are in this site’s post on AI capex and the stock market. The financed customers, OpenAI and Anthropic, have real and rapidly growing revenue, unlike the largely pre-revenue CLECs. And equity stakes fail differently than loans: a written-off investment caps the loss at the amount invested, with no lingering receivable to restate.
The rhymes are also real. The labs burn cash at historic rates and, per reporting on their own projections, do not expect to be cash-flow positive for years. The neocloud layer — CoreWeave and its peers — is substantially debt-financed against GPU collateral that depreciates far faster than 1990s fiber did. The commitments are an order of magnitude larger than telecom-era vendor financing. And the underlying economics still run through per-token pricing whose margins are anything but settled — how LLM API pricing works is the microeconomic end of the same question. The 1999 lesson is not that vendor financing guarantees a bust; it is that when the cycle turns, self-financed demand disappears faster than organic demand, and the vendor eats losses on both sides.
Disclosures worth reading
None of this requires insider access to track. The relevant facts live in public filings, and knowing where to look is an education in itself:
- Customer concentration. Annual reports disclose when single customers exceed 10% of revenue. Watch whether a vendor’s largest customers overlap with its investment portfolio.
- Related-party and equity-investment footnotes. Stakes, warrants, and their carrying values show up here, along with gains and losses that can swing reported earnings.
- Purchase commitments and backlog. Contracted obligations and remaining performance obligations reveal how much of the headline deal value is binding versus aspirational.
- Guarantees and off-balance-sheet structures. The CoreWeave backstop surfaced in a securities filing, not a press release. Guarantees, SPVs, and lease structures around data centers are where obligations hide from the headline numbers.
Bottom line
Circular AI deals are disclosed, legal, and rational for both sides: vendors lock in demand and upside, labs get capital and scarce supply. The revenue they generate is real cash. What they cost is clarity — every financed dollar of demand makes the industry’s growth harder to measure and its downside more correlated, which is exactly the trade Lucent and Nortel made in 1999 with worse balance sheets. The practical response is neither panic nor dismissal: read the filings, note which revenue is self-financed, and weight demand signals accordingly. And if your actual decision is which model to build on rather than what to think of the financing behind it, start with how to choose an LLM instead.
This post is educational content about market structure and corporate finance. It is not investment advice, and nothing here is a recommendation to buy or sell anything.