AI Vendor Financing: Legal, But Not Risk-Free

There's a chart that's been making the rounds. Arrows loop between Nvidia, Microsoft, and OpenAI. Money flows in, money flows out, and if you squint at it long enough, it starts to look like a perpetual motion machine — or a con. The Enron comparisons write themselves, and plenty of people have obliged.

Ani Björkström, who covers AI for a finance audience, thinks most of those people are wrong. But she also thinks the people waving the chart away deserve equal scrutiny. Her recent video on the subject is worth engaging with seriously, because she makes a careful distinction that most of the commentary — on both sides — has glossed over.

The word that's doing all the work

Start with the label. "Circular financing" isn't a term of art. It's a description of a shape — arrows forming a loop on a diagram — that journalists reached for because it was vivid. The Financial Times' Lex column, Björkström notes, uses a different phrase: "back-to-back." Less alarming, more accurate.

The actual accounting term for what's happening is vendor financing, and once you know what that means, the architecture of the AI investment web becomes considerably less mysterious.

Vendor financing is when a seller helps fund its own customer's purchase. A car dealer who offers you a payment plan is doing vendor financing. A pharmaceutical company that takes an equity stake in a small research lab — in exchange for a promise that the lab will use the pharma firm to manufacture whatever drug it develops — is doing a more sophisticated version of the same thing. The seller gets a customer; the buyer gets capital it couldn't otherwise access; both parties now share an interest in the outcome. As Björkström puts it: "Their fates are now tied together. Hold on to that — because it is the whole story. It is the upside and it is the risk in one move."

That's the structure. Microsoft invests in OpenAI; OpenAI spends that money on Microsoft's cloud infrastructure; Microsoft books revenue and holds equity in its own biggest customer. Amazon and Google have replicated the same arrangement with Anthropic. Nvidia sits at the center of the web, its chips the physical substrate on which all of this runs.

What it isn't, and what it is

The version of this that would constitute a crime has its own name: round-tripping. That's when two companies shuffle money between each other purely on paper — no goods move, no services are rendered, revenue figures inflate on both sides. Enron did this. Executives went to prison.

What's happening in AI is structurally different. Real chips get shipped. Real servers process real queries. When you use ChatGPT, compute cycles on Microsoft's infrastructure are consumed. Revenue is earned, not fabricated. "Vendor financing is not round-tripping," Björkström says plainly. "This isn't a dodgy accounting trick that only exists on paper."

That distinction matters, and it's been underplayed. The Enron comparison is wrong as a matter of accounting, and repeating it drowns out the legitimate concerns.

The numbers, carefully stated

Here's where the picture gets genuinely interesting rather than merely alarming. Microsoft's AI revenue run rate reached $37 billion as of April 2025, up 123% year on year. Working backward from that figure, the previous year's run rate was roughly $16.5 billion, or about $4 billion per quarter.

According to Ed Zitron's reporting in The AI Demand Bubble, OpenAI spent approximately $2.9 billion in a single quarter running its models on Microsoft's servers. Set those two figures side by side and you get an estimate — explicitly an estimate, drawn from reconstructed data rather than audited accounts — that OpenAI may account for something in the range of 70% of Microsoft's AI revenue. Björkström is careful to flag that OpenAI is private, these are reconstructions from leaks, and they should be treated as the right order of magnitude rather than gospel. That caveat isn't throat-clearing; it's the most important sentence in the analysis.

Microsoft has invested more than $13 billion in OpenAI. So the company is, in effect, financing its own largest customer to buy its own services. Each transaction is real. The circularity is also real.

This off-balance-sheet AI debt concentration is larger than most outside observers can measure — which is precisely the problem Björkström is circling.

Paid twice, or burned twice

The upside case for this structure is elegant. Every time a user runs a query through ChatGPT, Microsoft collects a hosting fee because the request runs on its servers. It also holds equity in OpenAI, so it participates in OpenAI's profits if and when those materialize. Two revenue streams from a single user interaction.

The downside is the mirror image. If AI confidence collapses, Microsoft loses twice simultaneously: the equity stake gets written off, and the cloud customer paying the bills disappears. "Every vendor in this web loses twice in the same moment," as Björkström puts it. Neither OpenAI nor Anthropic is currently profitable. The losses, by all accounts, are substantial.

Where the dot-com comparison actually holds

Björkström's most useful analytical move is the one where she disaggregates the bubble comparison. She argues the AI investment web is genuinely different from the dot-com era in one important respect: then, vendor financing was predominantly debt-funded, which is how financial contagion spreads — from failing tech companies into banks, pension funds, and insurers. Today, the hyperscalers are largely financing this from their own balance sheets.

If Microsoft and Google want to bet their own capital on AI and lose, that's a private matter with private consequences.

Except — and here's where I find myself less sanguine than Björkström — the data center buildout underpinning all of this is debt-financed. That's a separate risk channel, and she acknowledges it exists. The equity-funded investment web and the debt-funded infrastructure layer are two distinct stories running in parallel, and conflating them is as much an error as treating vendor financing as fraud.

The individual deal structures — equity stake, cloud commitment, compute access — are defensible when you examine them one at a time. I find that less reassuring than it might sound. The part of this that actually unsettles me isn't the mechanics of any single arrangement; it's the opacity layered on top of them. OpenAI remains private. The figures we're working with are reconstructions from leaked documents. We know the risk is concentrating. We cannot measure how much. Björkström's phrase for the shape of this — "concentration of risk" — is the right one. It amplifies the rewards if AI delivers, and it amplifies the damage if it doesn't. What it doesn't do is make the damage legible to anyone on the outside until it's already arrived.

Vendor financing is a bet. These companies have placed the same bet simultaneously, in multiple directions, at scale that is difficult to fully audit because one of the central players publishes no public accounts.

That's not Enron. It's also not nothing.

---

Bob Reynolds is Senior Technology Correspondent at BuzzRAG.