The Ouroboros: How $800 Billion Is Circling the Same Six Companies

I wasn’t going to write about circular financing. Everyone has a chart now.

For six weeks my feed has looked like a subway map drawn by a conspiracy theorist — arrows from Nvidia to OpenAI to Oracle and back again, each one labeled with a number ending in “billion.” The takes split cleanly into two camps. One says it’s fraud. The other says it’s genius. Both are lazy.

So I did what I always do when a narrative gets too loud: I went back to the last time it happened. Not metaphorically. The exact same structure, with the exact same justifications, ran twenty-six years ago. It has a name, a body count, and a set of financial statements you can still pull up. And the one number that killed it last time is sitting, right now, in a footnote almost nobody reads.

Let me show you the loop first, then the ghost, then the number.

Here is the mechanism stripped of the subway map. Company A — a chip maker or cloud vendor — invests in Company B, an AI lab. Company B spends that money buying Company A’s products. The money circles a small cohort, inflating apparent demand. Nvidia buys stock in OpenAI; OpenAI commits enormous sums to cloud providers like Oracle; those providers buy Nvidia GPUs to fulfill the contracts. Each leg books revenue or backlog from the same underlying spend.

By 2026, analysts pegged the web of interlinked commitments north of $800 billion. And notice where the scale lives: in the commitments, not the cash. That distinction is the whole story, and I’ll come back to it.

The bull framing isn’t stupid, to be fair. To supporters, the “circularity” critique misses a basic point: building AI is extraordinarily expensive, and the most advanced chips are still hard to get. In that kind of market, companies don’t just place orders — they lock in supply by pairing long-term buying commitments with financing. One large asset manager called it a “virtuous circle” that lines up suppliers, builders and customers to meet exploding demand for compute.

Virtuous circle. Hold that phrase. You’ve heard it before.

Here’s the detail that made me stop scrolling and start writing.

The single most-cited arrow on every one of those subway maps — Nvidia’s “$100 billion investment in OpenAI,” announced with great fanfare in September 2025 — was never a signed deal. It was a letter of intent. Nvidia said it “intends” to invest “up to” $100 billion, progressively, as each gigawatt of systems got deployed. No definitive agreement. No money moved.

Then it quietly came apart. By December 2025, Nvidia’s own CFO confirmed at an investor conference that no binding agreement existed, and the commitment was excluded entirely from Nvidia’s data-center bookings guidance. In January, reporting described the talks as “on ice,” with internal doubts at Nvidia about OpenAI’s business model. By February, still no contract, no cash. Jensen Huang himself called the full $100 billion “probably not in the cards.”

What actually happened was smaller and more revealing. In late February 2026, OpenAI announced a funding round — led by Amazon, Nvidia, and SoftBank — that closed at the end of March at a post-money valuation reported in the mid-$800-billions. Nvidia’s piece was roughly $30 billion of straight equity, not tied to gigawatt milestones. Alongside it, OpenAI made binding hardware commitments for several gigawatts of Nvidia’s next-generation systems.

Sit with that for a second. The number that anchored a thousand “AI is a circular bubble” charts was, for five months, a press release. The real transaction was a third of the size and structured completely differently. If the headline leg of the loop was mostly narrative, how much of the other $800 billion is commitment theater — and how much is real, contracted, cash-backed demand?

That’s the question worth answering. And the way to answer it is to find the leg of the loop you can actually audit, line by line, in public filings. But first, the ghost — because we’ve run this experiment before, and we know how it ends.

In the late 1990s, telecom equipment giants — Cisco, Nortel, Lucent — borrowed heavily to offer their customers financing deals that essentially manufactured sustained demand for telecom equipment. Because that gear was in short supply, customers — many of them startups building out fiber-optic networks — inflated their orders, contributing to a glut that left them reeling when, in 2001, it became clear the industry had wildly overestimated demand. The equipment makers were left holding bad debt while the startups went bust.

The dollar figures are almost comically on-the-nose. Lucent committed around $8.1 billion in vendor financing. Nortel extended roughly $3.1 billion. Cisco promised about $2.4 billion in customer loans. The pitch was airtight: lend money to cash-strapped carriers so they can buy your equipment. Everyone wins — until the merry-go-round stops.

Then it stopped. Dozens of carriers went bankrupt between 2000 and 2003, vendors wrote off billions in unpaid loans, and Cisco’s share price fell nearly 90% — never revisiting its 2000 peak, despite the company’s earnings today running many times higher than they were then.

Read that last clause twice. Cisco was right about the internet. Traffic did explode. The company grew into and far past its old earnings — and the stock still never came back, because the price had detached from the business by an order of magnitude. Being correct about the technology did not save the people who bought the stock at the top.

And here’s the part the bulls skip: it was never a demand problem. Demand was real and growing, just as it is with AI today. The problem was that the industry built far more infrastructure than profitable demand could fill. Carriers laid tens of millions of miles of fiber, much of which sat unused — “dark fiber.” Supply outran profitable demand, prices for bandwidth collapsed, and capacity became a commodity. The internet didn’t die. The financing structure did.

The bears have a name for the 2026 version of dark fiber. They call it dark GPUs.

I don’t want to sell you a crash. The honest position is that the differences are real, and the two biggest banks on the street have said so.

One argument: although the deals are vendor-financed, the hyperscalers still have enormous operating cash flows to support the obligations if needed — a buffer against structural collapse. Another: monetization looks different this time. Where early-internet companies built first and figured out revenue later, AI is generating revenue as it scales — measurable cloud consumption and productivity gains today, not someday. And the balance-sheet strength of a Microsoft, an Amazon, or an Alphabet is nothing like the over-leveraged carriers of 2000.

That’s true. Those three can eat a bad bet with operating cash flow and shrug.

But “the hyperscalers are fine” is a sleight of hand — because the hyperscalers aren’t the fragile part. The risk got quietly transferred to a set of companies most retail investors have never examined. And that transfer is the entire game.

Here’s the mechanism. Hyperscalers face acute compute shortages, so they rent capacity from specialized GPU cloud providers — “neoclouds” — that act as shadow clouds for AI workloads. This has produced staggering top-line growth across the neoclouds. But much of that “demand” is really a transfer of balance-sheet risk: the hyperscaler expands capacity without tying up its own capital, while the debt-financed neocloud absorbs the depreciation, the interest costs, and the refinancing pressure. Great for Big Tech’s capex optics. Capital-intensive and fragile for the neocloud.

There’s one neocloud whose public filings let you watch the loop tighten in real time. Nvidia owns a piece of it, sells it chips, and has signed a contract to buy back its unsold capacity through 2032. It has roughly $99 billion of backlog, about $25 billion of debt, and it burned nearly $5 billion of cash in a single quarter.

That’s where the audit begins — and where this free edition ends. I built a model around it. Here’s what’s inside.