Part 8 of a sourced series. Every deal here links to its source — company press releases, an SEC filing, and reputable reporting. The deals are facts. The "circular financing / bubble" reading is attributed opinion from named analysts — and I've put the rebuttals from the companies and from commentators who decline to call it fraud right next to it. Nothing here says anyone broke the law. My own opinion is marked. Corrections policy at the bottom; an evidence explorer lets you check every figure.
Money that goes in a circle
Here's the pattern that has analysts talking: Nvidia sells chips to AI companies, then invests in those same AI companies, who use the money to buy more chips. The dollars appear to loop.
Whether that loop is healthy financing or a demand illusion is the whole argument. So the boring, important thing first: the deals happened and are on the record; the interpretation that they could be inflating apparent demand is a debate, not a verdict. Let's take the deals one at a time.
Deal one: Nvidia → OpenAI
On 22 September 2025, Nvidia and OpenAI announced a letter of intent. The headline: Nvidia "intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed," tied to deploying at least 10 gigawatts of Nvidia systems (Nvidia Investor Relations; CNBC).
Read the words carefully, because they're load-bearing. "Up to." "Letter of intent." "As each gigawatt is deployed." This was conditional, milestone-linked intent — not a binding $100 billion check.
And it got smaller. On 2 December 2025, Nvidia CFO Colette Kress said the partnership "still" wasn't a definitive agreement: "We still haven't completed a definitive agreement" (Fortune). By early 2026 the headline number was scaled back — Jensen Huang said the amount would not reach $100 billion ("No, no, nothing like that") — with Nvidia reportedly finalizing a roughly $30 billion equity stake as part of OpenAI's funding round announced 27 February 2026 (Fortune). The exact round mechanics vary by source and are reported, not filed with the SEC.
Deal two: OpenAI → Oracle
OpenAI and Oracle announced a partnership to develop about 4.5 gigawatts of additional Stargate data-center capacity in the U.S. (OpenAI). The Wall Street Journal reported the deal as a five-year cloud agreement valued at roughly $300 billion (Data Center Frontier).
I want to be precise about that $300 billion. OpenAI confirms the 4.5-gigawatt partnership in its own words. The dollar figure is a WSJ report — Data Center Frontier itself flags it as reported, not fully confirmed by the parties. So treat "$300 billion" as journalism, not a disclosed line item.
Deal three: Nvidia ↔ CoreWeave
This one is the cleanest example of the loop, and it's in a regulatory filing. Nvidia agreed to an order with CoreWeave — initially valued at $6.3 billion, dated 9 September 2025, running through 13 April 2032 — under which Nvidia is obligated to purchase CoreWeave's residual unsold cloud capacity (CoreWeave 8-K, SEC; Data Center Dynamics).
Sit with the shape of that. Nvidia sells chips to CoreWeave. Nvidia also holds roughly an 11% stake in CoreWeave. And Nvidia agrees to be the backstop buyer for whatever cloud capacity CoreWeave can't sell to anyone else. Chip supplier, shareholder, and buyer of last resort — all the same company.
What the critics say
This is where the opinions start, and I'm labeling them as opinions.
Named analysts and academics have publicly called these arrangements "circular." In Global Finance, Columbia's Gregory Blotnick warned that "their interconnected AI investments mean trouble for one could rapidly cascade to others," and Minnesota's Andrew Odlyzko said "the big danger is if people wake up one morning and say… growth is slower than expected" (Global Finance). Bloomberg framed the worry conditionally: such "circular" deals "can create skewed incentives… and magnify losses if demand for AI fails to match today's lofty expectations" (Bloomberg).
A separate, sharper critique came from investor Michael Burry. In November 2025 he argued that AI hyperscalers — he named Alphabet, Amazon, Meta, Microsoft, and Oracle — understate depreciation by stretching the assumed useful life of GPUs to ~5–6 years when, in his view, it's closer to 2–3. He estimated roughly $176 billion of understated depreciation across 2026–2028 that he said inflates reported profits (CNBC; Motley Fool). That is Burry's thesis — an accusation he is making, not a finding anyone has proven. And it's the same lever I wrote about in Part 1: how long you assume a machine lasts quietly moves billions in profit.
What the other side says
Now the counter-view, because it's at least as important. Reputable commentators draw a hard line between this and illegal round-tripping. Economist Noah Smith put it bluntly: "It's perfectly normal and healthy for… General Motors to lend its customers the money they use to buy GM cars… This is called vendor finance. It's perfectly legal and commonplace" (Noahpinion). Real Investment Advice, even while raising the round-tripping question in its headline, stated flatly: "We are not accusing Nvidia, OpenAI, and other large tech companies [of fraud]" (Real Investment Advice).
The companies pushed back directly, too. Nvidia circulated a fact-check memo to shareholders; Jensen Huang called reports of friction with OpenAI "nonsense" and "I really love working with Sam" (Fortune). Sam Altman said OpenAI's revenue is "growing steeply" — "we're doing well more revenue than that" — and Microsoft's Satya Nadella said OpenAI had "beaten every business plan" it presented to Microsoft (Yahoo Finance).
So: real demand, or financed demand? The honest answer is that smart, sourced people disagree — and none of the ones cited here are alleging a crime.
My read
Opinion — Michael. I'm not telling you this is fraud, because I can't, and the careful people on both sides aren't either. What I see is concentration risk dressed as momentum. When the chip seller is also the investor and also the backstop buyer, demand and supply stop being independent signals — and a number that should be a thermometer starts acting like a mirror. That doesn't make it illegal; vendor financing is ordinary. It makes it fragile, in exactly the way Odlyzko describes: fine until the morning someone decides growth came in slow. And Burry's depreciation point rhymes with Part 1 — the same assumption that flatters this quarter sends the bill to a later one. Bad systems, not bad people — but a system this interlocked has fewer places to hide a wrong guess.
You don't have to take my read. The press releases, the 8-K, and the analysts on both sides are linked above. Judge the loop yourself.
Sourcing & corrections
The Nvidia–OpenAI letter of intent is quoted from Nvidia's own IR release; the downscaling from Fortune; the Oracle 4.5-GW partnership from OpenAI, with the ~$300B figure framed as WSJ-reported via Data Center Frontier; the CoreWeave backstop from CoreWeave's SEC 8-K and Data Center Dynamics. The "circular"/bubble characterization is attributed opinion (Blotnick and Odlyzko via Global Finance; Bloomberg; Burry via CNBC and Motley Fool); the counter-view and company rebuttals are from Noahpinion, Real Investment Advice, Fortune, and Yahoo Finance — each linked inline and matched in the explorer. Spot an error or something unfair? Email mpolzin@zimzap.com or message me on LinkedIn — I'll review and correct.
Next — Part 9: "The Narrative Machine."
🔎 Check it yourself: explore every claim →
Top comments (0)