Nvidia reported record demand two days ago. Today AI stocks are falling.
And between those two facts sits a report about a financing programme - and a response from Nvidia worth reading word for word.
Errors or inaccuracies are possible. Spotted something that looks wrong? Write to me and I will correct it.
What the Programme Was
In July 2026 Nvidia introduced a new business model. Per the report, this is how it worked:
| What Nvidia gave | Credit support to AI cloud providers |
| What it received in return | Leaseback of capacity the customer could not sell |
| And in addition | A share of revenue from sales of Nvidia-powered capacity |
| The potential, per the company | "billions of dollars over the medium to long term" |
And it is worth pausing to understand what that actually is.
A young AI cloud provider wants to buy hundreds of millions of dollars of chips. Banks price it as a risky borrower, so financing is expensive for it.
What Nvidia offered was to close that gap itself - to provide credit support, and so cheapen the money with which the customer buys Nvidia's product.
And that is a familiar model in industry. It is called vendor financing, and it has existed for decades - a medical equipment maker, an aircraft manufacturer or a telecoms equipment maker does it all the time.
What makes this one sensitive is the second clause: Nvidia did not only extend credit, it also committed to lease back capacity the customer failed to sell. That is, it took onto itself part of its own customer's demand risk.
What Was Reported Today
The Wall Street Journal reported that Nvidia paused some of the deals in the programme - less than two months after launching it. Among the participating companies are Sharon AI and Firmus Technologies.
And per the report, three reasons:
| Control | Internal concern about how much control Nvidia exerts over its customers' businesses and their choices |
| Antitrust | Concern about exposure to regulatory scrutiny |
| Customer resistance | Cloud providers objected to the requirement that they could rent chips only to approved customers |
And the third reason is in my view the important one, because it is different in kind from the other two.
Vendor financing is one thing. Determining who your customer is permitted to sell to is another entirely.
The first is a credit term. The second is control of the distribution channel.
And that is precisely where an antitrust regulator wakes up: a supplier with a dominant share that conditions financing on the customer selling only to parties the supplier approved extends its control one layer forward in the chain.
So it is no surprise that both internal counsel and the customers objected - each for their own reason, and both leading to the same conclusion: this programme is hard to run at scale.
And Nvidia's Response, Word for Word
A company spokesperson said:
"The new business model we introduced in July that opens up compute access to the fast-growing AI ecosystem is still in place and continues to evolve due to high demand."
Read that sentence again, because it is very carefully worded.
What it does say: the model exists, and it is evolving.
What it does not say: that deals were not paused.
Those are two different claims. A report saying "some deals were paused" and a response saying "the model is in place and continues to evolve" can both be true at once. And in fact, "continues to evolve" is a formulation that contains change within it.
I am not claiming Nvidia misled anyone. I am saying that anyone who read a headline along the lines of "Nvidia denied it" read something the response itself did not say.
And that is exactly why I quote it in full rather than summarise it.
And for context: two days earlier, on 26 August, Jensen Huang publicly defended Nvidia's financial involvement in the ecosystem. He argued that "the risk is low", and that infrastructure the company invests in can be shifted between customers if a particular project runs into difficulty.
In other words: the claim presented at the quarter was that the risk is managed. Today's report adds a layer - that even if the risk is low, it was the legal and commercial structure of the programme that created friction.
And the Market Reaction
The figures below are the closing prices of 28 August 2026, at 4:00 PM New York time.
| Price | Change | |
|---|---|---|
| IREN | $35.45 | -12.53% |
| Marvell | 216.62 | -10.28% |
| Arm | 239.05 | -6.33% |
| Nvidia | $217.55 | -4.57% |
| Vertiv | 257.08 | -4.53% |
| Nebius | 209.18 | -4.26% |
| Super Micro | 37.08 | -3.59% |
| Dell | 456.24 | -3.39% |
| CoreWeave | 84.23 | -2.96% |
| AMD | 465.58 | -2.33% |
| TSMC | 417.52 | -2.29% |
| Broadcom | 368.79 | -0.74% |
| Oracle | 150.85 | -0.72% |
| Micron | 932.86 | -0.27% |
And the first thing I see in this table is that it is not uniform - and that matters.
This is not a blanket sell-off of "AI stocks". Micron fell 0.27%. Broadcom 0.74%. Oracle 0.72%. Three giants fully exposed to the same investment cycle barely moved.
And who did fall hard? Companies tied to the financing and construction side of the build-out - AI cloud providers, power equipment, and chip intellectual property.
Which says the market did not price a fall in demand today. It priced something narrower.
But the Two Largest Falls Require a Caveat
The two stocks that fell hardest - IREN and Marvell - both reported last night. And each has its own reason, unrelated to the report about the financing programme.
Marvell posted record revenue and 37% growth - alongside a non-GAAP gross margin that fell to 58.9%, with guidance for a further decline to 57.5%-58.5%. I wrote about it this morning.
IREN posted its crossover quarter - AI revenue overtook mining for the first time - but its total revenue actually fell, and adjusted EBITDA dropped 68% in the quarter.
So their falls cannot be attributed to today's news alone. Anyone who does that is measuring two different events as one.
And what is worth noting is that at IREN the two do touch each other: it is precisely the company whose deployment economics are set by external financing.
And That Returns Us to the Question I Wrote About This Morning
In the IREN piece I published today I wrote about the figure that seemed to me the most important in its report: the company raised GPU financing at two different interest rates.
| Rate | |
|---|---|
| Financing backed by the Microsoft contract | 6.0% |
| Financing against non-investment-grade customers, led by Blue Owl and PIMCO | 9.0% |
And I wrote there, explicitly, that anyone wanting to know whether the AI bubble is inflating or stabilising should track that gap.
Today's news is precisely about that gap.
Because what Nvidia's financing programme did, in practice, was to narrow it - to put Nvidia's balance sheet where the customer alone would have paid 9%. And if Nvidia steps back from that, the gap stays open, and whoever prices it is the private credit market - not the chipmaker.
That does not mean the build-out stops. It does mean it becomes more expensive for anyone who is not a hyperscaler.
What I Will Track From Here
| The rates on the next financings | 6% against 9% - every new deal reveals who the customer is |
| What happens to the July programme | "continues to evolve" is an open formulation |
| The regulatory side | The approved-customer requirement is what attracts antitrust attention |
| The hyperscalers | Oracle, Broadcom and Micron barely moved today |
| The smaller cloud providers | That is where the risk sits, and where the sharp reaction was |
הזווית שלי
דעה אישית של אילן אברמוב - לא ייעוץ ולא המלצה
What happened today is not a story about demand. Nvidia reported a record quarter two days ago, and nobody denied it.
What happened is that one question was replaced by another. For two years the market asked "how much will they buy". Today it asked "who pays for it". And those are entirely different questions with entirely different answers.
And what I try to hold onto when reading a report and a response is a simple rule: compare what was claimed to what was denied, word for word. The report said deals were paused. The response said the model exists. Those are not the same claim, and anyone reading a headline about a "denial" gets an impression the response itself did not create.
And what I think is genuinely interesting in this story is not the financing but the approved-customer clause. A supplier financing its customers is ordinary in industry. A supplier determining who its customer may sell to is something else - and that is exactly the point at which a business model becomes a regulatory question.
And what I would not do is conclude from today that something broke. The price table says the opposite: Oracle, Micron and Broadcom barely moved. What moved is precisely the edge that depends on external financing.
And that, in my view, is not a sign of crisis - it is a sign that the market has begun to distinguish between companies inside the same story. And that distinction is usually a healthy thing, even when it hurts for a day.
(It is important to stress: this is my personal opinion only, and nothing herein constitutes a recommendation to take any action.)
