On Monday, 10 August 2026, NVIDIA announced the creation of compute infrastructure financing platforms together with six of the world's largest asset managers. We are writing about it today.
Errors or inaccuracies are possible. Spotted something that looks wrong? Write to me and I will correct it.
What Was Announced, Precisely
NVIDIA signed memorandums of understanding with six firms: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR.
The stated purpose: to establish independent compute infrastructure financing platforms creating dedicated pools of capital at significant scale and at attractive rates for NVIDIA customers - frontier AI labs, enterprises and AI clouds.
The target size: more than $500 billion of third-party capital, over time.
And three words in the release determine how to read everything else.
"Independent". The platforms are not NVIDIA's. They are established with the financial partners and run by them, meaning the capital does not sit on NVIDIA's balance sheet.
"Third-party capital". The money is neither NVIDIA's nor the customer's - it comes from institutional credit, insurance funds and private capital.
"Subject to execution of the final agreements". The release says so explicitly, and in the CNBC interview these were described as memorandums of understanding. This is an announcement of intent, not a signed transaction.
The Problem This Is Meant to Solve
Here you have to understand why NVIDIA is in the financing business at all.
The constraint on NVIDIA's growth is not demand. It is its customers' balance sheets.
Anyone buying GPUs in the billions has to put the money up front, and book an asset regarded as rapidly depreciating equipment. A mid-sized lab or cloud provider simply cannot fund that from its own balance sheet, and even the large hyperscalers are beginning to feel it - the rating agency Moody's has already warned that unprecedented capital expenditure is eroding free cash flow and pushing the technology giants into heavier debt.
So if the customer cannot fund it, NVIDIA's revenue stops - even when the demand exists.
What this announcement tries to do is move the financing off the customer's balance sheet and into the capital markets.
The Idea Itself: a GPU as Infrastructure
And this is the genuinely interesting part, and it is an accounting argument before it is a technological one.
Until now, a GPU has been treated as equipment. Equipment wears out and becomes obsolete when a new generation arrives, so a lender struggles to extend long-dated credit against it.
NVIDIA's claim is that this classification is wrong. Jensen Huang told CNBC:
"This is really the first time that technology chips have become an investable asset class"
And he lists four properties that he argues make GPUs financeable: they generate revenue, they are long-lived, they are fungible, and they are flexible. In the official release he adds that NVIDIA compute is broadly adopted, transferable across customers and operators, and continuously improved through CUDA software.
And the sentence that carries the whole thesis:
"In AI, compute is revenue"
His chosen comparison is not incidental: he says the computer has become part of the infrastructure, like electricity and like the internet, and so it has to be thought of as infrastructure.
Why this might actually work, and why the asset managers came in.
An asset you can borrow against needs three properties: it must produce ongoing income, it must be transferable to someone else if the borrower fails, and it must have a secondary market.
And on the claim, NVIDIA GPUs meet all three: they are rented and produce ongoing payment; the CUDA ecosystem ensures other customers could use them; and demand exceeds supply.
Jon Gray, president of Blackstone, put a number on it: AI usage across Blackstone's portfolio companies has grown sevenfold this year. And demand, he says, is outstripping supply.
And Jim Zelter of Apollo frames it in asset-class terms: modern compute, he says, has emerged as a scarce, mission-critical asset with compelling investment characteristics.
And David Solomon of Goldman gave CNBC a detail worth holding: it was Huang who approached the asset managers with the idea. This was not a Wall Street initiative.
The Comparison Fink Made Himself
And here comes the most important statement of the day, and it did not come from NVIDIA.
Larry Fink, CEO of BlackRock, told CNBC he believes this is the beginning of the "next future for financial engineering" - and compared it to the creation of mortgage-backed securities in the 1970s.
He meant it as a compliment. He added that some money has already been raised, that BlackRock will be raising considerably more, and that the money needs to be raised as fast as possible because he considers it imperative that the United States leads in AI.
And that comparison is accurate - which is exactly the problem.
Mortgage securitisation really did do what Fink describes. It took an illiquid asset, divided it into tradeable units, and channelled capital into it at a scale that had not been possible before. It financed American home ownership for three decades.
And then, in 2008, it did the other thing as well.
What broke was not the securitisation structure - it was the assumption about the collateral. As long as it was assumed that house prices do not fall nationwide, the structure held. When that assumption broke, every layer built on top of it broke with it.
So the only question that matters here is: what is the assumption about the collateral. And the answer is that NVIDIA GPUs hold their value over time.
And here it is worth pausing, because that assumption sounds absurd - and it is not actually the assumption being made.
So How Do You Finance an Asset You Know Will Depreciate
That is the first question to ask, and the answer is less intuitive than it seems.
Depreciation and financeability are not the same question. A truck loses value every year, and so does an airliner, a drilling rig and a printing press. All of them are financed with debt, at enormous scale, and have been for decades.
What a lender actually checks is not whether the asset will wear out - it is two other things:
- Whether the income the asset produces over the life of the loan repays the loan.
- Whether the pace of decline is predictable enough to price in advance.
A GPU that pays back its cost within two to three years, with debt amortised over that same period, is a perfectly reasonable financed asset - even if it is worth very little in year six. The danger is not the depreciation. It is a gap between the pace of decline and the tenor of the debt.
Which gives the precise statement of the assumption, and it differs from how it first sounds.
The assumption is not that GPUs do not depreciate. Nobody claims that. The assumption is that the decline is predictable, gradual, and slower than the amortisation schedule.
And that is a far weaker assumption than 2008's - where the assumption was that house prices do not fall nationwide, meaning the collateral does not erode at all. Here erosion is conceded up front, and only its pace is priced.
What is genuinely dangerous is not a gradual decline. It is a step.
And for GPUs that step is specific and describable: a previous-generation GPU stays economic as long as the revenue it produces exceeds the electricity it consumes. The moment a new generation is efficient enough to make the old one uneconomic to run, its value does not decline gradually - it collapses.
Which is exactly why the metric worth following is the second-hand price. It prices the distance to that step, and nothing else.
The Item That Is Not in the Release
And there is a detail worth flagging, precisely because it is not in the official release.
According to trade press reporting, NVIDIA may provide "residual-value support" of up to 25% of an opportunity, decided case by case, with the emphasis that the mechanism does not replace the financial institutions' own due diligence.
NVIDIA's official release says nothing about any capital commitment or guarantee on its part, so we present this as reporting rather than confirmed fact.
But if it is accurate, the implication is significant: a residual-value guarantee means NVIDIA carries part of the downside risk on GPU values. That turns the structure from "NVIDIA sells and others finance" into "NVIDIA sells, and if the value falls it absorbs part of it".
So Why Did the Stock Fall
The stock fell 2.86% on the day of the announcement, from $223.96 to $217.55, on heavier-than-usual volume.
And that reaction is the most instructive thing about how the market read it.
The positive reading was available: a company that removes the financing constraint from its customers opens itself a far larger market.
And the market chose the second reading: if your customers need $500 billion raised from outside in order to keep buying from you, then the demand you have been reporting was not fully funded.
הזווית שלי
דעה אישית של אילן אברמוב - לא ייעוץ ולא המלצה
I think this is a very important announcement, and also the easiest one to read wrongly.
Start with what it actually is: not a product announcement and not a demand announcement. It is an announcement about a change in how an entire industry is financed. And historically those deserve more attention than new chip launches - because what sets the pace of building is not what can be built, but who is willing to fund it.
And the insight behind the move is, to my eye, correct. NVIDIA identified that its constraint is not technological but balance-sheet, and that it sits with the customer rather than with itself. That is a sharp diagnosis, and the solution matches it precisely.
What convinces me most in the argument is actually the word "fungible". The difference between equipment and infrastructure is exactly the ability to hand it to someone else when the first user fails. An NVIDIA GPU, thanks to CUDA, does move between customers relatively easily - which makes it far better collateral than a purpose-built machine.
And now three things I am not willing to smooth over.
The first is Fink's comparison, which he made himself. When the CEO of BlackRock compares a new financing structure to mortgage securitisation and means it as praise, he is describing both the capability and the risk accurately. Securitisation did not fail in 2008 because of the structure; it failed because the assumption about the collateral turned out to be wrong. Here the assumption is that GPUs hold their value - and the entire history of the chip industry argues against it. A new generation arrives every 18 to 24 months, and each one lowers the value of the last.
The second is the reflexivity. NVIDIA sells GPUs, invests in the customers who buy them, and per reporting also partially guarantees their residual value. Each leg is entirely reasonable on its own. Together they create a loop in which it is hard to distinguish real demand from demand financed by the seller. I am not claiming that is what is happening - I am saying the structure makes it harder for anyone trying to check.
And the third is the simplest: this has not been signed. The release says explicitly "subject to execution of the final agreements", and these are memorandums of understanding. $500 billion announced is not $500 billion raised. Fink said some has already been raised, but named no figure.
And what I find most interesting is how this connects to the whole week. We wrote about Riot and Bitdeer converting power into data centres, and about Meta spending $130 to $145 billion a year on infrastructure. In every one of those cases the narrow link was not the idea - it was who funds the concrete, the power and the steel.
NVIDIA's announcement is the first large-scale attempt to solve exactly that. If it works, it accelerates everything we wrote about this week. And if the assumption about the collateral turns out to be wrong, it accelerates the opposite direction too.
And what I will watch is a single measure: the second-hand price of previous-generation GPUs. As long as it holds, this structure stands. The moment it breaks, every financing layer built on top of it gets repriced.
