Nvidia announced this week that Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR were willing to commit up to $500 billion to build AI data centers. That eye-popping figure grabbed all the headlines. But the bigger story hiding underneath is what Nvidia promised those financial giants in exchange: that its chips, used as collateral in these deals, will not lose value. Nvidia is putting its own money behind that guarantee.
I wrote about the first chapter of this story back in July — the $50 billion lease and the talks to back $250 billion of OpenAI data center financing. This is the sequel, and it is much bigger. Let me walk you through what changed, why the plan is genuinely clever, and the part that keeps me up at night.
The Deal, in Plain Numbers
On August 10, Nvidia signed memorandums of understanding with six of the biggest names in global finance to set up financing platforms for its customers. The stated goal: mobilize more than $500 billion in third-party capital so hyperscalers, frontier AI labs, and enterprises can build data centers and buy Nvidia hardware without tapping their own balance sheets.
This is not Nvidia spending $500 billion — as TechCrunch detailed, it is Nvidia convincing the market to spend it. It is Nvidia convincing the market to spend it. The company becomes the bridge between Wall Street money and AI infrastructure — and it is taking on a real obligation to make that bridge hold.
The Part Nobody Talked About: Aging GPUs
Here is the twist that makes this deal different from every financing arrangement before it. Nvidia has agreed that if the GPUs used as collateral in these loans do not retain their expected value, it will cover up to 25% of the difference. If a data center operator defaults and the lender has to sell the chips, and the chips sell for less than the books said they should, Nvidia eats a chunk of that loss.
Think about what that means. Nvidia is effectively creating a secondary market for used GPUs, then underwriting it with its own balance sheet. It wants aging hardware to keep trading at healthy prices, because that is what makes banks comfortable lending against it in the first place.
Jensen Huang made the pitch directly on CNBC: “This is really the first time that technology chips have become an investable asset class. These are revenue-generating assets now. They’re productive, they’re long-lived, they’re fungible, they’re flexible.” His framing treats AI servers like railroads or airlines — infrastructure that keeps earning for decades — rather than PCs that depreciate the moment you open the box.
“When needs change, the factory can be used by another customer, another cloud or another operator. This broad ecosystem gives NVIDIA compute a deep market of potential users and offtakers, helping protect residual value,” he promised.
The Word Everyone Is Using: Circular
Financiers call what Nvidia just signed up for “wrong way” risk. That is the uncomfortable situation where your obligations grow exactly when your business weakens. If AI demand softens, Nvidia’s revenue drops at the same moment its guarantee obligations balloon. The two curves move in the wrong directions for the same company.
The comparison that hangs over all of this is Lucent Technologies. Lucent lent its customers money to buy its own telecom equipment during the dotcom boom, and when the bubble burst, the loans and the equipment both collapsed with it. Nvidia is aware of the shadow — Huang has explicitly addressed it, and the bond markets got spooked enough after the announcement that he went on business TV to explain how the risk was limited.
“Is this circular financing?” Huang wrote on X. “This initiative is designed to address that concern. We are bringing independent, long-term institutional capital into the AI infrastructure market.”
That is a fair answer, and not just spin. Unlike Lucent, Nvidia is not lending its own money to customers. It is getting other people — very sophisticated people — to shoulder most of the capital and most of the risk, in exchange for a partial value guarantee on the collateral. Bloomberg has calculated that Nvidia has been involved in roughly $750 billion worth of deals this summer that revive the circular-financing debate. This new structure is Nvidia’s answer to that criticism, and it is a genuinely clever one.
But here is the uncomfortable context: the traditional funding taps are running dry. Oracle is deep in debt. Google issued a record $85 billion equity raise. Meta is burning through cash. Satya Nadella recommended a book called “1873” — about the railroad-era financial engineering that crashed the US economy — during Microsoft’s last earnings call. When CEOs start recommending books about historical financial crashes, the room is telling you something.
What Wall Street Says It Sees
The executives on the other side of the table are not naive. Goldman Sachs CEO David Solomon called it “a pivotal moment of a historic AI investment cycle.” Blackstone’s Jon Gray compared AI compute to housing — a “financeable asset class” in the way mortgage lenders look at homes — and said demand at Blackstone portfolio companies has surged sevenfold this year. BlackRock’s Larry Fink went further, calling the project the start of the “next future for financial engineering,” likening it to the creation of mortgage-backed securities in the 1970s.
Read that last comparison slowly. Mortgage-backed securities — the financial instrument that, when engineered irresponsibly, helped crash the global economy in 2008. Fink means it as a compliment: this is how you turn an illiquid asset into something institutions can invest in at scale. And he is right about the mechanics. The question is whether the asset underneath is as solid as everyone is betting it is.
Felix’s Take: This Is Not Academic for Me
As someone who manages IT budgets in a government institution in the Philippines, I have a very personal relationship with hardware depreciation. We do not buy GPUs in the hundreds of millions, obviously — but every equipment procurement I sign off on has a lifespan, a book value, and a moment when the accountants tell me the asset is “fully depreciated” even though it still works fine. The gap between accounting value and real value is something I think about constantly.
What Nvidia is doing is essentially arguing that for AI hardware, the accounting is wrong — that GPUs keep earning long after traditional depreciation schedules say they should. That argument benefits everyone who owns Nvidia hardware today, because it props up resale values. And it benefits anyone who might want to buy used GPUs tomorrow, because it gives the secondary market liquidity and legitimacy.
There is a real upside here for smaller players. A healthy market in aging GPUs means that the “two AI economies” I wrote about earlier — the frontier economy of billion-dollar clusters and the production economy of budget-conscious teams — get a bridge between them. A Filipino startup that cannot afford new B200s might one day lease or buy mid-generation hardware at a price that actually makes sense, backed by the same financing machinery that funds hyperscalers.
But I have been around long enough to know that when a company starts guaranteeing the value of its own products, it is either supremely confident or desperately trying to keep the music playing. Nvidia’s CUDA software layer genuinely does keep old chips useful — that is a real moat, and I wrote about how that moat is under pressure from AMD and open approaches. The question is whether software improvements can keep pace with the sheer scale of hardware being financed.
The Other Half of the Story: Pruning
Here is the part that makes this week feel like a pivot point. On the same day the Nvidia story broke, Microsoft announced it was killing off a bunch of its AI features — AI-generated podcasts, Group Chats, Deep Research in the consumer Copilot, the Mico character — and merging its consumer and business Copilot apps. The message from an internal memo was blunt: Copilot had to earn “the right to exist.”
Meanwhile, Databricks raised $5 billion at a $190 billion valuation after its CEO said the quiet part out loud: “AI is expensive.” The company wanted to raise $1 billion, saw $15 billion of interest, and took $5 billion anyway. Even a profitable, cash-flow-positive company growing 80% year over year feels the need to stockpile capital because compute costs are that brutal.
Put the two stories side by side and the picture gets sharp. On one side, the industry is building the largest financing machine ever created to fund AI infrastructure. On the other, it is aggressively cutting features that do not generate enough value to justify their token costs. That is not a contradiction — that is a market maturing. Build-out money is still flowing, but it is flowing with terms, and everything that cannot prove its worth is being pruned.
So Where Does That Leave Us?
If Nvidia’s plan works, compute becomes a long-lived, bankable asset, and the whole AI economy gets more stable. More capital, better secondary markets, cheaper access to older hardware. If it does not work — if demand plateaus, if a new architecture makes today’s GPUs obsolete faster than expected, or if the financing itself becomes the bubble — then the losses will be bigger and more interconnected than anything we have seen in tech since 2008.
The honest answer is that nobody knows which way it breaks. But the direction of travel is clear: the AI buildout has entered its finance phase. The era of “build it and the money will find you” is ending, replaced by the era of “show the bankers how it pays back — the question I keep circling around in the $3 trillion AI question.” For those of us watching from the cheap seats — running local models, budgeting for inference, waiting for the hardware we can actually afford — that shift might be the best thing that has happened to the market. Or the thing that breaks it.
When you are playing a game this big, you learn to watch the other player’s balance sheet as closely as your own.