Nvidia Is Now the AI Buildout's Lender of Last Resort

How much is a used AI chip worth? The people who lend money against one have settled on an answer, and it is not Nvidia's. Banks write a GPU down over three to four years, and to lend against one at all they demand higher rates and stronger guarantees than they would for ordinary equipment.

AI compute is a productive, durable and fungible asset that can support long-term financing. — Nvidia

Nvidia's own view is that its chips keep earning for the better part of a decade ^[1]. That is not a depreciation-table detail. It is the whole question underneath the past year, because the one party willing to price an AI chip optimistically has turned out to be the company that makes it. The gap is not academic. One of the people standing on the wrong side of it described it plainly this year — and he is himself, as of this week, one of Nvidia's financed customers ^[2].

debt markets have almost entirely shut off access to money for any company that doesn't have substantial revenue independent of AI data centers. — Elon Musk

A central bank is a lender of last resort: it steps in when no one else will. Over twelve months Nvidia has taken on that job for the middle layer of the AI buildout — the labs, the rented-computing firms known as neoclouds, and the chip-buyers that banks and bond markets now refuse. The disclosures of this week alone — an $18 billion equity stake in SpaceX, credit guarantees for a West Virginia campus, an Indonesian loan syndicated on Nvidia's name — are only the latest entries in a ledger that has been filling since last fall ^[2][3][4]. It started, in the cleanest accounting, last October. OpenAI agreed to pay Nvidia cash for its chips, and Nvidia agreed to take a slice of that cash and buy non-controlling shares in OpenAI — a loop in which the customer's payment comes back as the customer's funding ^[5]. The groundwork was already on the books: a $700 million stake in Nebius over the summer, and a portfolio so concentrated in CoreWeave that by September nearly nine dollars in ten sat in that single company, alongside a pledge to buy whatever compute capacity CoreWeave could not rent ^[6][7]. Vendor financing — a seller lending its customer the money to buy the seller's own product — got its historical rhyme from the critics right away.

I have to say the words ‘vendor financing’ do not carry nice reflections to somebody of my age. — Steve Austin

By November, after Michael Burry put a billion dollars against the stock, Nvidia answered the circular-financing charge with a seven-page memo ^[1]. The defense rested on one phrase.

catastrophically overbuilt supply and nowhere near enough demand — Michael Burry

Then the arithmetic ran past the phrase. By May, $90 billion was committed to backing cloud providers, suppliers, and developers ^[8]. By August the instrument had changed from equity to guarantees, Nvidia putting its own balance sheet behind the labs and neoclouds — by its own account, to hold market share for its next chip platform, Vera Rubin, against custom silicon ^[9]. Those guarantees now carry a stated cap of $108.5 billion ^[10]. This month came the $18 billion SpaceX stake, sitting beside the $40 billion SpaceX borrowed to buy Nvidia GPUs ^[2]; the talks to backstop construction and electricity at a West Virginia campus built for Anthropic ^[3]; and an Indonesian firm, described in the paperwork as "backed by Nvidia," syndicating a $3.1 billion loan for GPUs across five banks ^[4]. The scope is bounded by who still needs the help. Meta just raised up to $97.5 billion for its buildout and was oversubscribed five times; Nvidia put $3 billion into that offering, AMD a billion — suppliers buying equity in their own biggest customer ^[11]. Meta can fund itself. The lender-of-last-resort role concentrates in the middle layer, the firms that cannot borrow, not in the giants that still can. The natural end point of the whole project is to stop lending entirely — to make the chip itself acceptable collateral at an ordinary bank. That is the $500 billion plan Nvidia has been shopping, modeled on aircraft leasing, in which developers would pledge GPUs against long-term loans ^[12]. It is an attempt to settle the depreciation argument by declaration, and it is exactly where the banks' three-to-four-year answer comes back in. The plan is still a proposal under revision, and the $250 billion lease-support figure attached to it is rumored rather than confirmed. The refusals run alongside. OpenAI's finance chief floated asking Washington for a backstop of last resort; the Treasury's answer was that developers are responsible for their own risks ^[10]. Alex Karpovsky put the request in its bluntest form.

to go to the government and say, nationalize us, please — Alex Karpovsky

The government declined. And Oracle — which had been playing banker for the buildout through its cloud backlog — this week capped its AI financing at $20 billion through 2030 and promised to lean on vendor financing and customer prepayments instead ^[13]. The stock jumped 12% on the promise to stop being the banker. The exit from the role points straight back at the one party left holding it. None of this is hidden, and not everyone reads it as a bubble. The analysts who map the whole pattern — suppliers from Nvidia to Broadcom lending to startups to lock in orders — still rate Nvidia a Strong Buy, on the ground that today's leaders have real revenue and real earnings ^[14]. And JPMorgan's Stephanie Aliaga does different arithmetic on the debt ceiling.

We think that the market is very capable of absorbing this new issuance. And if anything, it might just enable this AI boom to be sustainable. — Stephanie Aliaga

The pattern may well be survivable. One limit even Nvidia's guarantees may not cover is insurance: clustered data centers concentrate risk, and Aon's chief executive warns that if insurance gets scarce, the financing structure changes entirely.

If it's more difficult to get insurance, that means the financing structure is going to change completely. You can't access the debt markets. — Greg Case

But the hedges are the tell, and they sit in plain sight. CoreWeave trades about 60% below its highs despite Nvidia's $2 billion stake and its pledge to backstop up to $6.3 billion of unsold capacity ^[15]. The short seller Vincent Daniel stated the open question.

To own this equity, an investor must underwrite that CoreWeave can raise tens of billions of dollars in incremental capital on accretive terms—not a bet we are willing to take. — Vincent Daniel

The analysts tracking the financing partnerships keep a sell rating on the stock, citing circular-financing risk ^[16]. And the customers taking Nvidia's money are building exits from its chips. Anthropic's IPO filing shows a single compute pool spread across four chip families — Amazon's, Google's, AMD's, and Nvidia's — expressly to avoid depending on any one supplier ^[17]. Even Musk, who said SpaceX would build exclusively on Nvidia's Vera Rubin architecture ^[18], is standing up a chip factory of his own.

We’re very grateful to our existing supply chain, to Samsung, TSMC, Micron and others. But there’s a maximum rate at which they’re comfortable expanding. That rate is much less than we would like. And we need the chips, so we’re going to build the Terafab. — Elon Musk

The backstops run to 2032 ^[7]. By then a chip bought this year will have had six years to answer the only question that mattered — and the three-to-ten argument will no longer be an argument.