Unscarcity
Sign in for free: Preamble (PDF, ebook & audiobook) + Forum access + Direct purchases Sign In

Unscarcity Research

Compute as Collateral: When Chips Become an Asset Class

Nvidia lined up $500B of Wall Street capital and may backstop $125B of it. When the seller guarantees the buyer's debt, the loan decides who gets compute.

15 min read 3395 words Updated September 2026 /a/compute-as-collateral

Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

Compute as Collateral: When Chips Become an Asset Class

Or: the company that sells the machines now offers to guarantee the loans that buy them, and the interest rate on that loan is quietly deciding who will be allowed to rent intelligence.

The Day Chips Became Bonds

On August 10, 2026, Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build “financing platforms” that would mobilize more than $500 billion of third-party capital for what the press release calls AI factories: data centers full of Nvidia systems, financed with other people’s money. Jensen Huang told CNBC the chips were “revenue-generating assets now. They’re productive, they’re long-lived, they’re fungible, they’re flexible.” Then came the detail that turns a sales channel into something stranger: Nvidia may backstop up to 25% of the deals, about $125 billion, absorbing part of the loss if the machines are worth less than the loans when the music stops.

Larry Fink, whose firm manages more money than any other on Earth, reached for the analogy himself: “This is the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s.”

He meant it as a compliment.

Two things are worth holding onto before the analysis. First, these are memorandums, not contracts. Nvidia signed a $100 billion memorandum to invest in OpenAI in September 2025; the Wall Street Journal later reported the deal was on ice, and Nvidia ended up putting $30 billion into OpenAI’s next round instead. Announcements are cheap. Second, even as a trial balloon, the announcement tells you what a company that has at times been valued above $5 trillion believes the constraint on its growth now is. It is not demand for intelligence. It is the supply of credit to buy the machines that produce it.

That is the moment a technology stops being equipment and becomes a financial instrument. This note is about what changes when it does, why we have seen this loop before, and why an asset class built on scarce compute is structurally at odds with the abundance it is supposed to deliver.

Three Moves That Turned a GPU Into a Security

The asset class did not appear on August 10. It was assembled over about three years, in three moves.

Move one: lend against the chip. In August 2023, CoreWeave borrowed against its GPUs at a floating rate of about 15%, the price a lender charges when it is nervous. On March 31, 2026, the same company closed an $8.5 billion delayed-draw term loan secured by comparable hardware, and Moody’s rated it A3 while DBRS rated it A (low). The floating portion priced at SOFR plus 2.25%, the fixed portion at about 5.9%. It was the first investment-grade rating ever assigned to debt backed by GPUs, and investment grade is a threshold, not a compliment. Pension funds and insurers operate under rules that let them buy only above a certain rating, so the day that paper was rated, retirement savings became eligible to fund graphics cards.

Read the fine print, though, and the collateral is not really the silicon. CoreWeave’s own announcement describes the facility as secured by the hardware “and an associated customer contract.” The rating reflects the creditworthiness of whoever signed that contract. As one analyst told The Verge, in most deals so far the chips alone weren’t enough; lenders also needed contractual cash flow on those chips. The chips were the story. The tenant was the collateral.

Move two: move the debt off the balance sheet. In October 2025, Meta financed its Hyperion campus in Louisiana with a structure familiar to any project-finance banker and alien to anyone who thought of Meta as a company that pays cash. Morgan Stanley arranged about $27 billion of debt and $2.5 billion of equity into a special purpose vehicle, Blue Owl Capital took 80% of it, and Meta kept 20%. Meta is the anchor tenant, paying rent to a vehicle it does not consolidate, with a lease of at least four years and, by one trade publication’s reading, financial exposure that runs sixteen. The debt exists. It is just not on the sheet where you would look for it.

Move three: the seller guarantees the buyer. This is the move that makes the AI buildout unlike the ordinary infrastructure booms it gets compared to. In September 2025, Nvidia agreed to buy up to $6.3 billion of CoreWeave’s unsold capacity through April 2032, a promise to be the customer of last resort for a customer that buys its chips. It took stakes in OpenAI, CoreWeave, Nebius, and Anthropic, each of which then spent the capital on Nvidia hardware. Axios reported talks about Nvidia guaranteeing financing for a quarter-trillion-dollar OpenAI data center. And now it offers to backstop a quarter of a $500 billion lending program against its own products.

An analyst at Vested Finance described the position as an insurer writing earthquake policies on buildings in its own city. That is unkind but not wrong. Nvidia profits enormously if demand stays strong and is exposed, through backstops it wrote itself, precisely when demand weakens. MacroMicro’s phrase for the resulting role is “an AI Federal Reserve”: a private lender of last resort for compute, with none of a central bank’s accountability and all of its power to decide who gets liquidity.

The Number That Decides Everything

Every one of those structures rests on one assumption: how long a chip keeps earning.

That assumption is a choice about accounting, and the choice has swung wildly. At Nvidia’s own conference in March 2025, Huang said that once Blackwell shipped in volume, “you couldn’t give Hoppers away.” Seventeen months later, the same Hoppers are “long-lived” collateral. Michael Burry, who made his name on the last asset class Larry Fink was comparing this one to, estimated in November 2025 that the big cloud companies were understating AI depreciation by about $176 billion between 2026 and 2028 by stretching the assumed useful life of their accelerators. Nvidia argues the reverse: that CUDA software keeps improving the installed base, that A100 chips from 2020 still draw multi-year commitments, and that useful life can stretch toward a decade.

The secondary market has an opinion, and it is a mixed one. An H100 that sold for roughly $30,000 in 2023 changed hands for about $8,000 by mid-2026, a 73% loss in three years. Its rental rate peaked near $8 an hour, collapsed to between $1 and $2, then recovered to about $2.35 on annual contracts as demand outran the supply of new chips. Silicon Data projects that rents on older chips keep rising through 2028; one cloud provider nearly doubled its Blackwell B200 price on a renewal. So both sides are right about different things. Resale value fell like a phone’s. Rental income held up like a warehouse’s, because there are not enough warehouses.

Notice what that means for the asset class. The chips are worth financing only as long as compute stays scarce. The moment the shortage clears, whether because fabs catch up, because Chinese open-weight models keep needing less compute for the same output, or because a frontier lab stops being able to pay its rent, the residual value that lenders modeled evaporates and the backstop gets called. An asset priced on scarcity cannot be the instrument that delivers abundance. The two are the same number with opposite signs.

We Ran This Loop in 1999

The telecom equipment industry ran the identical structure a generation ago, and it is worth being precise about how it ended.

By the end of 2000, nine equipment vendors (Lucent, Nortel, Cisco, Alcatel, Ericsson, Motorola, Nokia, Qualcomm, and Siemens) had about $25.6 billion of loans to their own customers on their books. For the five North American firms, vendor financing equaled 123% of their 1999 pretax earnings. Roughly a third of it went to start-ups: competitive local exchange carriers, fixed-wireless providers, dot-coms. Nortel’s financing offers reached as high as 130% of the equipment’s cost. One contemporary account put it plainly: at some point Lucent wasn’t selling equipment anymore, it was giving it away and labeling it a sale.

Lucent committed $2 billion to WinStar Communications. When WinStar struggled, Lucent refused a final $90 million extension. WinStar filed for bankruptcy, and Lucent wrote off $700 million. Its provisions for bad customer debt reached $2.2 billion in 2001 and $1.3 billion in 2002. Roughly four dozen CLECs went bankrupt between 2000 and 2003, and the unwind was synchronized, because the same funding source dried up for all of them at once. The early warning, visible before any default, was the loosening of terms: a vendor’s willingness to finance a larger share of the purchase on weaker security to keep a shaky customer buying.

A residual-value guarantee on a quarter of the collateral is a loosening of terms. So is a backstop for unsold capacity. So is investing in your customer’s equity so it can afford your invoice. The vocabulary is new; the mechanism is 1999, at roughly ten times the scale, with the pension system invited in through the investment-grade door that CoreWeave opened.

There is one hopeful reading of the telecom bust, and it matters for the book’s argument. The fiber laid in 1999 did not disappear when its owners went bankrupt. It was bought out of bankruptcy for cents on the dollar, and the glut became the cheap bandwidth that made the next twenty years of the internet possible. Overbuild, default, and abundance for everyone else: that is the abundance J-curve in its purest form. The problem is that glass does not depreciate and silicon does. A strand of fiber from 1999 still carries light. A GPU from 2026 will be a space heater in 2031. If this cycle busts, the stranded asset will be far less useful to the people who pick it up, which means the abundance dividend of an AI overbuild is smaller than telecom’s, and the write-off is larger.

Who Is Allowed to Rent Intelligence

Set aside the question of whether it ends badly. Suppose it works. Financing structure still decides something the book cares about more than returns: who gets access.

Usage-linked debt needs a creditworthy user. The loan is really against the contract, and the contract has to be signed by someone a rating agency trusts. That is why CoreWeave’s anchor customers are Microsoft, Meta, and OpenAI, why Anthropic pays SpaceX $1.25 billion a month for Colossus, and why Google rents bridge capacity from a competitor: the tenants of the compute economy are titans, because only titans make the paper investment grade. A financing platform designed to “broaden access” to AI factories, in Nvidia’s words, will lend most cheaply to the customers who already have the most, since they are the ones whose signatures make the collateral good.

The scale of the debt makes this a public question rather than a corporate one. The five hyperscalers issued a record $121 billion of bonds in 2025, more than four times their 2020-2024 average, and had already issued about $220 billion in the first eight months of 2026, according to LSEG. JPMorgan expects the year to end near $317 billion, and Nvidia itself sold $25 billion of bonds in June. Global corporate bond issuance hit a record $4.9 trillion by early September, and the AI buildout is the reason. Morgan Stanley puts hyperscaler capital spending at roughly $800 billion this year and $1.1 trillion next; UBS estimates Amazon, Alphabet, and Microsoft will spend about 102% of their cloud revenue on capex in 2026. McKinsey’s ceiling for the decade is $7 trillion. Against all of that, OpenAI, the marquee tenant, took in about $13 billion of revenue in 2025 while losing about $21 billion at the operating level, and its chief executive has spoken of $1.4 trillion in infrastructure commitments over eight years.

China has noticed which lever matters. Commentators there greeted Nvidia’s announcement as a setback for domestic chips and called for a national computing-power exchange, with futures, options, and eventually securitized compute assets of its own. When your rival turns its product into a bond, a better chip is no longer a sufficient answer. You need a bond market.

The Unscarcity Read

The book’s framework has three things to say here, and none of them requires the financiers to be villains.

Truth Must Be Seen. The second of the Five Laws is a transparency mandate, and the AI capital stack is being built in a way that defeats it. Off-balance-sheet vehicles, non-recourse project companies, vendor backstops disclosed in footnotes, useful-life assumptions that move $176 billion of reported profit: each is legal, and together they make it impossible for a citizen, a regulator, or a pension trustee to see where the risk actually sits. The fix starts with disclosure rather than prohibition: publish the residual-value assumption behind every rated compute loan, publish the backstops, consolidate the SPVs in the tenant’s accounts. If chips are an asset class, they can carry the reporting obligations of one.

Power Must Decay. A vendor that supplies the hardware, invests in the customers, guarantees their lenders, and buys their unsold output has become a private central bank for the most important input of the century, and there is no mechanism by which that position weakens over time. The book’s third law says concentrated authority must have a half-life. The historical answer for infrastructure this essential was common carriage: the railroads, then the telephone network, then the grid were forced into posted, nondiscriminatory pricing once they became the thing everyone needed. Compute financed by retirement money should carry the same condition. If pension-eligible debt funds an AI factory, the factory takes tenants on published terms.

Stewardship, not ownership. Chapter 9 of the book replaces ownership of productive assets with stewardship. A Mission Guild holds its factories the way a church holds its buildings, oriented toward the mission rather than the balance sheet, and the MOSAIC can step in if the Guild drifts from that mission; Transition Trusts convert financial ownership into stewardship rights over a decade. The compute buildout is the first test of that idea at scale, and it arrives with a ready-made transition. If the 1999 pattern repeats and the compute glut is sold out of bankruptcy at residual value, the buyer of that stranded capacity can be a public trust rather than the next landlord, and the glut can become the Foundation’s compute layer: metered, priced at cost, open to anyone. That is how fiber became abundance last time. Nobody planned it. This time we could.

What to Watch

  • Whether the memorandums become platforms. The OpenAI $100 billion memorandum did not. Watch for signed vehicles with named borrowers, not press releases.
  • Whether chips alone get financed. Every rated deal so far has needed a customer contract behind the hardware. The day a lender accepts GPUs without a hyperscaler’s signature is the day the collateral has genuinely changed, and the day the risk has, too.
  • The useful-life number. Hyperscaler depreciation schedules and the residual-value assumptions in rated compute debt are the load-bearing figures of the whole structure. Any lengthening should be read as a loosening of terms.
  • Backstop drawdowns. The first time Nvidia actually buys unsold CoreWeave capacity, or pays out on a residual-value guarantee, is the first crack.
  • Rental price indices. Silicon Data’s H100 index is the closest thing to a market price for compute. Falling rents on old chips while new supply lands is the 1999 signal.
  • Who is in the paper. Pension funds and insurers can now buy GPU-backed debt. If this cycle cracks, the losses will not sit with banks that remember 2008. They will sit on balance sheets that were told this was infrastructure.

Sources

Every infrastructure boom ends with the same question: who owns the stranded capacity, and on what terms do they let the rest of us use it. In 1999 nobody planned the answer, and we got cheap bandwidth by accident. This time the collateral depreciates like a phone, the lenders include your pension, and the seller has written the insurance. The accident will not repeat itself. The plan has to.

Share this article:

Share Share Share