Silicon as Collateral

Fungible. Wheat is fungible. Crude oil is fungible. The word means the units are interchangeable — priced as a class, not a batch, because what matters is the thing, not whose hand it's in. No chipmaker used it while a chip was equipment, something a buyer depreciated on its own books as it aged toward obsolescence. Then, this year, Nvidia began describing its products in the language of a commodity warehouse.

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

The third adjective is the announcement. A lender doesn't finance a depreciating machine; it finances an asset it can seize and resell to the next owner. Nvidia was no longer describing chips. It was describing collateral. The shift was forced before it was marketed. Capital spending on AI infrastructure now outruns revenue at Amazon, Google, Meta and Microsoft, draining cash reserves and pushing all but Microsoft into negative free cash flow ^[1]. The giants closed part of the gap in the bond market — Amazon and Alphabet sold record amounts of debt in euros, yen and sterling to fund an estimated $725 billion in 2026 spending ^[1]. Apollo's chief economist, Torsten Slok, is watching what the credit market thinks of that, and an early verdict is already visible: the gap between what it costs to insure hyperscaler debt against default and what it costs to insure bank debt has widened to roughly 60 basis points since last October ^[2].

What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets. — Torsten Slok

Consider two loans. In February, Blue Owl tried to raise $4 billion for a data center anchored by CoreWeave, and lenders walked — not because of the building, but because CoreWeave's S&P rating is B+, below investment grade, leaving a half-billion-dollar bridge loan hanging ^[3]. This month, SpaceX — a name nobody discounts — moved to raise about $40 billion in debt to buy Nvidia chips, and even there the security was the chips: the money sits in a special-purpose vehicle, a shell company that holds the GPUs as collateral ^[4]. Weak name or strong, the loan now runs against the silicon. So Nvidia became a bank. It has committed $36 billion in take-or-pay demand guarantees — six-year pledges to pay its partners for capacity whether or not it is used — which its cloud customers use to secure their own borrowing ^[5][6]. It agreed to backstop up to $105 billion of the Portsmouth, Ohio data center where OpenAI is expected to buy $350 billion in chips ^[5]. Through its DSX program it hands startups GPUs with no upfront cash, taking a share of their revenue instead ^[7]. And it has put nearly $50 billion into the AI labs that are also its customers ^[5]. Huang knows exactly how this looks.

We recognize the scale of this support, and we know some will call this circular financing. We see it differently. — Jensen Huang

That is the chief executive naming his own critique and answering it in a breath. The chain doesn't stop at a supplier financing its buyers. Google is now extracting equity warrants — options to buy shares later at a fixed price — from chipmakers AMD and Marvell as a condition of purchase commitments, the customer financing the supplier, the money flowing both ways ^[8]. Beneath that, Wall Street built a credit market around the chips themselves. The centerpiece is a $500 billion program to package GPU-backed loans into securities — a process called securitization — and sell them to insurers and private-credit funds ^[9]. Banks that once lent against a corporate name now lend against a rack of H100s, and Morgan Stanley says it plainly.

This is a capital-heavy, capital-dependent industry, and we are here to help clients raise, syndicate and underwrite that capital needs, and to find offsets for the risks. — Morgan Stanley

Then the descent continues. There are futures on GPU rental prices now — Kalshi and Polymarket list contracts on Nvidia H200s, and CME plans compute futures — though Kalshi's notional volume is still around $100,000, so this rung is embryonic ^[10]. There is tokenized GPU debt: Bullish, a crypto exchange, extended a $100 million facility to USD.AI, which makes non-recourse loans — loans secured only by the chips, never the borrower — against GPU assets ^[11]. USD.AI's chief executive, David Choinière, says his backer sees it plainly.

Bullish recognizes that compute is becoming a credit market in its own right. — David Choinière

Nebius prices its GPU capacity by auction and takes customer prepayments, cutting its payback period to a year and ten months ^[12]. A company founded by ex-Coinbase employees, backed by Pantera and Coinbase Ventures, sells H200 rigs to university researchers at 30% down with five-year financing ^[13]. A decentralized marketplace lets households rent out idle gaming-PC GPUs at half to a fifth of cloud prices ^[14]. From a $500 billion securitization to a professor financing a rack like a car, crypto-market people and money staff the rungs. Every structure in that stack is priced off one number: what a used chip is worth. And it is the one number nobody knows. Huang's pitch is that a GPU earns for a decade, the way a leased aircraft does. The banks don't buy it. They have no historical loss data for silicon and no deep secondary market in which to resell it, so they demand three-to-four-year depreciation schedules — and their insistence on higher rates and stronger guarantees has already forced Nvidia to revise the $500 billion plan ^[15]. The aircraft comparison has a flaw: aircraft have decades of loss data and a functioning resale market. A five-year-old GPU has neither. Each side has its exhibit. CoreWeave has signed contracts renting 2020-era A100s — chips now six years old — into 2029, at or above their original prices, and its chief executive insists the old silicon will outlast expectations ^[16].

We expect demand to meaningfully exceed supply for years. — Mike Intrator

Against that stands Michael Burry, who claims the hyperscalers understated depreciation by $176 billion ^[17]. Alphabet says its own TPU chips deliver 1.5 times the compute per dollar of rivals — and if custom silicon wins, it flattens the resale value of exactly the chips being collateralized ^[18]. Corporate buyers are already flinching at token bills and routing work to cheaper models or their own hardware, shrinking demand for premium compute ^[19] — the labs' scramble to pay these bills before they come due is the other half of this story ^[20]. Bain puts the whole buildout as a question of timing.

The question is whether the applications arrive in time to pay for it — Bain & Company, Inc.

One giant is the exception that keeps the picture honest: Microsoft stays free-cash-flow positive, holds a fixed-rate debt portfolio, and monetizes AI through subscriptions — paying as it goes while the others borrow ^[21]. The machinery has already survived one scare. In December 2025, a failed Oracle–Blue Owl data-center deal was read as the first pop of an AI infrastructure bubble ^[22]. The buildout did not shrink. It flinched, then accelerated — you can see both inside April's record $16.3 billion package for an Oracle data center in Michigan, where PIMCO anchored roughly $10 billion after other U.S. banks withdrew over AI-demand concerns ^[23]. The banks stepped back and the money still found its way in. The number the entire stack is priced off is sitting in a data center right now: an A100, five years old, the oldest silicon in the fleet, carrying rental contracts that run to 2029 and loans that don't care what the chip can do — only what the next buyer will pay for it. The securitizations, the futures, the tokenized debt, the rent-to-own racks all assume the answer is "enough." Nvidia chose the word fungible. The question nobody has answered yet is whether a used chip is fungible the way wheat is — or just a machine getting older.