Nvidia Is Turning Compute Into Collateral

A fully optimized data center running Nvidia's latest chips costs roughly $25 billion a year to rent and produces $23 billion in output. Before anyone builds a financial product on top of it, the infrastructure is already underwater ^[1]. The machine Nvidia has assembled over the past year does not change that arithmetic. It changes who holds the loss if the arithmetic never improves. The first instrument is a vendor-financing program called DSX, rolled out in July. It lets AI cloud providers and startups acquire chips and token credits without paying upfront, in exchange for a share of whatever their products earn ^[2]. The customer gets compute with no capital outlay; Nvidia gets a stake in the customer's revenue. Then came the platform built with six Wall Street firms — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — mobilizing more than $500 billion. The structure is explicit about who holds the risk: external consortium partners take the primary financial risks and capital commitments, keeping the liabilities off Nvidia's balance sheet, while Nvidia provides only limited residual support ^[3]. Where Nvidia does step in, it caps its exposure at a slice of a project's value — potentially guaranteeing up to 25%, measured against the residual worth of the financed GPUs ^[4]. Above that sits a derivatives layer. CME Group is launching a GPU futures product, Silicon Data and Ornn are building compute futures markets with the major exchanges, and a Bermudan exchange is already trading compute while filings are in for ETFs that track it ^[5][6]. A GPU can now be bought, collateralized, and hedged like a barrel of oil. Nvidia's chairman names what all of this is for.

This is really the first time that technology chips have become an investable asset class. — Jensen Huang

In other words, a GPU has moved from fast-depreciating hardware to collateral with a residual value and a derivative market. And where there is collateral and a derivative market, there is a way to move the risk. One analyst puts the comparison without euphemism: the structure shifts risk from core companies to global investors — the same architecture as the mortgage-backed securities of 2005 to 2009 ^[7]. Except Nvidia has not fully escaped. Consider the deal it cut to secure OpenAI as an anchor tenant: a $105 billion backstop and residual value guarantees on a data center leased to OpenAI for twenty years ^[8]. Rating agencies have flagged these off-balance-sheet commitments as phantom liabilities, roughly $70 billion across the major AI companies ^[9]. And the hardware underneath is proving more expensive to keep whole than the asset-class story assumes: Nvidia's GPU warranty costs rose more than 1,000% in 2025, from $81 million to $894 million ^[10]. Michael Burry argues the assets themselves may be overstated — that hyperscalers understated GPU depreciation by $176 billion, which, if true, is the exact mechanism of a collateral bubble ^[11]. The market is beginning to price what the structure moves. CoreWeave, a neocloud provider carrying $14 billion in debt, saw its credit default swaps climb nearly 60% in two months ^[12]. The borrowing is crowding out everyone else: CDS spreads on LVMH, Sanofi, and BAE Systems have risen more than 10% as hyperscaler issuance floods the investment-grade market ^[13]. The Bank of England is now reviewing lending practices in the sector.

Some data centers may be rendered uneconomic, and some owners may go bankrupt. — Oaktree Capital Management

Sam Altman, whose company plans $1.4 trillion in infrastructure spending, has said the wave reminds him of the dot-com era.

someone’s gonna get burned there. — Sam Altman

Nvidia has built a machine that keeps the revenue from selling chips and distributes the risk that those chips won't earn enough to cover the debt they back. The one piece of risk it kept for itself — the guarantees and the backstops — is the piece that turns systemic if the revenue falls short.