On August 11, Nvidia joined forces with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to stand up independent financing platforms aimed at mobilizing more than $500 billion in third-party capital for AI infrastructure. Four days later, facing a market skeptical of "circular financing," Jensen Huang stepped in personally: Nvidia may provide a residual-value support mechanism of up to 25% per project, assessed case by case.
Read as a vendor helping customers finance purchases, the news misses its actual weight. Huang is not selling GPU units; he is redefining AI compute from an operating cost into an investable asset class. When the chipmaker itself underwrites the residual value of its own hardware, the AI buildout has reached the point where financial engineering is what keeps the capital loop turning.
The 25% backstop: what it actually guarantees
The residual-value backstop exists to ease lenders' fear that AI hardware will be worth less tomorrow than the financing model assumes. If a project defaults and assets are liquidated, and the realized value of the GPUs falls short of the model's expectation, Nvidia's mechanism can cover part of that shortfall. It does not mean Nvidia carries the financing risk of the whole project.
Strip down the structure: participating institutions independently underwrite each deal (customer profile, compute demand, utilization, cash flow, residual value); Nvidia supplies the AI-factory platform — GPUs plus high-speed networking, system software, AI frameworks and the CUDA ecosystem; long-duration capital comes from the institutions. In some cases Nvidia will cover up to 25% of the residual-value shortfall, assessed project by project, complementing rather than replacing the lenders' own risk analysis.
Functionally, the 25% works as a first-loss cushion. It makes senior lenders comfortable enough to fund the rest at better rates, while the remaining exposure sits inside special purpose vehicles (SPVs) — legal shells that hold the assets and debt off the parent's balance sheet. Analysts have compared the plumbing to mortgage securitization: the underlying assets are GPUs and data-center capacity, and investors are buying the future cash flow of compute rent.
From buying chips to buying compute: Huang's asset-duration argument
Huang's case deserves a fair hearing, not because he is powerful, but because he is asserting a premise the industry had quietly assumed false: a GPU is a long-lived asset, not a consumable.
"These systems are not like our PCs, not like our phones," Huang told CNBC. "These are revenue-generating assets now. They're productive, they're long lived, they're fungible, they're flexible." Three pillars hold the argument up:
First, reusability. An AI factory is more than GPUs — it bundles high-speed networking, system software, AI frameworks and the CUDA ecosystem, and can run language, vision, speech, biology and robotics workloads. When one customer's demand shifts, capacity can in principle be redeployed to another tenant, cloud or operator. The asset is not single-tenant.
Second, software extends economic life. Nvidia points to A100: launched in 2020, still in production for training, fine-tuning, inference and HPC in 2026, with customers signing multi-year capacity deals. Software iterations let "old" chips keep earning rent.
Third, compute is revenue, not cost. "In AI, compute is revenue." Once compute is defined as a revenue generator, it migrates from an expense line on the P&L to an asset line on the balance sheet — the accounting foundation of an "investable asset class."
Moat, or circular financing with a nicer name?
The market's skepticism is equally real. Nvidia's stock dipped about 3% after the announcement — a signal that investors do not fully buy the "independent underwriting" framing. Three objections stand out:
Circular financing. Nvidia sells chips to data-center operators; operators pledge those chips as collateral for financing; the financing buys more Nvidia chips. The money flows back to Nvidia while risk is financialized out to third parties. Nvidia's answer: the platforms run on independent long-duration capital, each deal is underwritten by institutions, and Nvidia does not decide who gets funded. Critics counter that independent capital is not independent demand — capital can extend the buildout, but it cannot create downstream usage.
Asset-duration mismatch. This is the sharpest attack. A data-center GPU is obsolete in 3-5 years; a toll road or power grid is a 30-50-year asset. Pricing five-year hardware with infrastructure financing, if demand disappoints or a new architecture renders today's fleets obsolete faster than depreciation schedules assume, turns the 25% promise into real money against declining revenue. Huang's railroad-and-airline analogy is a bet on a specific timeline. Buggy whips also had a long-term story until they didn't.
Is the backstop even a guarantee? Huang has not said how the mechanism works or who takes the first loss if a deal fails. Until contracts are public, this is an intention, not a guarantee. Still, the signal is real: outside money may only show up because Nvidia is willing to cover part of the downside.
The same coin as the $2 trillion off-balance-sheet wave
Zoom out and this sits on one narrative with the AI debt wave — roughly $2 trillion hidden off balance sheets we analyzed recently. Two faces of the same macro story:
On the demand side, AI giants finance by issuing debt, letting leverage grow off the balance sheet and keeping risk out of view. On the supply side, the chipmaker actively underwrites residual value to keep capital flowing. Together they show the AI buildout's capital loop is now deeply financialized: debt upstream, securitization midstream, compute rent downstream.
This is not a binary "bubble vs. solid" call. It is a subtler signal: when the industry's most powerful vendor needs financial engineering to sustain its expansion narrative, the capital cycle has slid from the technology-dividend phase into the financial-engineering phase. Financial engineering is not inherently bad — it can bring capital into AI infrastructure at lower cost. But its premise is that the underlying asset is worth the duration being financed. Whether a GPU can be held like a toll road for 30 years decides whether this is a great financial innovation or an elaborate Ponzi structure.
Who the financialization of AI capex touches
Whatever the eventual outcome, the move is already reshaping the ecosystem:
NeoCloud operators get cheaper money. Previously they borrowed on their own credit ratings; now Nvidia's backstop and standardized platforms unlock more attractive rates. Nebius posted 454% revenue growth last quarter, and players like CoreWeave get fresh ammunition for their expansion story.
Enterprises and institutional investors see compute appear for the first time as an "investable asset" — no longer a capex game only giants can play, but an allocation that can sit in institutional portfolios. This is the capital face of Huang's tokenomics: tokens are currency, compute is an income-producing asset.
Regulators should watch the structure closely. SPV isolation plus a residual-value backstop plus securitization analogies is a combination with a familiar smell — the structured-finance toolbox that preceded 2008. The difference: the underlying asset is a GPU, and GPUs obsolesce an order of magnitude faster than real estate.
Huang's one-line summary of the shift: "We have moved from an era in which companies bought chips and built data centers project by project to one in which AI factories can be financed as productive infrastructure." Whether that holds is not decided by how compelling the narrative is, but by whether the pledged GPUs are still earning rent five years from now.
