Key Takeaways:
Nvidia is selling Wall Street on a new idea: that a GPU is a revenue-generating asset a lender can size, seize, and resell — not a chip that loses half its value every 18 months.
Nvidia signed memorandums of understanding on Aug. 10 with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build compute financing platforms that aim to mobilize more than $500 billion of third-party capital for AI infrastructure, turning the chipmaker into an underwriter of its own hardware.
"The reason for that is because it's productive, it's revenue generating, it is fungible, it's used by just about every cloud service provider," Nvidia Chief Executive Jensen Huang said of the AI factory platform, framing compute as an investable asset class. Goldman Sachs Chairman David Solomon called it "a market for credit backed by NVIDIA compute."
The structure works like a leveraged lease. Nvidia will guarantee up to 25 percent of the residual value of GPUs used as collateral in certain deals, a credit enhancement that lets lenders underwrite depreciation they otherwise cannot forecast. Borrowers — the CoreWeaves, Nebiuses and sovereign clouds that sit between hyperscaler cash flow and venture capital — repay debt from future compute rental income, closing a loop of collateralized financing, GPU purchases, and rental servicing.
The financing bet rests on a recovery in rental prices. Hourly rates for Nvidia's main GPUs have rebounded from late-2025 lows, with the H100 climbing from about $2 to $2.72, the H200 at $3.29, and the Blackwell-architecture B200 holding at $5.61 on initial capacity scarcity, according to Silicon Data's Neo-Cloud price index. The estimated residual value of a single H100, based on a 36-month forward rental curve, has recovered from about $14,000 in October to roughly $20,000 today.
The model already has a live proof point. CoreWeave mortgaged its H100 cluster to Blackstone and Magnetar Capital for $2.3 billion in debt financing in August 2023, and its shares have risen about 18 percent since its second-quarter beat and more than 1.65-fold since its March IPO. In May, CoreWeave closed DDTL 5.0, a $3.1 billion public syndicated GPU-collateralized facility with OpenAI and Cohere as underlying borrowers — the first of what the market now calls collateralized compute obligations, or CCOs.
The bear case is that GPU depreciation is unpredictable because Nvidia controls the roadmap. Used H100s with about three years of use trade at only 20 to 30 percent of peak new prices, and Amazon shortened server depreciation from six years to five, adding $1.4 billion to its depreciation expense. Short seller Michael Burry has said cloud providers understate AI chip depreciation.
Supporters counter that inference demand extends GPU life. The A100, six years after launch, has left the high-end training market but still holds stable rental prices in inference and fine-tuning workloads, with KKR's Waldemar Szlezak noting A100s remain revenue-generating six or seven years after launch. The DSX reference design, Nvidia's standardized AI-factory blueprint, is quietly becoming the underwriting document — a lender can price projected token revenue per megawatt against a validated build, and an appraiser can mark the asset against a fleet of identical systems.
The stakes are large. Futurum Group estimates $7.9 trillion in global data center capex from 2026 to 2030, against which $500 billion covers roughly 6 percent of the projected debt stack — a wedge sized to prove the asset class rather than fund the buildout. The risk is concentration: about $1.05 trillion of the $2.1 trillion in remaining performance obligations across four hyperscalers comes from OpenAI and Anthropic, and OpenAI's reliance on ever-higher valuations — about $852 billion by spring 2026, with an IPO pushed to 2027 — makes it the cycle's most exposed borrower. Microsoft removed its structural support in April, and Morgan Stanley estimates investment-grade hyperscaler debt leverage has doubled to about 1.8 times in a year.
Nvidia shares have recovered after early volatility as the market digested that the $500 billion is a financing target, not committed spending, subject to definitive agreements. The competitive field now includes cost of capital: AMD has no equivalent third-party financing platform, and an operator choosing between architectures will weigh financing spreads alongside tokens per megawatt. If Nvidia-backed paper prices inside everyone else's, the financing advantage compounds every quarter it goes unanswered — and the party that most accurately prices the economic life of compute will hold pricing power for the next phase of AI infrastructure.
This article is for informational purposes only and does not constitute investment advice.
