Nvidia has committed to cover up to 25% of hardware value losses across its AI infrastructure financing push — a backstop that could reach $125 billion, with much of that risk landing on insurers' balance sheets, according to Axios.

On August 10, Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build financing platforms meant to mobilize more than $500 billion in third-party capital for AI infrastructure, Axios reported. The structure is simple: raise money from pension funds, sovereign wealth funds and insurers, buy Nvidia GPUs, build data centers, and lease capacity to AI companies for long-term payments.

The sweetener for investors who don't normally underwrite chip depreciation is buried in the deal terms. Nvidia said it "may provide a residual-value support mechanism for up to 25% of an opportunity, assessed carefully on a project-by-project basis." That guarantee covers part of the shortfall if the hardware is worth less than expected when a lease ends. Aggregated across every deal Nvidia is courting, the exposure could total $125 billion, per Axios.

Who actually holds the risk?

None of the $500 billion sits on Nvidia's balance sheet, and neither does most of the credit risk behind it. Special purpose vehicles buy the GPUs and issue the debt. Increasingly, that debt is placed with insurance and retirement capital managed by firms like Apollo and KKR, according to CNBC reporting from April 2026.

The match is not accidental. Data centers are decades-spanning projects that pair naturally with the long-dated liabilities insurers already carry — annuities and pension obligations. The risk doesn't disappear. It changes owners and arrives on an insurer's books relabeled as investment-grade fixed income.

CNBC's April report quotes the phrase "GPU debt treadmill," coined by AI commentator Dave Friedman, to describe the core mismatch. GPUs last roughly seven years. The buildings housing them last 20 to 30. Anyone underwriting a 20-year facility is betting that today's chips, or their replacements, still hold value in year fifteen — after the hardware has been swapped out two or three times.

How fast is the debt growing?

The leverage underneath this build-out has scaled quickly:

• Incremental borrowing funded about 9% of hyperscaler capital spending in fiscal 2024.
• By mid-2026, that figure reached roughly 32% on a trailing basis.
• U.S. data-center debt issuance roughly doubled to about $182 billion in 2025, per reporting Axios cited.

Capacity itself is a live underwriting problem. One insurance executive told CNBC that insuring a single $20 billion data center campus was nearly impossible to price in 2023. By 2026, in that executive's words, it has become a weekly conversation. Insurers have responded with new products: cover for credit losses, cover for declines in chip resale value, and cover for contract breaches tied to power outages or cooling failures, according to the same CNBC report.

What justifies the spending?

Bain & Company said this week that the AI industry must generate $6 trillion a year in new revenue by 2031 to justify data center spending already committed. Existing AI services might cover $1.8 trillion, leaving a $4.2 trillion gap to fill from businesses that barely exist yet, such as autonomous robotics and AI-driven drug discovery. Bain separately projects $5 trillion to $6.5 trillion in data center spending through 2030.

Those figures explain why insurers are being asked to carry so much weight. Traditional lenders and Nvidia's own balance sheet cannot absorb spending at that scale alone.

Is any of the money committed?

The arrangement doesn't eliminate risk; it reroutes it from Nvidia's income statement and the banks into the reserves of companies whose business model depends on pricing tail risk decades ahead. If AI demand compounds the way Nvidia's revenue guidance assumes, the residual-value guarantee sits mostly unused — a marketing device that made institutional capital comfortable enough to write the check. If a major AI lab pulls back capacity, or a cheaper chip generation strands the current fleet, the 25% backstop gets tested for real.

None of the six firms named in the MOUs has disclosed how much capital they have actually committed, or whether the residual-value terms are finalized. Nvidia's own release says the partnerships remain subject to final agreements. The market has already priced in $500 billion of financing infrastructure and a $125 billion guarantee ceiling — and so far, not a single dollar of it is confirmed as spent.