Invest1 publisher2 min readPublished
Nvidia weighs absorbing a quarter of GPU value loss to get lenders into AI factories
The six asset managers that signed August memoranda for more than $500 billion of AI infrastructure capital still need loan documents that can value a rack of chips a successor part may supersede in 18 months.
The Investor · Invest desk
What happened
- Nvidia said on August 10, 2026 that it had signed memoranda of understanding with six major asset managers, including Apollo and BlackRock, targeting more than $500 billion of third-party AI infrastructure capital.
- Nvidia's AI Factory Research Center sits on Digital Realty's Northern Virginia campus, and the two are developing the Omniverse DSX blueprint for gigawatt-class sites aimed at customers other than the cloud giants.
- Nvidia disclosed about $530 billion of gross off-balance-sheet guarantees as of the second quarter of fiscal 2027, tied to a growing web of AI and cloud project commitments.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint The capital cannot move faster than the paperwork: loan documents, valuation methods and a resale market for used accelerators all have to be written before a debt tranche exists at all.
- exposure Nvidia's disclosed guarantee book is already about $30 billion larger than the whole pool the memoranda are meant to raise, so the company's contingent liabilities are where this buildout lands first.
- decision A lender at the table has to decide whether it is underwriting the shell and the land or the chips inside, and those two answers support very different loan sizes.
- precedent If vendor residual support becomes a condition of closing, the chip maker becomes a standing credit counterparty in every non-hyperscale data centre financing that follows.
An 18-month supersession window is a tenor problem before it is a valuation problem. To recover principal inside it, a loan has to amortize at roughly 67 percent of face a year [1]. Cryptobriefing.com notes that the H100, which commanded premium prices not long ago, has already seen its market value drop sharply as newer chips arrived [6].
Nvidia's answer, according to cryptobriefing.com, is residual-value support of up to 25 percent on specific projects, absorbing part of the downside if the hardware falls faster than expected [8]. That leaves 75 percent of any decline with the lender [2]. Run the 25 percent across the entire $500 billion and the contingent figure is $125 billion, about 24 percent on top of the guarantees Nvidia has already disclosed [4]. The support is described as project-specific, so treat that as the ceiling case.
Digital Realty's vehicle shows where the strain sits. It closed $3.25 billion in March 2026 and says the fund is designed to support more than $10 billion of new AI infrastructure investment, which is about 3.1 times the equity, or $2.08 of non-equity capital behind every dollar committed [4][5]. The equity is closed. The other $6.75 billion is the part that needs a lender willing to hold a view on chip residuals, and for scale, $3.25 billion is 0.65 percent of the $500 billion Nvidia is trying to organize [6].
The target customers here are enterprises, governments and research institutions that need serious GPU capacity without building their own sovereign clouds [2], sitting one tier below the mega-campuses, and their landlord in this case is a listed REIT working from Nvidia's Omniverse DSX blueprint [3].
This can go three ways. Accelerators get underwritten the way aircraft are, with observable resale prices and residual-value insurers, and the 25 percent backstop turns out to be cheap. Or the memoranda convert slowly, project by project, and the non-hyperscale tier gets built mostly on equity and vendor paper. Or lenders keep declining chip collateral and the $500 billion stays a target. I would put the weight on the second, because cryptobriefing.com reports that the loan documentation, the valuation frameworks and the secondary market for AI hardware as collateral are all still ahead [9].
What would move me off that view: a syndicated facility secured primarily on GPUs, at size, priced without vendor residual support and with a tenor longer than 18 months.
What to watch
- Conversion of any of the six August memoranda into a signed and funded commitment, and at what project size.
- The terms of the first loan facility that takes GPUs as primary collateral, and whether Nvidia's residual-value support appears in the documents.
- Nvidia's next disclosure of gross off-balance-sheet guarantees, measured against the $530 billion reported for the second quarter of fiscal 2027.