GPU financing: the five structural issues lenders' counsel keep raising, and what each one asks for

Operators ask what their GPUs are worth. Lenders' counsel ask a different question. In White & Case's words, the questions are "What supports their value and repayment, and which protections will remain available if the transaction goes wrong?" (White & Case, 9 November 2026).

The alert, by Todd K. Wolynski and Brian Schmelkin, sets out five structural issues. Each is below, with what other credible voices have said about the same point and what a lender is likely to ask you for.

The five issues at a glance

1. Economic life: commercial, not technical

White & Case puts it first: the question is how long the GPUs stay economically useful for the workloads that repay the loan, "not simply how long they will remain technically functional." Its answer is structural. These assessments "should then inform advance rates, amortization, tenor and any balloon payment," backed by "periodic appraisals and loan-to-value covenants specifying the valuation methodology, testing frequency and consequences of a shortfall."

Others reach the same place from different directions:

• Quinn Emanuel (litigation alert, 13 March 2026) expects valuation disputes, because "no one agrees on what GPUs are actually worth." It adds: "That mismatch between accounting life, engineering life, and market life means that the same GPU portfolio may appear solvent on one balance sheet and distressed on another." (Quinn Emanuel)
• Pinsent Masons (Out-Law analysis, 18 September 2026): "There is currently no widely accepted framework for assessing the long-term value and durability of GPU compute capacity," and economic life "affects amortisation schedules, assumptions about residual value, and the appropriate tenor for any financing." (Pinsent Masons)
• Norton Rose Fulbright (October 2025) describes the market response: "AI infrastructure financings tend to have a short term (typically max. five years) and will typically amortise, either fully or with a relatively small tail." (Norton Rose Fulbright)
• Pillsbury (8 October 2026) offers the counterweight: "shorter technology cycles do not necessarily translate into a corresponding decline in residual value." (Pillsbury)
• KBRA frames the post-contract question as re-leasing risk, with revenue tied to "the economic competitiveness of the compute capacity." (KBRA AI Credit hub) Its July 2026 release says "the central question is whether the hardware and broader operating platform can continue producing competitive, cash-generating compute." (KBRA research press release, 27 July 2026)

Private credit is splitting on this point. On Apollo's third-quarter call (3 November 2026), President Jim Zelter said chip financing "that fully amortizes over a fixed period assumes no residual value and is further supported by guarantees from IG counterparties" is what Apollo will lend against (official FactSet corrected transcript). On Ares's call (29 October 2026), CEO Michael Arougheti said that, for the time being, Ares is sticking to data centers and power, the half of the market it understands best, and that he does not know how to underwrite the depreciation of that type of technology. (transcript, The Motley Fool; also reported by Bloomberg via Yahoo Finance and Semafor). UBS Asset Management's view is that structure can make the debt "worth owning where the structure matches the useful life of the asset and the lender is paid for what remains uncertain" (UBS, 29 September 2026).

For an operator, that means a lender will either size the loan so it does not depend on residual value, or test residual value on a schedule. Either way, it will want to know how value will be measured.

2. Repossession and recovery: a route, not a right

White & Case's test is practical: "who will locate, access, disconnect, package, transport, test and remarket the GPUs." It wants a recovery model that "accounts for downtime, unpaid facility charges, professional fees, deinstallation and transport costs, and any work required before resale." It also wants relocation limits, records of replacements "including make, model and serial numbers," arrangements with the host, and warranty transfers confirmed "before funding, rather than assume they accompany ownership of the hardware."

• Pillsbury lists what lenders should negotiate where the GPUs and the data center are financed separately, including "step-in rights and the ability to replace the data center operator, compute operator or other critical service provider" and "coordination of lender remedies so that enforcement against one project does not unnecessarily strand or destroy the value of the other."
• Quinn Emanuel expects recovery itself to be contested: borrowers "will contest whether the lender's valuation of the collateral was accurate, whether the lender waived the covenant breach by not acting sooner, and whether the lender's disposition of seized collateral was commercially reasonable."

What a lender will ask you for: a named site, who holds access, what the hosting agreement says on default, and whether the GPUs can be sold in place.

3. Revenue: committed, not forecast

White & Case's questions: "How much revenue is committed rather than forecast? Are customers required to pay for reserved capacity, or only for what they use?" And: "A long-term financing supported by shorter-term customer commitments should explicitly address the resulting renewal and repricing risk during the term of the financing." Where contracts are central, it points to "lockbox arrangements and financial coverage covenants."

• Pinsent Masons: "there is not yet a market-standard equivalent for GPU compute" to a long-term power purchase agreement.
• Norton Rose Fulbright notes that AI data center customers "are often new market entrants with a credit profile and track record which differs substantially from the traditional hyperscalers" and that "customer contracts tend to be shorter."
• UBS names the weak point: "Issuers to avoid are the ones funding long-lived infrastructure with short-lived technology and thin protection."

The public filings show what these protections look like in practice. CoreWeave's facilities are "entered into by bankruptcy-remote, special-purpose consolidated subsidiaries formed to hold the financed infrastructure and the related customer contracts," with debt sizing tied to "the depreciable cost of computing equipment, projected debt service coverage and project-level conditions" (CoreWeave 10-Q, quarter ended 30 June 2026, Note 10). We walk through those packages in what a GPU lender holds if offtake weakens.

4. Know exactly what is owned and financed

White & Case asks the documents to say whether the financed property is "individual cards, servers, networking equipment or a wider integrated system," backed by "an asset register supported by serial numbers or other reliable identifiers," reconciled to delivery, installation, acceptance and payment. Its test: "If the borrower or lessee defaults, can the financier identify the financed assets, establish its rights and distinguish those assets from third-party property?"

• Pillsbury says a repeatable GPU securitization would need "clear identification of the financed GPU systems by deployment location, serial number or otherwise" and arrangements "that trace the related customer payment rights and collections to the financed pool."
• Mayer Brown, writing on acquisitions, flags the obligations that sit beside the hardware: "Capacity reservations, pre-payment arrangements, residual value guarantees, and lease obligations that have not yet commenced can represent material future cash commitments that are invisible in standard financial statements." (Mayer Brown, June 2026)

What a lender will ask you for: the serial-number register, delivery and acceptance records, and the host's and landlord's position on the equipment.

5. Insurance: confirm it exists

White & Case: "A contractual insurance requirement, however, does not establish that the required coverage is available in the market." Insurers "may be unwilling to cover the full value of GPUs housed in a single data hall or in nearby halls exposed to the same incident." Any material uninsured exposure should shape the decision, including whether "a funded cash reserve" is needed.

The insurance market is saying the same:

• Clyde & Co (law firm, 23 March 2026): "Developers are seeking loss limits of US$10 billion for projects valued at c.US$30 billion—excluding GPUs. With GPUs included, exposure could reach $60 billion. Construction market capacity typically caps at US$4–4.5 billion." (Clyde & Co)
• Swiss Re Institute (27 March 2026) says financing institutions demand insured limits covering the full cost of construction, and "The re/insurance industry can only support a fraction of this limit at competitive rates for traditional construction risk policies." (Swiss Re sigma insights 07/2026)
• Marsh (13 July 2026): "While GPUs are arguably the single most critical asset in a data center, we do not yet have a mature market geared toward insuring these critical pieces of infrastructure." (Marsh)
• Lockton's James Nelson, as reported by Insurance Business (7 August 2026), said the global market may offer only about $15 billion to $20 billion of capacity in some way, shape, or form. (Insurance Business)
• Moody's 16 July 2026 webinar agenda (4:00 PM ET) posed as a discussion question: "How do emerging risks like adequate insurance coverage, lease transferability, and behind-the-meter power solutions affect credit?" The page is an agenda, not a Moody's finding that cover is unavailable. (Moody's)

Where a price index fits, and where it doesn't

White & Case asks for LTV covenants that specify "the valuation methodology, testing frequency and consequences of a shortfall." A test like that needs a value that is named, dated and repeatable. Published GPU price series now exist that a document can point to:

Be careful about what these measure. Most are rental prices, not resale values for used hardware. A rental index is an input to an income-based view of value and to repricing risk on renewal (issue 3). It is not, on its own, an appraisal of the collateral. Ornn says OCPI "measures on-demand rentals only," so a reserved multi-year block will price differently. A covenant can name an index as one input, but it still needs a stated method to turn that print into a collateral value.

Where Frontier Desk fits

Frontier Desk reads a GPU project the way a lender will, against these five issues, and is building a record of the named price series above so a value can carry a source and a date. As an advisor, we work on the structure around the hardware. The GPU financing tool on /compute shows what a lender is likely to ask for next, and the reading order for a full underwrite is in how lenders underwrite a GPU cluster loan. The lenders are in who lends against GPUs. Start with a free first look.

Questions

Is a six-year depreciation schedule what a lender will use? Not necessarily. White & Case asks financiers to size repayment to a conservative economic life, and Quinn Emanuel's alert shows how far apart accounting, engineering and market lives can be. Expect the lender to set its own view.

Will a lender lend against residual value at all? Some will not. Apollo says it lends on structures that assume no residual value, with investment-grade guarantees. Ares says that, for the time being, it is sticking to data centers and power, the half of the market it understands best. Others test value on a schedule through an LTV covenant.

My host has a lien. Does that matter? Yes. White & Case asks financiers to investigate "the contractual rights of suppliers, facility operators and other creditors, as well as the local rules governing landlord rights." Clear the host's position before the lender asks.

Can a price index replace an appraisal? No. A named index gives the test a dated source, but most published series track rental prices, not used-hardware sale values. The valuation method has to say how the print becomes a collateral value.

This guide summarizes public commentary. It is not legal advice.

Sources

All read on 10 November 2026 (ET) unless noted.

1. White & Case, Todd K. Wolynski and Brian Schmelkin, "GPU financing: Five structural issues every financier must address," 9 November 2026. https://www.whitecase.com/insight-alert/gpu-financing-five-structural-issues-every-financier-must-address
2. Quinn Emanuel, "Client Alert: Emerging Litigation Risks in Financing AI Data Centers Boom," 13 March 2026. https://www.quinnemanuel.com/the-firm/publications/client-alert-emerging-litigation-risks-in-financing-ai-data-centers-boom/
3. Pillsbury, Philip J. Tendler, "Financing Compute: Legal Risks and Opportunities in the Emerging GPU Asset Class," 8 October 2026. https://www.pillsburylaw.com/en/news-and-insights/financing-compute-legal-risks-opportunities-gpu-asset-class.html
4. Pinsent Masons (Out-Law), "Artificial intelligence: can GPU compute become a financeable asset class?", 18 September 2026 (written by Jessica Loy). https://www.pinsentmasons.com/out-law/analysis/artificial-intelligence-gpu-compute-financeable-asset-class
5. Norton Rose Fulbright, "Data centre financing: A European perspective," October 2025. https://www.nortonrosefulbright.com/en/knowledge/publications/754217e6/data-centre-financing-a-european-perspective
6. Mayer Brown, "Portability of AI Compute Infrastructure in AI Acquisitions," June 2026. https://www.mayerbrown.com/en/insights/publications/2026/06/portability-of-ai-compute-infrastructure-in-ai-acquisitions
7. Clyde & Co, Corey Greenwald, Paul Lowrie, Alice Hodgson and Ben Keatinge, "Data Centres at the Core: Navigating Insurance Challenges and Mega-Valuation," 23 March 2026. https://www.clydeco.com/en/insights/2026/03/data-centres-at-the-core-navigating-insurance-chal
8. KBRA, AI Credit Ecosystem hub (re-leasing risk report summary). https://www.kbra.com/ai-ecosystem
9. KBRA, "KBRA Releases Research – AI Credit: AI Compute Infrastructure—Cross-Sector Credit Considerations for GPU-, TPU-, and Accelerator-Intensive Assets," research press release, 27 July 2026. https://www.kbra.com/publications/RPXWSDPR (related report: https://www.kbra.com/publications/KPZcbgKd)
10. Moody's, "Credit trends in recent US data center and GPU financings," webinar agenda page, 16 July 2026, 4:00 PM ET. https://events.moodys.com/2026-mie28330-us-data-center-and-gpu-financings
11. Apollo Global Management, official FactSet corrected transcript, Q3 2026 earnings call, 3 November 2026, James Zelter (President). https://ir.apollo.com/_assets/_271fa7e209453f6962fa6c1a8cfa2ce7/apollo/db/2247/22889/webcast_transcript/APO_3Q%2726_Earnings_Call_Transcript.pdf
12. Ares Management, Q3 2026 earnings call, 29 October 2026, transcript by The Motley Fool. https://www.fool.com/earnings/call-transcripts/2026/10/29/ares-ares-q3-2026-earnings-call-transcript/
13. Bloomberg via Yahoo Finance, "Ares CEO Says It's Avoiding Backing AI Chips, Citing Complexity," 29 October 2026. https://finance.yahoo.com/technology/ai/articles/ares-ceo-says-avoiding-backing-193819449.html
14. Semafor, "AI chip financing hits a wall," 9 November 2026. https://www.semafor.com/article/11/09/2026/ai-chip-financing-hits-a-wall
15. UBS Asset Management, Greggory Price and Matthias Dettwiler, "Webs of AI debt," 29 September 2026. https://www.ubs.com/us/en/assetmanagement/insights/investment-outlook/the-red-thread/trt-end-year-2026/articles/webs-of-ai-debt.html
16. Swiss Re Institute, "sigma insights 07/2026: Insuring AI: data centre value accumulation risks," 27 March 2026. https://www.swissre.com/institute/research/sigma-research/sigma-insights-07-2026-insuring-ai-data-centre-risks.html
17. Marsh, "The market has appetite. The question is how much capacity is there?", 13 July 2026. https://www.marsh.com/en/industries/technology/insights/the-market-has-appetite.html
18. Insurance Business, "Data center buildout pushing insurers toward reinsurance and global capacity," 7 August 2026. https://www.insurancebusinessmag.com/us/news/construction/data-center-buildout-pushing-insurers-toward-reinsurance-and-global-capacity-585392.aspx
19. CoreWeave, Form 10-Q for the quarter ended 30 June 2026, Note 10. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm
20. Ornn, OCPI methodology. https://data.ornn.com/methodology
21. Silicon Data, SiliconIndex product page. https://www.silicondata.com/products/silicon-index
22. S&P Global Platts, "Platts to launch US GPU Compute Prices, Nov. 23," 5 November 2026. https://www.spglobal.com/energy/en/pricing-benchmarks/our-methodology/subscriber-notes/110526-platts-to-launch-us-gpu-compute-prices-nov-23