Three Conditions That Will Determine Whether Nvidia-Backed Debt Succeeds

PUBLIC

INDUSTRY ANALYSIS

Three Conditions That Will Determine Whether Nvidia-Backed Debt Succeeds

We tested Nvidia’s $500 billion financing platform against three conditions drawn from the 1970 Ginnie Mae precedent. The first condition now has a case that meets it. The second is not yet met. The third is being filled in by contract.

PUBLISHER

Bitplanet Research Lab

DATE

2026 · 08 · 19

CONFLICT OF INTEREST DISCLOSURE The publisher of this report is a listed company that holds bitcoin as a treasury asset and pursues businesses in digital assets and AI and data center infrastructure. Interests may bear on the choice of topic and on its interpretation.

EXECUTIVE SUMMARY

Of the three conditions, the recovery market is the least tested

WHAT HAPPENEDNvidia signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on August 10. The plan is to mobilize more than $500 billion of third-party capital into a credit market secured by Nvidia compute[1]. Two days later, CEO Jensen Huang said on the company blog that Nvidia may provide residual-value support for up to 25% of a given project[49]. That is not a guarantee of principal and interest; it covers part of the shortfall left after the collateral is sold.

WHY IT MATTERSThree tests decide whether a new collateral asset class works. First, is collateral value sufficiently separated from the borrower’s credit? Second, does a market exist in which repossessed collateral can be resold? Third, is there a party willing to provide credit enhancement until the market has been tested? In March, CoreWeave became the first financing pledging both GPUs and an investment-grade customer contract to secure an investment-grade rating (A3)[4]. A structure that loosens the link between borrower credit and collateral value has now been rated investment grade. What does not yet exist is a market with a record of repossessing and reselling this collateral, and Nvidia’s residual-value support is the device that stands in for that missing market.

WHAT TO WATCHCondition 2 is the one to track, and two indicators will show whether it is being met: actual volume in the CME compute futures due to list on October 5 and whether a benchmark reflecting secondary-market GPU sale prices emerges, and how the CoreWeave loan repays. Secondary indicators are the counterparties on the first transaction and whether the 25% residual-value support ceiling shows up in actual contract terms and disclosures[3][5].

$500B+

Platform target (third-party capital · no deadline stated · 2026-08-10)

[1]

25%

Residual-value support per loan (CEO remarks · not in filings · 2026-08-11)

[3]

A3

First investment grade for GPU and contract collateral (CoreWeave · 2026-03-31)

[4]

CONTENTS

01The compute-backed financing Nvidia and six managers are building 02Deriving three conditions from earlier precedents 03Testing the current structure against the three conditions CConclusion RSources AAppendix

01

THE PRODUCT

How Nvidia’s compute-backed financing is structured

Nvidia announced on August 10, 2026 that it had signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR[1]. The goal is to mobilize more than $500 billion of third-party capital to build AI infrastructure[6]. Definitive agreements have not been signed, and the types of borrower, pricing, site locations and the timing of first funding remain undisclosed[1][3].

In its press release, Nvidia set out a plan to turn compute and full-stack AI infrastructure into an asset class global investment capital can invest in[1]. Goldman Sachs CEO David Solomon said a new credit market could form around Nvidia compute[1]. Blackstone President Jon Gray explained that lenders will come to appraise compute infrastructure as an independent collateral asset, much as they appraise a house today[3]. The $500 billion Nvidia cited is the aggregate third-party capital the six platforms aim to mobilize over time. It does not represent Nvidia revenue, the size of any single fund, or a commitment to a particular customer[49]. The financial institutions decide whether to lend after weighing customer credit quality, demand outlook, utilization, cash flow and residual value[49][1].

The financing structure and its main components

Role Party or asset Function

Capital supply Third-party institutions such as insurers and pension funds Providers of the capital the platforms mobilize over time

Platforms Apollo · BlackRock · Blackstone · Brookfield · Goldman Sachs · KKR Each firm builds and runs its own platform and underwrites independently

Borrowers AI infrastructure operators and dedicated entities Build and operate AI infrastructure with the capital raised

Collateral Compute equipment, facilities and power, revenue from customer contracts The precise composition and weighting will be set in the definitive agreements and is not public

Support Nvidia May provide residual-value support of up to 25% on some transactions / applied case by case

Source: NVIDIA press release[1], company blog[49]

Announcements and reports on compute-backed financing came in quick succession from late July. On July 26 it was reported that Nvidia was discussing a financial backstop of roughly $250 billion tied to a 10-gigawatt (GW) campus in Pike County, Ohio[7]. By an August 14 report the amount under discussion had been cut to less than $120 billion, and the scope had narrowed from the full campus to its first phase[54]. That report cited investor concern about risk exposure as the reason for the reduction[54]. Nvidia then disclosed its residual value guaranty agreements with SB Energy in a Form 8-K on August 17. The guaranties cover leases for approximately 4.25 gigawatts (GW) of IT load under the initial commitment, with a cumulative payment cap of $105 billion. Each guaranty takes effect in sequence from 2028, when its lease commences[16]. On July 29, staff of the Division of Corporation Finance at the U.S. Securities and Exchange Commission (SEC) gave a view on a specific set of facts: a data center securitization structure in which the issuer owns the facility outright and repays principal and interest from that facility’s net operating income does not fall within the Exchange Act definition of an asset-backed security (ABS)[8]. The letter also states that this is neither a Commission rule nor a formal approval, and that the conclusion can change with the facts[8]. INTERPRETATION For issuers weighing similar structures, this reads as a signal that some regulatory uncertainty has been removed. Nvidia announced its platforms on August 10, and CME followed the next day with plans to list compute futures[1][5].

Set against total data center funding demand, the scale of Nvidia’s $500 billion target comes into sharper focus. Morgan Stanley estimated data center capital expenditure through 2028 at about $2.9 trillion. Of that, roughly $1.4 trillion would be covered by hyperscalers’ own cash flow, leaving about $1.5 trillion to be funded externally[9]. Nvidia’s $500 billion target equals about one-third of that external need (our estimate). Measured against the roughly $150 billion expected to come from securitization, it is about 3.3 times as large[9]. The two figures do not cover the same period. The Morgan Stanley estimate is cumulative through 2028, while Nvidia’s $500 billion target carries no stated deadline. Nvidia has also not disclosed the timing or pace of the fundraising[49]. The one-third ratio is therefore a simple comparison of two magnitudes and does not imply a future market share.

BlackRock Chairman Larry Fink likened the plan to the formation of the mortgage-backed securities (MBS) market he worked in during the early 1970s[3]. Recalling the last financial crisis, the analogy can be read as a warning about risk.

INTERPRETATION / AUTHOR’S INFERENCE

This report takes that analogy as its starting point for analyzing what a new collateral asset class needs in order to form, and compares the conditions visible in past cases with the compute-backed financing structure now taking shape.

02

THE PLAYBOOK

Three tests drawn from earlier precedents

The three conditions below are an analytical framework derived by sorting and comparing precedents after the fact. They are not verified necessary or sufficient conditions. We did not exhaustively check whether cases exist that met all three and still failed, or that missed one and still succeeded. Observability of collateral prices and maturity structure are not treated as separate conditions; they sit as sub-elements of Condition 2 and Condition 3 respectively.

The structures we compared fall into three broad types. First, financings secured by cash flows arising from contracts. Mortgage-backed securities and Dunkin’ royalty-backed financing belong here. Second, financings secured by the residual value of physical assets that can be resold. Aircraft and shipping containers are the standard examples. Third, vendor financing, in which a supplier lends directly to the customers buying its own products. The financing telecom equipment makers used between 1998 and 2001 is the reference case, and unlike the first two types it often involved no separate collateral at all. Comparing the three types brings out what separated success from failure.

Successes. The first MBS was issued by Ginnie Mae in 1970, pooling government-insured mortgages, with the U.S. government guaranteeing timely payment of principal and interest[10][11].

Dunkin’ Brands shows how collateral structure can lift a bond’s credit standing. In 2006 Dunkin’ raised $1.7 billion against the royalty income franchisees pay the parent[12]. The company’s bank loan rating at the time was speculative grade (B+ on the S&P scale), yet the notes backed by those royalties were rated investment grade (BBB), five notches above the bank loan rating[12][13]. The reason lay in payment priority. Royalties are paid out of franchisee sales ahead of other costs, so they can hold up reasonably well even when the parent’s own results deteriorate. Issuance continued at $2.6 billion in 2015, $1.2 billion in 2019 and $900 million in 2025, carrying the structure through 19 years[13][14][15].

Similar structures have been applied to other infrastructure assets. In telecom towers, SBA Communications and Vertical Bridge raised $2.07 billion and $1.94 billion respectively on the same day in October 2024[17]. In fiber, Ezee Fiber issued $880 million of notes against $12 million of monthly recurring revenue[18]. In data center securitization, the closest analogue to the present structure, 88 transactions totaling about $48.7 billion have priced since the market opened in 2018[19]. As of December 2024, 84% of the 42 deals S&P had rated carried an A category rating, though that distribution comes from a period dominated by cloud-lease structures[20]. Container lease securitization has likewise run for decades on collateral that is highly mobile[21].

Failures and disappointments. Nine suppliers active in telecom equipment vendor financing between 1998 and 2001 carried combined exposure of roughly $25.6 billion[22]. In a single year the impaired share jumped from 2.6% to 60% at Lucent, from 25.5% to 80% at Nortel and from 6.7% to 57% at Motorola[23]. The aircraft securitization market survived, but in 2020 Fitch downgraded 32 of the 86 tranches it rated[24][25].

Comparing the successes with the failures, three factors separated them.

Condition 1. Is collateral value separated from the borrower’s credit? Every failure above shared a tight link between the two. INTERPRETATION In the 2008 mortgage-backed securities crisis, the core problem was not that houses served as collateral. Losses widened because borrowers’ capacity to repay weakened at the same time house prices fell. The crisis clearly had more causes than this one: rating practices, securitization structure and leverage all played a part, and this analysis focuses on the correlation between collateral value and borrower credit. Vendor financing ran into the same problem, because carriers’ ability to repay and the value of telecom equipment were both driven by the same variable, the telecom capex cycle[22]. The successes, by contrast, used structures that weaken that correlation. Data center securitization secures investment-grade hyperscalers as principal tenants, which lowers the risk that tenant credit and building value move together on the same factor[20][19]. Dunkin’ achieved a comparable credit enhancement effect by giving royalty cash flows payment priority[12].

Condition 2. Is there a market in which repossessed collateral can be resold? This is what decides whether secured lending actually works. Aircraft have decades of accumulated cases of an airframe being repossessed and re-leased to another carrier[26], and container proceeds from secondhand sales also serve as a source of repayment[21]. Vendor financing, largely unsecured, left nothing to repossess when loans went bad[22]. One thing has to be true for any of this to work: the market needs an objective price benchmark for the collateral. MBS has house price indices; aircraft have established appraisal standards and practice[27]. Without a reliable price benchmark, even repossessed collateral is hard to value properly.

Condition 3. Is there a party to provide credit enhancement until the first two conditions have been tested? Early investors in a new asset class have to carry risk that has not been tested. Two devices have covered that early uncertainty. One is ‘anchor credit’ from a government or a highly rated corporate. In the first MBS, the government guarantee served as the credit foundation until the market could price the risk on its own[10][11]. The Dunkin’ notes cleared the company’s bank loan rating by five notches for the same reason: priority over the royalty cash flows worked as credit enhancement[12][13]. The other device is maturity structure. Container securitizations set legal final maturities at 20 to 25 years, keeping them inside the economic life of the collateral[21], while aircraft appraisals apply a downturn scenario from the outset and discount residual values by a further 40% to 75%[27]. A guarantee shares loss risk with a third party; maturity structure shortens the period of exposure itself. The approaches differ, but both cushion early uncertainty.

▲ A note on sources and how to read them

The success cases draw mainly on materials from issuers’ counsel, rating agencies and federal agencies, while the failure cases rely more heavily on after-the-fact reporting and filings[28][15][19][22][23]. The two sets of material were therefore written at different times and for different purposes.

03

THE SCORECARD

Scoring the three conditions: the recovery market remains untested

We assessed the current structure against the three conditions derived in the previous chapter, on a three-step scale of ‘met’, ‘partially met’ and ‘not met’. A condition is ‘met’ where the structure exists and has a record of working in practice, ‘partially met’ where the structure or contract terms exist but working practice has not been observed, and ‘not met’ where neither is evident.

Two principles governed the scoring. Contract terms alone do not establish that a market works in practice, and futures that are due to list are not treated as having proven price discovery until trading actually begins.

On this basis Conditions 1 and 3 are ‘partially met’ and Condition 2 is ‘not met’. Condition 3 rests on the thinnest evidence of the two partial scores, because neither the contract language nor the loss attachment order has been made public.

Condition Score Basis What would prompt a rescore

1. Is collateral value separated from the borrower’s credit Partially met CoreWeave secured an A3 rating on a financing collateralized by both GPUs and an investment-grade customer contract. This is one private loan, and whether the structure repeats has not been established [4]. Disclosure of the counterparties, collateral mix and rating outcome on the platforms’ first transaction

2. Is there a market in which repossessed collateral can be resold Not met No actual repossession of GPU collateral has been observed. In the comparable bitcoin mining rig cases, collateral value fell short of the loan balance [53]. CME compute futures are due to list in October and reference rental rates rather than secondary-market sale prices [5]. Traded volume and open interest in compute futures, emergence of a benchmark for secondary-market GPU sale prices, and an observed case of the third-party takeover and operation clause being exercised

3. Is there a party to provide credit enhancement until the first two conditions have been tested Partially met Nvidia has signalled possible residual-value support of up to 25%. It is discretionary and applied case by case rather than a contractual obligation, and the specific contract language and loss attachment order are undisclosed. The scope is also narrower than Meta’s 16-year full guarantee and Broadcom’s commitment to cover 100% of the senior shortfall [49][50][51]. Confirmation of the specific contract language, loss attachment order and whether the support appears in filings

Condition 1: partially met. The first financing to secure an investment-grade rating is now on the record. Many borrowers in this market are unrated AI operators[3], and both their capacity to repay and the value of their GPUs depend heavily on demand for AI compute. If that demand falls, borrower revenue and collateral value can decline together. Facility-backed securitization lowered that risk by adding contracted cash flows from highly rated tenants. In compute equipment financing, anchor tenant participation, overcollateralization and a sufficiently deep subordinate tranche can play a similar credit enhancement role.

CoreWeave’s A3 rating rests on a structure that pledged GPUs together with the contracted revenue of a large investment-grade customer[4][29]. Rather than fully separating borrower credit risk from collateral value, it supplements the collateral with a strong customer’s credit. That customer’s own credit standing is not entirely independent of AI investment flows either. The diversity of lenders that Zelter pointed to may also mitigate risk[3]. But six institutions lending against the same asset class does not by itself lower the correlation among collateral values. Taking this case into account, Condition 1 can be raised from ‘not met’ to ‘partially met’. The evidence is still a single private loan, so whether the same structure repeats in the platforms’ actual transactions has yet to be seen.

Condition 2: not met. Residual-value support cannot substitute for a recovery market. No GPU-secured loan has yet produced an actual repossession, because no borrower default has led a creditor to seize the collateral. There is therefore too little evidence to judge whether repossession and disposal procedures work smoothly in real conditions.

CoreWeave could nonetheless become the first case to test the recovery structure. The $7.5 billion loan facility secured by GPUs and customer contracts began repaying in January 2026[30]. Over the same period the market value of the H100s in the collateral pool was being sharply repriced[31][32], though sources on the H100 price trend conflict. CoreWeave then completed a refinancing on March 31 through the $8.5 billion DDTL 4.0 facility[4][33]. It was the first financing secured by high-performance compute infrastructure together with a customer contract to receive an A3 from Moody’s and an A (low) from DBRS.[4] Pricing came at SOFR+2.25% on the floating rate tranche and about 5.9% on the fixed rate tranche, roughly half the approximately 11% average on the 2024 loan facility[30][29]. This was a private loan rather than a public securitization, and the collateral includes the contracted revenue of a large customer alongside the GPUs[29]. Refinancing successfully without a default tells us nothing about whether the repossession machinery works.

Bitcoin mining rigs, compute equipment of a similar character to GPUs, have already produced real repossessions. Equipment-secured lending to miners once reached roughly $4 billion, and rig prices collapsed in 2022. Counting only publicly known cases, defaults totalled about $227 million to $238 million[53]. In some of the impaired deals, the pledged rigs went to the lenders. Stronghold returned about 26,200 rigs to extinguish $67 million of debt, and Core Scientific handed over 27,403 rigs to extinguish $38.6 million[53].

Three things stand out from those cases. First, collateral value can fall short of the loan balance even when every unit is recovered. The rigs Core Scientific delivered were appraised at $25 million against a loan balance of $38.6 million[53]. Prices for new rigs fell about 85% over the same year[53]. Second, a lender may be unable to sell the equipment it recovers and end up operating it. NYDIG, Galaxy and Foundry all ran the machines they took back rather than leave them idle, and Foundry acquired the associated facilities through a bankruptcy auction[53]. Third, bulk sales bring additional price pressure. Large disposals frequently cleared at discounts of 10% to 25% to fair value[53].

GPUs and bitcoin mining rigs differ in how broadly they can be used. A mining rig serves one purpose; a GPU can run many kinds of compute work. That is why Nvidia emphasizes the fungibility and redeployability of its GPUs[49]. Even so, the mining rig cases show substantial value lost in the actual process of recovering compute equipment collateral.

Nvidia’s residual-value support is a device for a world in which the recovery market has not yet formed. It covers part of the shortfall when disposal proceeds fall short. INTERPRETATION The support can be read as having been designed on the premise that liquidity and price discovery in the recovery market are not yet proven. Had that liquidity been ample, the need for such support would have been correspondingly lower. Residual-value support therefore does not by itself satisfy Condition 2, because it does not create a resale market for the collateral; it shifts part of the disposal loss onto the supplier. A record of repeated real trades and sufficient market liquidity can only be established once actual transactions accumulate.

CoreWeave financing terms compared

Item May 2024 loan facility March 2026 DDTL 4.0

Size $7.5 billion $7.5 billion initially / up to $8.5 billion once stabilization conditions are met

Pricing About 11% on average SOFR+2.25% floating / approximately 5.9% fixed

Rating Unrated Moody’s A3 / DBRS A (low)

Collateral GPUs and related assets GPU clusters and revenue from customer contracts

Structure Non-recourse Non-recourse / assets and liabilities ring-fenced in a dedicated entity

The effective interest rate reported in the quarterly filing for the period ended March 2026 is calculated separately; the rates shown here are the contractual rates at the time of funding. Sources: CoreWeave IR[4], 8-K[33], compiled DDTL 4.0 terms[29].

Nvidia is also putting in place mechanisms to reduce uncertainty in the recovery process. Borrowers must adopt the system architecture Nvidia specifies, and the design allows a third party to take over and operate the equipment if problems arise[3]. A contract can establish a repossession procedure; it cannot supply a record of repossessions actually happening. Aircraft-secured financing earned market confidence not because leases contain repossession clauses but because those clauses have been exercised repeatedly in real defaults[26]. Nvidia’s structure is at the stage of laying the contractual groundwork on which such a record could later be built.

A recovery market also needs an objective basis for valuing collateral. For facility assets, that basis already exists. S&P published rating criteria for data center securitizations in June 2024[34], and Moody’s dedicated methodology of February 2025 sorts assets into five groups and applies revenue volatility of 5% to 23% by group[35]. Both sets of criteria address data center facilities. Fitch has treated AI training facilities less favourably than general cloud facilities[20].

The market for facility collateral is expanding quickly. The first deal was Vantage’s $900 million securitization in February 2018, and through 2020 there were only three or four issuers[36]. Issuance broadened from 2021, reaching about $25 billion in 2025 with $40 billion to $70 billion expected in 2026[36]. That growth belongs to the facility collateral market. Valuing equipment such as GPUs objectively is a separate problem.

CME Group and Silicon Data announced on August 11 that they would list two compute futures contracts on October 5, subject to regulatory review[37][5]. The contracts are the H100 Rental Index Futures and the B200 Rental Index Futures, listed under NYMEX rules and cash-settled[5]. Silicon Data, which publishes the underlying indexes, is backed by the trading firm DRW[5].

Three limits stand between these contracts and a working benchmark for GPU recovery value. First, listing a future does not automatically bring adequate volume and liquidity. Second, the underlying index reflects GPU rental rates and does not directly measure secondary-market sale prices. Third, if collateral pools later extend to other GPU generations, what the index measures and what actually sits in the pool may diverge. The residual value assumptions used in securitization depend on what can actually be recovered on disposal. In the first half of 2026, residual values moved on separate drivers even while rental rates were rising[38][32]. Residual value estimates themselves differ by source. Some put the decline at $30,000 down to $8,000[31], while others show a subsequent rebound[32]. What the sources agree on is that prices can be repriced sharply over short periods rather than eroding at a steady, gentle pace[39]. Silicon Data, which reported the rebound, is also the index provider for the CME contracts.

Condition 3: partially met. Credit enhancement exists, but its scope is uncertain. Nvidia has offered residual-value support as its external credit enhancement[49]. Applying 25% to the $500 billion platform target gives $125 billion, but that is not an amount Nvidia has committed; it is a simple conversion applying the support ratio to the entire target[2][40]. Whether support is provided, and at what size, is decided project by project, and which projects will receive how much has not been disclosed[49].

To gauge capacity to bear that support, Nvidia reported fiscal 2026 revenue of $215.9 billion and net income of $117.0 billion, with year-end cash and marketable securities of $62.6 billion[41]. The shape of its existing guarantees is partly visible in filings. When first disclosed in the third quarter of fiscal 2026, maximum guarantee exposure was $860 million with $470 million held in escrow[42]. Cumulative maximum exposure on the facility lease guarantees disclosed since then is $3.5 billion. That exposure declines as partners make payments to lessors, and $712 million sits in escrow. The guarantees run five to seven years, and Nvidia assessed their fair value as not material to its financial statements[43].

Because the actual size of the residual-value support, the conditions that trigger it and the total exposure are all unsettled, the information now public does not allow a judgment on how much Nvidia could absorb. The $125 billion figure above is a simple conversion rather than a commitment, so setting it against $62.6 billion of cash and marketable securities is not an appropriate comparison either. The guaranty structure for the Ohio project was set out in the disclosure of August 17. The guaranties as filed cover leases for the initial 4.25 GW, with a cumulative payment cap of $105 billion[7][54][16]. That runs above the $62.6 billion of cash and marketable securities at year-end, but each guaranty takes effect in sequence from 2028 and the tenant reimburses Nvidia for amounts actually paid, so the payment cap and Nvidia’s net exposure need to be read separately[16].

Nvidia’s residual-value support can be judged on three dimensions: whether it is mandatory, how far it reaches, and how losses are shared.

First, it is not a contractual obligation. The government guarantee on the first mortgage-backed securities was a legal duty to ensure payment of principal and interest, and investors held the right to claim payment under it[10][11]. Nvidia has said only that residual-value support is possible; whether it applies is decided through case-by-case review[49]. Even as the need for support grows, in other words, providing it remains Nvidia’s call.

Second, the scope is limited compared with other deals struck around the same time. Meta guaranteed the residual value of the entity it established with Blue Owl for the first 16 years of operation. If the lease is not renewed or terminates early, Meta covers any shortfall left after asset sale proceeds fail to repay principal and interest in full. On that structure, $27.294 billion of notes were rated A+ by S&P and priced at 225bp over Treasuries[50]. Broadcom agreed to cover the entire shortfall for senior investors on the platform it built with Apollo and Blackstone. If Anthropic fails to meet its lease payment obligations, the special purpose vehicle disposes of the chips, and Broadcom pays senior investors any shortfall the proceeds do not cover[51][52]. The scope Nvidia has described is considerably narrower[49], and losses beyond it fall to investors.

Third, the specific contract language and the loss attachment order have not been made public. Which loss layer the support applies to, where it sits between senior and subordinate positions, and what conditions trigger it are all undisclosed. None of this appears in the platform press release or in Nvidia’s periodic SEC filings; it was published only through the company blog[49].

The nature of the credit enhancement provider differs too. Ginnie Mae is a government agency; Nvidia is a private company. Its capacity to enhance credit is finite, and its own credit standing is itself linked in part to AI demand. The more the support is drawn on, the more Nvidia’s own risk exposure grows, and the build-up of off-balance-sheet leverage flagged by the U.S. Senate relates to exactly this kind of risk transfer[45]. In having a supplier enhance the debt of customers buying its own products, the structure also resembles telecom equipment vendor financing of the late 1990s[22]. Back then lending ran to as much as 130% of equipment value with no separate collateral, whereas the present structure caps residual-value support at 25% and includes collateral together with third-party takeover and operation terms[49]. Whether those differences in safeguards translate into different loss rates and different market stability has yet to be tested.

Maturity structure matters alongside credit enhancement. Estimates of the economic life of a GPU diverge between two to three years[46] and four to six years[47][1]. In 2025, Amazon and Meta adjusted the useful lives of similar GPU assets in different directions[46]. Using an asset of uncertain economic life as collateral for long-dated debt has its limits. That risk can be reduced by adjusting the maturity structure without changing the collateral itself: setting bond maturities inside the economic life of the asset, as container securitizations do[21]. Shorter expected repayment periods, faster principal amortization and ample overcollateralization can all serve. Aircraft appraisals apply the same logic when they build in a downturn scenario from the start and discount residual values by a further 40% to 75%[27].

C

CONCLUSION

Of the three conditions, the recovery market is the least tested

The success cases reviewed above share three features. These three conditions are less an arbitrary standard than a set of common characteristics observed while comparing those precedents and then organized into an analytical framework. Conditions 1 and 3 currently score as ‘partially met’. The evidence behind Condition 1 is still a single private loan. Condition 3 has residual-value support of up to 25%[49], but it is not a contractual obligation and neither the specific contract language nor the loss attachment order has been disclosed. Against Meta’s 16-year full guarantee and Broadcom’s commitment to cover 100% of the senior shortfall, both struck around the same time, the scope is also relatively narrow[50][51]. Of the three, only Condition 2, the market for repossessing and reselling collateral, scores ‘not met’.

Three things drive that score. First, there are no actual repossessions. No case has yet been observed in which a borrower default on a GPU-secured loan led a creditor to repossess and dispose of the collateral. Aircraft finance has decades of accumulated repossession and re-lease cases; GPU-secured financing has none[26]. CoreWeave’s successful refinancing is not evidence that the recovery structure works. No default occurred, so nothing about the repossession machinery was tested.

Second, a price benchmark is forming but it measures something else. The underlying index for the futures due to list in October reflects GPU rental rates, while repossession calls for secondary-market sale prices[5].

Third, a recovery market cannot be built out of contract terms. Condition 1 can be partly addressed by adding the contracted cash flows of strong customers to the collateral, and Condition 3 by adding credit enhancement clauses. Nvidia can likewise design a contract structure allowing a third party to take over and operate the equipment if a borrower fails to meet its obligations[3]. But the existence of such a clause and the formation of an actual recovery market are separate matters. Only when repossession and resale genuinely occur after a default can their effectiveness be confirmed. GPU-secured financing has no such case yet. Mining rig financing, similar in character, did produce real repossessions, and in at least one instance appraised value fell short of the loan balance even after every unit was recovered[53].

INTERPRETATION Nvidia’s residual-value support cannot on its own resolve the absence of a recovery market. Forming that market and covering losses during the period before it forms are two different functions.

The largest uncertainty among the three conditions sits with Condition 2. It cannot be addressed quickly through capital or contract structure alone. Conditions 1 and 3 respectively loosen the link between borrower credit and collateral value and cushion early loss risk, but neither substitutes for the recovery market itself. The more that market develops, the less the structure needs additional credit enhancement. If it fails to develop, the burden on the credit enhancement provider persists.

INTERPRETATION / AUTHOR’S INFERENCE

Whether compute financing establishes itself as an independent collateral asset class appears likely to hinge on how Condition 2 resolves. If the recovery market does not form, the structure may remain one that depends heavily on Nvidia’s credit enhancement.

Meeting Condition 2 would not guarantee the market’s success either. The evidence behind Condition 1 is still a single case, and structurally similar vendor financing has its own record of failure[22][23]. Regulators are also watching the growth of the associated debt and off-balance-sheet exposure[48][45].

The risk of leaving Condition 2 untested is equally clear. Until repossession cases accumulate, the disposal value of the collateral remains an unverified assumption.

INTERPRETATION / AUTHOR’S INFERENCE

Where loss-absorbing devices such as residual-value support and subordinate capital prove insufficient, losses can pass through to end investors according to the loss attachment order set in the contracts. Because the tranche structure and that order are undisclosed, the loss range investors would actually bear cannot be assessed at present.

Issuers have different structural options for responding to this uncertainty. They can shore up Condition 1 by including the contracted cash flows of strong customers in the collateral, lower early risk by shortening maturities and pulling principal amortization forward, or raise capital chiefly in the private market until the recovery market has been tested.

The indicators that would change this assessment can be specified. Three things would raise the score on Condition 2: meaningful volume and open interest building after the October futures listing; the emergence of a benchmark reflecting secondary-market GPU sale prices; and the third-party takeover and operation clause actually being exercised in a GPU-secured loan default. Conversely, the current assessment would need reconsidering if the platforms’ first transaction is built around facility assets rather than GPUs, or if it turns out that the required ratings cannot be achieved without residual-value support. Condition 1 will be rescored once the counterparties, collateral mix and rating analysis on the first platform transaction are disclosed. Condition 3 will be rescored on the specific contract language for residual-value support, the loss attachment order and whether the support appears in filings.

▍ KEY CONCLUSION

Two indicators will determine whether Condition 2 is met. First, whether adequate volume develops after the October compute futures listing and whether a benchmark reflecting secondary-market GPU sale prices emerges[5]. Second, how the CoreWeave loan repays and whether the third-party takeover and operation clause works in any future borrower default[30]. As secondary indicators, Condition 1 can be tracked through the counterparties and collateral mix on the first platform transaction, and Condition 3 through the specific contract language for residual-value support and the loss attachment order.

How quickly Condition 2 is tested will be the central measure of whether compute-backed financing settles into an independent asset class or stays in a risk structure resembling past vendor financing. Where deals are done privately and terms stay confidential, some of these indicators will be hard to observe directly. In that case, rating agency pre-sale reports, revisions to rating methodologies and trustee reports can serve as secondary indicators.

R

REFERENCES

Sources

# Title · URL Publisher Date

[1] NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital Primary https://nvidianews.nvidia.com NVIDIA Newsroom 2026-08-10

[2] Nvidia partners with Wall Street giants to raise $500 billion for AI buildout (Reuters wire, syndicated version) Business Recorder 2026-08-11

[3] Wall Street just endorsed Jensen Huang's 'big concept' for AI. What now? CNBC 2026-08-11

[4] CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment-Grade Rated GPU-backed Financing Primary https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx CoreWeave IR 2026-03-31

[5] CME Group and Silicon Data joint press release on the compute futures listing Primary PR Newswire 2026-08-11

[6] Nvidia builds KRW 710 trillion AI infrastructure financing platforms with six Wall Street firms (roundup) (엔비디아, 월가 6개사와 710조원 AI인프라 금융플랫폼 구축(종합)) Yonhap News (연합뉴스) 2026-08-11

[7] Nvidia and OpenAI in talks for up to $250 billion backstop (confirming WSJ reporting of 2026-07-26) CNBC 2026-07-27

[8] Latham Letter Delivers Regulatory Clarity for Data Center Securitizations Latham & Watkins 2026-08

[9] How Data Center ABS and CMBS Fit in a Broader Financing Ecosystem (restating Barclays and Morgan Stanley data) https://structuredfinance.org Structured Finance Association 2026-07-23

[10] Ginnie Mae Multiclass Securities Program Primary https://www.hud.gov/hudprograms/Ginnie_Mae_MSP HUD 2018

[11] Michael Bright (Ginnie Mae EVP and COO), written testimony to the House Financial Services Committee Primary https://financialservices.house.gov/uploadedfiles/hhrg-115-ba04-wstate-mbright-20171129.pdf U.S. House Financial Services Committee 2017-11-29

[12] Dunkin' Brands Returns to Securitization Market Asset Securitization Report 2015-01-28

[13] Whole Business Securitization: Cash Flow Structure & Valuation Corvid Partners 2024

[14] Dunkin' takes third helping of whole-biz ABS in $1.2B deal Asset Securitization Report 2019-03

[15] Dunkin' 2025-1 $900 million whole business securitization offering (counsel to the issuer) Davis Polk 2025-11

[16] NVIDIA Form 8-K (nvda-20260817), residual value guaranties Primary https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm SEC EDGAR 2026-08-17

[17] Vertical Bridge secures $1.94B in tower financing Wireless Estimator 2026-02

[18] A fiber network company makes $880 million ABS debut Asset Securitization Report 2026-07

[19] Data Centers: A Comparison of ABS and CMBS Structures (rating agency on the transaction) Primary https://www.kbra.com/publications/wtXJZSZt KBRA 2025-05-30

[20] Data center ABS: Risks, yields, and ratings https://media.crai.com/wp-content/uploads/2025/12/03163400/Insights-Data-center-ABS-%E2%80%93-Risks-yields-and-ratings-December2025.pdf Charles River Associates 2025-12

[21] Quick notes on container securitizations https://portfolio-strategy.apsec.com APSEC Portfolio Strategy 2023-01

[22] Vendor Financing Loops: What 1999 Telecom Tells Us About 2026 AI Michel Johannsen 2026-03-15

[23] The Stupid Loan Bubble (citing academic estimates built from SEC filings) Newsweek 2001

[24] Fitch Continues Downgrade Actions Across Aviation ABS Deals Asset Securitization Report 2020-12

[25] Fitch notes four ABS 'double downgrades' Airfinance Global 2020-09-10

[26] Alternative and Emerging Asset Class Spotlight: Aircraft ABS Structured Finance Association 2020-03

[27] Credit FAQ: Rating Considerations For New Aircraft ABS Securitizations Primary S&P Global Ratings 2021-03-16

[28] Securitising the Company: Inside Whole Business Securitisation (contributed by counsel to issuers) ICLG 2026-05-18

[29] DDTL 4.0 terms compiled (maturity 2032-03, SOFR+2.25% floating and approximately 5.9% fixed, non-recourse, collateral = GPU clusters plus contracted revenue) StockTitan · Global Data Center Hub 2026-03~06

[30] Emerging Litigation Risks in Financing AI Data Centers Boom Quinn Emanuel 2026-03

[31] Nvidia Signs $500B Wall Street GPU Bond Deal; Rating Standard Unresolved TechTimes 2026-08-11

[32] GPU Residual Value Report: 2026 Outlook American Compute 2026-06-15

[33] CoreWeave Form 8-K Ex-99.1 Primary https://www.sec.gov/Archives/edgar/data/1769628/000176962826000129/ex991.htm SEC EDGAR 2026-03

[34] Data Center Securitizations: Global Methodology And Assumptions Primary S&P Global Ratings 2024-06-13

[35] Data Center e-Primer (includes a description of the Moody's methodology) CREFC 2026-02

[36] The infrastructure revolution: understanding data center securitization https://www.rbccm.com/en/insights/2025/12/the-infrastructure-revolution-understanding-data-center-securitisation RBC Capital Markets 2025-12

[37] CME Group Says It Will Launch First Regulated Compute Futures Contracts Finance Magnates 2026-08-11

[38] GPU Tech Refresh (citing the SemiAnalysis index) American Compute 2026-04-02

[39] The Illusion of Stability: Unpacking H100 GPU Market Value Trends (provider of the underlying index for the CME futures) Silicon Data 2026-01-30

[40] Nvidia, finance firms to raise $500bn for AI build-out RTE (Irish public broadcaster) 2026-08-11

[41] NVIDIA FY2026 results (revenue $215.9 billion, net income $117.0 billion, cash and securities $62.6 billion; secondary media source restating the 10-K figures) Tom's Hardware 2026-07

[42] NVIDIA Form 10-Q FY2026 Q3 (nvda-20251026), guarantee note Primary https://www.sec.gov/Archives/edgar/data/1045810/000104581025000230/nvda-20251026.htm SEC EDGAR 2025-11

[43] NVIDIA Form 10-Q FY2027 Q1 (nvda-20260426), guarantee note Primary https://www.sec.gov/Archives/edgar/data/1045810/000104581026000052/nvda-20260426.htm SEC EDGAR 2026-05

[45] Warren, Colleagues Press FSOC to Launch Probe into Financial Stability Risks of AI Debt Bubble Primary U.S. Senate Banking Committee minority 2026-01-22

[46] Deep Quarry: Useful Lives of GPUs National Law Review 2026

[47] The question everyone in AI is asking: How long before a GPU depreciates? CNBC 2025-11-14

[48] Financial Stability Report July 2026 Primary Bank of England 2026-07

[49] NVIDIA AI Factory Compute Is Becoming an Investable Asset Class Primary https://blogs.nvidia.com/blog/nvidia-ai-factory-compute/ NVIDIA official blog 2026-08-12

[50] Financing Package: $27.29 billion of notes from Blue Owl Capital / Beignet Investor (issuance terms and the 16-year residual value guarantee) IFR 2025-12

[51] Broadcom, Apollo, and Blackstone Establish AI XPV Platform Primary Broadcom · PR Newswire 2026-06-09

[52] AI XPV tranche structure and the terms covering 100% of the senior shortfall (secondary citation based on Bloomberg reporting; primary report not obtained) TradingKey · Capacity 2026-06

[53] Unsustainable Bitcoin Mining Debt Led to Record Defaults, Buyouts in 2022 https://hashrateindex.com/blog/untenable-bitcoin-mining-debt-led-to-defaults-acquisition-in-2022/ Hashrate Index 2023-01

[54] Nvidia scales back funding guarantee for Ohio OpenAI data center (citing WSJ reporting) Reuters 2026-08-14

A

APPENDIX

Data limitations and the basis for our estimates

① Limitations

Limited disclosure of contract structure: because the definitive agreements for the six platforms are not public, we cannot confirm the precise definition of collateral, the tranche structure or the loss attachment order for residual-value support. Conditions 1 and 3 are therefore provisional scores reflecting the limits of what has been disclosed.

Extent of confirmation on residual-value support: the 25% ceiling on residual-value support was cross-checked against the CEO's remarks and independent reporting that cited them, but it does not appear in the company's formal filings.

Limits of after-the-fact classification: the three conditions are an analytical framework derived by sorting and comparing past cases retrospectively. We did not exhaustively check whether cases exist that met all three and still failed, or that missed one and still succeeded. In the course of the analysis we consolidated five elements into three conditions, classifying price observability and maturity structure as sub-elements of the related higher-level conditions. Nor did we exhaustively check whether cases exist in which price observability and maturity structure operated as independent variables.

Limits of the analogy: in comparing the Ginnie Mae precedent with Nvidia, the body of the report states the difference between a government agency's mandatory guarantee and a private company's discretionary support. The comparison with the 2008 housing market and the discussion of the downturn phase in the container market are classified as general interpretation without individual sources, and no specific figures were used.

Character of the sources: the success cases draw mainly on materials from law firms, rating agencies and federal agencies, while the failure cases rely comparatively more on after-the-fact reporting and filings.

Differences between data compilers: annual issuance figures are not presented as a single continuous series because compilers differ in what they include. SFA forecast $8 billion of issuance for 2025 in May of that year, while RBC counted about $25 billion for the year in December 2025. The two may differ in geographic and structural coverage, so the body uses only the RBC figure, whose measurement date can be stated. We could not obtain consistent data on default rates and rating downgrades in data center securitization. The 84% share of A category ratings is as of December 2024 and reflects a period dominated by cloud-lease transactions.

Conflicting sources: estimates of GPU residual value and economic life differ by source, so the conflicting figures are presented side by side in the body. Some of the institutions producing residual value data are the same as the provider of the underlying index for the CME compute futures.

Details drawn from secondary sources: the pricing, maturity and collateral composition of DDTL 4.0 were compiled from several secondary sources. The ratings, deal size and first-investment-grade status were confirmed against the company's IR release and the 8-K. Nvidia's fiscal 2026 revenue, net income and year-end cash and marketable securities were cross-checked through secondary sources citing the 10-K.

② Formulas for Our Own Estimates

For every figure marked ‘our estimate’ in the body, we set out the formula and the basis of calculation. Where sources use different methods, or where the specific formula cannot be verified, that limitation is stated alongside.

Measure Formula (numerator / denominator) As of Price and FX source

Residual-value support of about $125 billion $500 billion × 25% (per-loan ratio applied to the whole target) 2026-08-11 Based on remarks [2][3]; formula confirmed as stated alongside in reporting

Target = about one-third of external funding needs $500 billion ÷ $1.5 trillion (the denominator is a cumulative estimate through 2028 while the numerator is a target with no stated deadline, so the periods do not match) Target 2026-08-10 Funding need estimated through 2028 [1][9]

Target = about 3.3 times the expected securitization share $500 billion ÷ $150 billion (the denominator is an estimate of the share securitization ultimately absorbs; the same period mismatch applies) Target 2026-08-10 Estimate [9]

Scores by condition (Condition 1 partial, Condition 2 not met, Condition 3 partial) The condition-by-condition scores in Chapter 3 (criteria: met = structure plus a working record, partial = structure or contract terms only, not met = neither) 2026-08-19 Our own assessment, against the three conditions in Chapter 2

③ Methodology

Source priority: official company and regulator materials, then primary data and compiler databases, then specialist media, then syndicated media, then community sources.

Source classification: a source is classified as primary only where we read the original document. Where we could not obtain the original, we mark the citation as secondary and state that the primary report was not obtained.

As-of and retrieval dates: for values fixed at a point in time we state the as-of date; for dashboard figures that move in real time we state the retrieval date.

Handling conflicting figures: where figures differ, we present both. Where a reference value must be chosen, we weigh disclosure of the calculation scope, clarity of the retrieval point and whether components can be separated ahead of a publication's prominence, and we state the basis of the choice in the body.

Our own estimates: each of our own estimates is presented individually in Appendix ② with its formula, as-of date and limitations.

Legal and regulatory interpretation: regulator materials are cited only to explain the classification of general structures and their potential application, and are not extended into an official classification or approval of any specific product. Descriptions of product structure and regulatory interpretation are kept in separate paragraphs.

Conflict of interest disclosure: every report carries the same disclosure wording in a fixed position at the top of the document.

PUBLISHED BY Bitplanet Research Lab

WRITTEN BY Taewon Kim

REVIEWED BY Sooyoung Kim

DISCLAIMER This material is industry analysis prepared for information purposes and is not investment advice. It does not recommend buying or selling any security and offers no price forecasts or target prices. The author is neither a lawyer nor a licensed investment adviser, and all figures are drawn from public sources (as of the date of writing, and subject to change thereafter).