Three Conditions That Will Determine the Success of NVIDIA's Secured Bonds – Bitplanet

[Block Media Bitplanet] NVIDIA's $500 billion financial platform has been examined in light of the 1970 Ginnie Mae case.

EXECUTIVE SUMMARY

The Recovery Market Verification is the Most Lacking Among the Three Conditions

Key Points NVIDIA signed a memorandum of understanding on August 10 with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The plan is to attract over $500 billion in third-party capital for a credit market secured by NVIDIA Compute [1]. Two days later, CEO Jensen Huang stated on the company blog that they could support residual value up to 25% on a project-by-project basis [49]. This is not a guarantee of principal and interest but a mechanism to cover shortfalls arising after the disposal of collateral.

Core Significance The key criteria for assessing the success of this new class of secured assets are threefold. First, is the collateral value sufficiently separated from the borrower's credit? Second, is there a market for reselling the recovered collateral? Third, is there an entity willing to provide credit enhancement until the market is sufficiently validated? CoreWeave secured an investment grade (A3) for the first time with a financial structure that offered GPU and investment-grade customer contracts as collateral [4]. This means that a structure that reduces the linkage between the borrower's credit and the collateral value has been recognized with an investment grade. However, a market with accumulated performance from the recovery and resale of collateral has not yet formed, and NVIDIA's residual value support is a mechanism prepared for a situation where the recovery market is not sufficiently established.

Key Verification Indicators The verification of Condition 2 is crucial and can be confirmed through two indicators. The actual trading volume of CME Compute futures, which are set to be listed on October 5, and the formation of indicators reflecting the resale prices of used GPUs, as well as the repayment progress of CoreWeave loans. A secondary indicator is whether the composition of the counterparties in the first transaction and the maximum 25% residual value support are reflected in the actual contract terms and disclosures [3][5].

$500B+ Target Amount for the Platform (Third-party Capital; Deadline Not Specified; 2026-08-10)
25% Individual Loan Residual Value Support (CEO Statement; Not Reflected in Disclosure; 2026-08-11)
A3 First Investment Grade for GPU and Contract Secured Financing (CoreWeave; 2026-03-31)

CONTENTS

• 01 Compute Secured Financing Promoted by NVIDIA and Six Asset Managers
• 02 Derivation of Three Conditions Through Precedent Comparison
• 03 Current Structure Check Based on Three Conditions
• C Conclusion
• R References
• A Appendix

01 THE PRODUCT

Structure of Compute Secured Financing Promoted by NVIDIA

NVIDIA announced that it signed a memorandum of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR on August 10, 2026 [1]. The goal is to attract over $500 billion in third-party capital for AI infrastructure development [6]. The final contracts have not yet been signed, and details such as borrower types, interest rates, facility locations, and funding start times have not been disclosed [1][3].

In a press release, NVIDIA proposed making compute and full-stack AI infrastructure an asset class that global investors can invest in [1]. Goldman Sachs Chairman David Solomon assessed that a new credit market could be formed based on NVIDIA Compute [1]. Blackstone's John Gray explained that in the future, lenders will evaluate compute infrastructure as independent collateral assets, just as they do with housing [3]. The $500 billion mentioned by NVIDIA is the total target size of third-party capital they aim to attract across six platforms in the long term. This does not refer to NVIDIA's revenue or the size of a single fund, nor does it indicate specific amounts for particular clients [49]. Financial institutions decide on lending based on a comprehensive review of the client's creditworthiness, demand outlook, utilization rates, cash flow, and residual value [49][1].

Source: NVIDIA Press Release [1], Company Blog [49]

Since the end of July, there have been consecutive announcements and reports related to compute secured financing. On July 26, reports emerged that NVIDIA was discussing providing financial guarantees of about $250 billion related to a 10-gigawatt (GW) campus in Pike County, Ohio [7]. Subsequently, on August 14, reports indicated that the scale of negotiations had been reduced to less than $120 billion, and the scope was limited to Phase 1 of the campus rather than the entire campus [54]. The reports mentioned that concerns about investors' risk exposure were cited as the background for the reduction in the scale of negotiations [54]. On August 17, NVIDIA disclosed a residual value guarantee contract with SB Energy through an 8-K filing. The guarantee covers an initial IT load lease of approximately 4.25 GW, with a cumulative payment limit of $105 billion. Individual guarantees will take effect sequentially starting from 2028, when the lease commences [16]. On July 29, staff from the U.S. Securities and Exchange Commission (SEC) Corporate Finance Division provided opinions on specific factual circumstances. It stated that a data center securitization structure that directly owns the facility and repays principal and interest with the facility's net operating income does not fall under the definition of asset-backed securities (ABS) under the Securities Exchange Act [8]. However, this is not a rule or official approval from the commission, and it is also noted that judgments may vary based on factual circumstances [8]. [Interpretation] This can be interpreted as a signal that lowers regulatory uncertainty for issuers considering similar structures. Following NVIDIA's platform announcement on August 10, CME also announced plans to list compute futures the next day [1][5].

NVIDIA's $500 billion target becomes even clearer when compared to the overall data center funding demand. Morgan Stanley estimates that capital expenditures needed for data centers will be approximately $2.9 trillion by 2028. Of this, about $1.4 trillion will be covered by hyper-scalers' own cash flow, while the remaining approximately $1.5 trillion is expected to be raised externally [9]. NVIDIA's target of $500 billion corresponds to about one-third of this external funding need (self-calculated). Compared to the approximately $150 billion expected to be raised through securitization, it is about 3.3 times larger [9]. However, the two figures are not based on the same time frame. Morgan Stanley's estimate is a cumulative amount up to 2028, while NVIDIA's $500 billion target does not specify a deadline for achievement. NVIDIA has also not disclosed specific timing and pace for fundraising [49]. Therefore, the ratio of about one-third is a value obtained from a simple comparison of the two scales and does not imply future market share.

Larry Fink, the chairman of BlackRock, likened the current situation to the formation of the mortgage-backed securities (MBS) market he experienced in the early 1970s. This analogy can be read as a warning about risks when recalling past financial crises.

02 THE PLAYBOOK {#rps-6}

Three Evaluation Criteria Confirmed from Precedents {#rps-7}

The following three conditions are analytical criteria derived from the retrospective classification and comparison of precedents, and they are not verified necessary or sufficient conditions. There are cases that have failed despite meeting all three conditions, and there are also successful cases that did not meet some conditions; these have not been exhaustively verified. The observability of collateral prices and the maturity structure are included as sub-elements of conditions 2 and 3 without separating them as distinct conditions.

The financial structures being compared can be broadly divided into three types. First, there is the method of utilizing cash flows generated from contracts as collateral. This includes mortgage-backed securities and Dunkin's royalty-backed financing. Second, there is the method of using the residual value of physical assets that can be resold as collateral, with aircraft and containers being representative examples. Third, there is vendor financing, where suppliers directly provide funds to customers purchasing their products. A notable example of this financing method was used by telecommunications equipment companies from 1998 to 2001, often without setting up separate collateral unlike the first two types. Comparing these three types reveals the factors that determined their success or failure.

Successful Cases. The first MBS was issued in 1970 by Ginnie Mae, bundling government-backed home loans, with the U.S. government's guarantee also applying to principal and interest payments.

The case of Dunkin' Brands illustrates how the collateral structure enhances the creditworthiness of bonds. Dunkin' raised $1.7 billion by using royalty income paid by franchisees to the headquarters as collateral. At that time, the company's bank loan rating was speculative grade (B+ according to S&P), but the bonds issued based on that royalty received an investment grade (BBB). The bond rating was five notches higher than the company's bank loan rating. The key was the priority of cash flow payments from royalties. Royalties are cash flows that are paid before other costs from franchise sales, allowing them to remain relatively stable even if the headquarters' performance deteriorates. Subsequent issuances of $2.6 billion in 2015, $1.2 billion in 2019, and $900 million in 2025 continued this structure for 19 years.

Similar financial structures have been utilized for other infrastructure assets. In the telecommunications tower sector, on the same day in October 2024, SBA Communications and Vertical Bridge raised $2.07 billion and $1.94 billion, respectively. In the fiber optic cable sector, Easy Fiber issued bonds worth $880 million backed by monthly recurring revenues of $12 million. In the data center securitization market, which is most similar to this structure, a total of 88 transactions worth approximately $48.7 billion have been completed since its introduction in 2018. As of December 2024, 84% of the 42 rated by S&P were rated A. However, this distribution was during a period dominated by cloud leasing structures. Container lease securitization has also been operated for decades using highly mobile physical assets as collateral.

Failure and Underperformance Cases. The exposure of nine suppliers participating in vendor financing for telecommunications equipment from 1998 to 2001 totaled approximately $25.6 billion. The proportion of defaults surged within just one year, with Lucent rising from 2.6% to 60%, Nortel from 25.5% to 80%, and Motorola from 6.7% to 57%. The aircraft securitization market has continued, but in 2020, 32 out of 86 tranches rated by Fitch were downgraded.

Comparing successful and failed cases, the factors that determined success or failure can be summarized into three conditions.

Condition 1. Is the collateral value separated from the borrower's credit? The failed cases examined earlier all had a high correlation between collateral value and borrower credit. [Interpretation] The core issue in the 2008 mortgage-backed securities market crisis was not the mere fact of using homes as collateral. The simultaneous deterioration of borrowers' repayment ability and the decline in housing prices exacerbated losses. Of course, this factor alone cannot explain the causes of the financial crisis. Multiple factors, including credit ratings, securitization structures, and leverage, interacted, and this analysis focuses on the correlation between collateral value and borrower credit. A similar issue appeared in vendor financing as the repayment ability of telecommunications companies was tied to the same variable as the value of telecommunications equipment, which was the telecommunications investment cycle. In contrast, successful cases utilized structures that reduced the correlation between collateral value and borrower credit. Data center securitization secures investment-grade hyperscalers as major tenants, thereby lowering the risk of tenant credit risk and building value being influenced by the same factors. Dunkin' also prioritized cash flows from royalties to achieve a similar credit enhancement effect.

Condition 2. Is there a market to recover and resell the collateral? This is a key condition determining the effectiveness of collateralized loans. There are decades of accumulated cases where aircraft have been repossessed and re-leased to other airlines, and containers can also be sold as repayment resources. In contrast, vendor financing often lacks collateral structures, leaving no collateral to recover in the event of defaults. There is one prerequisite here: an objective market indicator is needed to observe collateral prices. MBS has a housing price index, and aircraft have established appraisal standards and practices. Without reliable price benchmarks, it is difficult to assess the appropriate value even if collateral is recovered.

Condition 3. Is there an entity that can provide credit enhancement until the first two conditions are sufficiently verified? Initial investors in new asset classes must bear risks that are not yet sufficiently validated. There were two main devices to mitigate this initial uncertainty. One is the 'anchor credit' provided by the government or reputable companies. The first MBS served as a credit base with government guarantees until the market could adequately assess the risks. The reason Dunkin bonds could receive ratings five notches higher than the company's bank loan rating was due to the priority of royalty cash flows acting as credit enhancement. The other is the maturity structure. Container securitization set the legal maturity at 20-25 years, limiting it within the economic lifespan of the collateral, while aircraft evaluations reflected recession scenarios from the outset, applying a 40-75% additional discount to residual values. While guarantees share the risk of loss with a third party, the maturity structure reduces the duration of exposure to risk. Although the approaches differ, the goal of buffering initial uncertainty remains the same.

03 THE SCORECARD {#rps-8}

Evaluation of Three Conditions and Unverified Recovery Market {#rps-9}

Based on the three conditions derived in the previous chapter, the current structure was evaluated. The evaluation is divided into three stages: 'met', 'partially met', and 'not met'. If the structure actually exists and operational cases are confirmed, it is rated as 'met'; if the structure or contractual terms exist but actual operation is not confirmed, it is rated as 'partially met'; if neither is confirmed, it is rated as 'not met'.

Two principles were applied in the evaluation. It was determined that the actual operation of the market cannot be verified solely by contractual terms, and that futures scheduled for listing are not considered to have validated price discovery functions until actual trading begins.

As a result of the evaluation, Condition 1 and Condition 3 were rated as 'partially met', while Condition 2 was judged as 'not met'. However, Condition 3 has the most limited basis among the 'partially met' due to the lack of disclosure regarding contractual wording and loss burden ranking.

Condition 1: Partially Met. The first financial case securing an investment grade has been confirmed. Among the borrowers in this market, many AI operators lack credit ratings, and their repayment ability and the value of GPUs are significantly influenced by AI computation demand. If AI computation demand decreases, both the borrower's revenue and collateral value may decline simultaneously. Facility collateral securitization has combined the contractual cash flows of high-credit tenants to lower these risks. In collateralized financing for computing equipment, the participation of anchor tenants, the establishment of over-collateralization, and a sufficiently sized junior tranche can serve similar credit enhancement roles.

CoreWeave's A3 rating is based on a structure that combines not only GPUs but also the contractual revenues of investment-grade large customers as collateral. In other words, rather than completely separating the borrower's credit risk and collateral value, it is closer to a structure that supplements it by adding the credit of reputable customers. However, it is also difficult to consider the creditworthiness of these customers as entirely independent of AI investment flows. The diversification of loan participants mentioned by Zelter can also be a factor that mitigates risk. However, just because six financial institutions lend to the same asset class does not mean that the correlation between collateral values decreases. Reflecting this case, Condition 1 can be rated up from 'not met' to 'partially met'. However, the currently confirmed case is only one private loan, so it is necessary to verify whether the same structure is repeated in actual transactions on the platform in the future.

Condition 2: Not Met. Residual value support alone cannot replace the recovery market. In GPU collateral loans, no actual collateral recovery cases have been confirmed yet. There have been no cases where creditors recovered collateral due to borrower defaults. Therefore, there is insufficient basis to verify whether the collateral recovery and disposal procedures operate smoothly in actual situations.

However, CoreWeave could be the first case to verify the recovery structure in the future. A loan agreement of $7.5 billion secured by GPUs and customer contracts began repayment in January 2026. At the same time, the market value of the H100 included as collateral was being significantly reassessed. However, data on the price trends of the H100 are conflicting. Subsequently, CoreWeave completed a refinancing of $8.5 billion through DDTL 4.0 on March 31. This is the first case to receive Moody's A3 and DBRS A(low) ratings among financial products secured by high-performance computing infrastructure and customer contracts. The interest rate was set at SOFR + 2.25% for variable rates and approximately 5.9% for fixed rates, which is about half the average interest rate of around 11% for loan agreements in 2024. However, this is not a public securitization but a private loan, and the collateral includes not only GPUs but also the contractual revenues of large customers. The mere fact that refinancing was successful without default does not verify the operation of actual collateral recovery procedures.

In Bitcoin mining machines, which are similar in nature to GPUs, actual collateral recovery cases have already occurred. The scale of equipment collateral loans for mining companies once reached about $4 billion, and in 2022, the prices of mining machines plummeted. Based on publicly available cases alone, the scale of defaults is estimated to be around $227 million to $238 million. In some transactions where defaults occurred, the mining machines set as collateral were recovered by the major shareholders. Stronghold returned about 26,200 mining machines and settled $67 million in debt, while Core Scientific delivered 27,403 machines and settled $38.6 million in debt.

This case highlights three key points. First, even if all collateral is recovered, its value may fall short of the outstanding loan balance. The assessed value of the mining equipment provided by Core Scientific was $25 million, while the outstanding loan balance at that time was $38.6 million. During the same period, the price of new mining equipment dropped by approximately 85% within a year. Second, there may be situations where the lender cannot sell the recovered equipment and must operate it directly. NYDIG, Galaxy, and Foundry chose to operate the recovered equipment instead of leaving it idle, and Foundry even acquired related facilities through bankruptcy auctions. Third, additional price discount pressures may arise during bulk sales. In practice, bulk sales often occur at prices discounted by 10-25% compared to fair value.

There are differences in asset utilization between GPUs and Bitcoin mining machines. Mining machines have limited uses, while GPUs can be utilized for various computational tasks. This is why NVIDIA emphasizes the interchangeability and redeployability of GPUs. However, in the case of mining machines, significant value loss occurred during the actual recovery process of the computational equipment.

NVIDIA's residual value support is a mechanism prepared for situations where the recovery market is not sufficiently established. It is structured to cover part of the shortfall if the recovery amount is insufficient after collateral disposal. [Interpretation] This can be seen as a mechanism established on the premise that the liquidity and price discovery function of the recovery market have not yet been sufficiently validated. If the liquidity of the recovery market were sufficient, the need for such support would be relatively low. Therefore, it cannot be said that condition 2 is met solely by residual value support. Residual value support is not a mechanism for forming a resale market for collateral but rather a structure that requires the supplier to bear part of the losses incurred during the collateral disposal process. Repeated actual transaction records and sufficient market liquidity must accumulate to validate actual transactions.

The effective interest rate stated in the quarterly report at the end of March 2026 is calculated separately, and the rates presented in the table are based on the contract rate at the time of procurement. Source: CoreWeave IR, 8-K, Comprehensive Data on DDTL 4.0 Conditions.

NVIDIA is also establishing mechanisms to reduce uncertainties in the recovery process. Borrowers must apply the system structure designated by NVIDIA, and in case of issues, a third party is designed to take over and operate the equipment. While recovery procedures can be established through contracts, actual recovery cases cannot be secured. The reason aircraft collateral financing gained market trust is not because lease contracts included recovery clauses, but because there are accumulated cases where those clauses operated repeatedly in actual default situations. Currently, NVIDIA's structure is at a stage of preparing a contractual basis for future actual recovery cases to accumulate.

For a recovery market to form, there is also a need for objective standards to evaluate collateral prices. Evaluation criteria have already been established for facility assets. S&P announced data center securitization evaluation criteria in June 2024, and Moody's classified assets into five groups in its dedicated methodology in February 2025, applying revenue volatility of 5-23% for each group. However, both evaluation criteria target data center facilities. Fitch evaluated AI training facilities relatively unfavorably compared to general cloud facilities.

The data center facility collateral market is expanding rapidly. The first transaction was a $900 million securitization issued by Vantage in February 2018, and by 2020, there were only 3-4 issuers. However, since 2021, issuance has significantly expanded, with annual issuance expected to reach approximately $25 billion by 2025 and $40-70 billion by 2026. However, this growth pertains to facility collateral markets. The issue of objectively evaluating market prices for equipment like GPUs is separate.

CME Group and Silicon Data announced on August 11 that they will list two compute futures on October 5, subject to regulatory approval. The planned products are H100 rental index futures and B200 rental index futures, which will be listed on NYMEX and traded in cash settlement. Silicon Data is the institution that calculates the related indices with the support of trading company DRW.

However, there are three limitations before these futures can function as a benchmark for evaluating the recovery value of GPUs. First, just because futures are listed does not automatically secure sufficient trading volume and liquidity. Second, the underlying index of the futures reflects GPU rental prices but does not directly measure resale prices. Third, if the collateral targets expand to GPUs of different generations in the future, the measurement targets of the underlying index may not align with the actual collateral assets. The assumptions of residual value applied in securitization depend on how much can be recovered during actual collateral disposal. In the first half of 2026, while rental prices were rising, the residual value showed different movements due to separate factors. Estimates of residual value also vary across sources. Some sources estimate a drop from $30,000 to $8,000, while others indicate a subsequent rebound. Commonly confirmed is that prices can be reassessed significantly in a short period rather than declining steadily at a constant rate. Silicon Data, which suggested a rebound, is also the institution that calculates the underlying index for CME futures.

Condition 3: Partially Met. There are credit enhancement devices, but the scope of application is uncertain. NVIDIA proposed residual value support as an external credit enhancement measure. Applying 25% to the platform target amount of $500 billion gives $125 billion, but this is a simple conversion applying the support ratio to the total target amount, not the actual amount NVIDIA has committed. The actual support and scale will be determined on a project-by-project basis, and it is not disclosed which projects will receive what amount of support.

To gauge the support capacity, looking at NVIDIA's financial status, the revenue for the fiscal year 2026 was $215.9 billion, with a net profit of $117 billion, and cash and securities at the end of the period were $62.6 billion. The structure of existing guarantees can be partially confirmed through disclosures. At the time of the initial disclosure for the third quarter of the fiscal year 2026, the maximum exposure of guarantees was $860 million, and the escrow deposit was $470 million. The cumulative maximum exposure of facility lease guarantees disclosed later is $3.5 billion. This exposure decreases by the amount paid by partners to the lessor, and $712 million is deposited in escrow. The guarantee period is 5-7 years, and NVIDIA assessed that the fair value of these guarantees does not significantly impact the financial statements.

However, it is difficult to assess NVIDIA's burden capacity based solely on the currently available information, as the actual scale of residual value support, triggering conditions, and total exposure amount have not been determined. The previously calculated $125 billion is merely a simple conversion figure and not the actual contracted amount, making it inappropriate to directly compare it with the $62.6 billion in cash and securities. The guarantee structure for the Ohio project was clarified through a disclosure on August 17. The actual disclosed guarantee targets an initial lease of 4.25 GW, with a cumulative payment limit of $105 billion. This exceeds the end-of-period cash and securities of $62.6 billion, but since individual guarantees will take effect sequentially from 2028 and the lessee will repay the actual payment amounts, it is necessary to distinguish between the payment limit and NVIDIA's net exposure.

NVIDIA's residual value support can be evaluated from three aspects: obligation, support scope, and loss burden structure.

First, it is not a contractual obligation. The government guarantee applied to the initial mortgage-backed securities was a legal obligation to guarantee principal and interest payments, and investors had the right to claim payments based on the guarantee. In contrast, NVIDIA has merely stated that residual value support is possible, and the actual application will be determined on a case-by-case basis. This means that even in situations where the need for support increases, the actual decision to provide support rests with NVIDIA.

Second, compared to other transactions concluded around the same time, the scope of support is limited. Meta guaranteed the residual value of the entity established with Blue Owl for the initial 16 years of operation. If the lease is not renewed or is terminated early, Meta compensates for any shortfall in repaying principal and interest with the proceeds from asset sales. Based on this structure, bonds worth $27.29 billion received an S&P A+ rating and were issued with a spread of 225 basis points over U.S. Treasuries. Broadcom has committed to fully covering any shortfall for senior investors in the platform established with Apollo and Blackstone. If Anthropic fails to fulfill its lease payment obligations, a special purpose company will dispose of the chips, and any shortfall not covered by the proceeds will be fully paid by Broadcom to senior investors. The scope of support indicated by NVIDIA is significantly more limited than these transactions, and losses occurring outside the scope of support will be borne by the investors.

Third, specific contractual language and loss burden ranking have not been disclosed. It has not been made clear which loss intervals the support applies to, at which position among senior and subordinated ranks it operates, and under what conditions it is triggered. Relevant information has not been included in platform press releases or SEC regular disclosures, and has only been disclosed through the company's official blog.

The nature of the credit enhancement entity is also different. Genimay is a government agency, while NVIDIA is a private company. NVIDIA's capacity for credit enhancement is limited, and the company's creditworthiness is partially linked to AI demand. As guarantees are executed, NVIDIA's risk exposure may increase, and the issue of off-balance-sheet leverage accumulation pointed out by the U.S. Senate is also related to this risk transfer structure. There is a structurally similar aspect to the vendor financing of telecommunications equipment in the late 1990s, where suppliers backed the debts of customers purchasing their products. At that time, loans were made up to 130% of the equipment price without separate collateral, but this structure has a residual value support cap of 25% and includes collateral and third-party acquisition and operation conditions, which is a significant difference. Whether these safety measures lead to actual loss rates and market stability differences will need to be verified in the future.

Along with credit enhancement, the maturity structure is also important. Estimates of the economic lifespan of GPUs vary between 2-3 years and 4-6 years. In 2025, Amazon and Meta also adjusted the useful life of similar GPU assets differently. There are limitations to using assets with uncertain economic lifespans as collateral for long-term bonds. However, the maturity structure can be adjusted to reduce this risk without changing the collateral assets. This can be done by setting the bond maturity within the economic lifespan of the assets, similar to container securitization. Short expected repayment periods, rapid principal amortization, and sufficient over-collateralization can be utilized. Reflecting recession scenarios from the beginning in aircraft evaluations and additionally discounting the residual value by 40-75% follows the same principle.

C CONCLUSION

The Verification of Recovery Markets is the Most Lacking Among the Three Conditions

Comparing the successful cases examined earlier reveals three common characteristics. These three conditions are not arbitrarily set criteria but rather a summary of common characteristics identified through the comparison of previous cases. Currently, conditions 1 and 3 are assessed as 'partially met.' There is only one case supporting condition 1, which is a private loan. Condition 3 has a maximum residual value support of 25%, but it is not a contractual obligation, and specific contractual language and loss burden ranking have not been disclosed. Compared to Meta's full guarantee for 16 years or Broadcom's 100% coverage of senior shortfalls, the scope of support is relatively limited. Among the three conditions, only condition 2, which pertains to the collateral recovery and resale market, is assessed as 'not met.'

The reasons for evaluating condition 2 as 'not met' are threefold. First, there are no actual collateral recovery cases. There have been no confirmed cases where creditors have recovered and disposed of collateral due to borrower defaults in GPU collateral loans. While there have been decades of accumulated recovery and re-leasing cases in aircraft financing, such cases have not yet been confirmed in GPU collateral financing. CoreWeave's successful refinancing does not serve as evidence for the verification of the collateral recovery structure, as there has been no default, and thus the actual operation of the recovery process remains unverified.

Second, while price benchmarks are being formed, the measurement targets differ. The underlying index of the futures scheduled to be listed in October reflects GPU rental prices, but what is needed for collateral recovery is the resale price of used equipment.

Third, the recovery market cannot be formed solely by contractual provisions. Condition 1 can be partially supplemented by including the cash flows of high-quality customers in the collateral, and condition 3 can be partially supplemented by adding credit enhancement clauses. NVIDIA can also design the contract structure to allow a third party to take over and operate the facilities if the borrower fails to fulfill its repayment obligations. However, the existence of such clauses and the actual formation of the recovery market are separate issues. The effectiveness can only be confirmed if collateral recovery and resale actually occur after a default. There are still no such cases in GPU collateral financing. In contrast, in similar mining equipment collateral financing, actual recoveries have occurred, but there have been cases where the total recovered collateral value fell short of the loan balance.

[Interpretation] NVIDIA's residual value support alone cannot resolve the absence of a recovery market. The formation of a recovery market and the compensation for losses during the period when the recovery market has not been sufficiently established serve different functions.

The greatest uncertainty among the three current conditions lies in Condition 2. Condition 2 cannot be easily supplemented in the short term through capital injection or contract structure alone. Conditions 1 and 3 serve as mechanisms to reduce the linkage between borrower credit and collateral value, thereby mitigating initial loss risks, but they cannot replace the recovery market itself. As the recovery market becomes sufficiently established, the reliance on additional credit enhancements may decrease. Conversely, if the market does not form, the risk burden on credit enhancement entities may persist.

However, even if Condition 2 is met, the success of the market is not guaranteed. There is currently only one case supporting Condition 1, and there are also structural failures in vendor financing that are similar in nature. The regulatory authorities are also monitoring the expansion of related debts and off-balance-sheet risk exposures.

The risks arising from the unverified status of Condition 2 are also clear. Until actual recovery cases are accumulated, the disposal value of collateral remains based on unverified assumptions.

In response to these uncertainties, the structures that issuers can choose also change. It is possible to include the cash flows from contracts with high-quality clients as collateral to supplement Condition 1, shorten maturities, and accelerate principal amortization to lower initial risks, or to raise funds primarily in the private market until the recovery market is sufficiently validated.

Indicators that can change future evaluations can also be specifically established. There are three grounds for raising the evaluation of Condition 2. These include the formation of significant trading volume and open interest after the October futures listing, the establishment of indicators reflecting the resale prices of used GPUs, and the actual execution of third-party acquisition and operation clauses in the event of default on GPU collateral loans. Conversely, if the first transactions on the platform are composed of facility assets rather than GPUs, or if it is confirmed that securing the required credit rating is difficult without residual value support, it may be necessary to reconsider the current evaluation. Condition 1 will be reassessed once the counterparties and collateral composition of the first transaction on the platform, as well as the credit evaluation details, are disclosed. Condition 3 will be reevaluated by confirming the specific contractual wording of residual value support, loss burden ranking, and whether it is reflected in disclosures.

How quickly Condition 2 is verified will be a key criterion in determining whether compute collateral financing establishes itself as an independent asset class or remains within a risk structure similar to past vendor financing. However, if transactions are conducted in a private manner and specific conditions are not disclosed, some indicators may be difficult to verify directly. In this case, the issuer's pre-issuance report, revisions to the evaluation methodology, and reports from custodial institutions can be used as supplementary indicators.

R REFERENCES {#rps-12}

Source

{#rps-13}

A APPENDIX {#rps-14}

Limitations of Data and Calculation Basis {#rps-15}

① Limitations of Analysis and Data

• Limitations of Contract Structure Disclosure: The final contracts of six platforms have not been disclosed, preventing the identification of specific collateral definitions, tranche structures, and the order of loss burden for residual value support. Therefore, Conditions 1 and 3 are provisional assessments reflecting the limitations of currently available information.
• Scope of Residual Value Support Confirmation: The maximum 25% residual value support has been cross-verified through statements from the CEO and independent reports citing these statements, but it has not been confirmed in the company's official disclosures.
• Limitations of Post-Classifications: The three conditions are analytical criteria derived from retrospective classification and comparison of past cases. There has been no comprehensive verification of cases that failed despite meeting all three conditions or cases that succeeded without meeting some conditions. During the analysis, five elements were integrated into the three conditions, and price observation and maturity structure were classified as sub-elements of the relevant higher conditions. There has also been no comprehensive verification of whether price observation and maturity structure act as independent variables.
• Limitations of Inference: When comparing the Genie May case and NVIDIA, the differences between mandatory guarantees from government agencies and discretionary support from private companies were explicitly stated in the text. The comparison with the 2008 housing market and the description of the container market's recession phase were classified as general interpretations without presenting separate individual sources, and specific figures were not used.
• Characteristics by Data Source: Successful cases primarily utilized data from law firms, appraisers, and federal agencies, while failed cases relied relatively more on post-reports and disclosures.
• Differences in Figures by Aggregation Agency: The issuance scale by year varies by aggregation agency, and thus was not presented as a single continuous time series. SFA projected the issuance amount for the same year to be $8 billion in May 2025, while RBC estimated it to be approximately $25 billion annually in December 2025. The two sources may differ in the scope of regional and transaction structures, so only the RBC figures, which can specify the aggregation timing, were used in the text. Consistent data showing the default rates and credit rating downgrades of data center securitizations were not secured. The 84% A-grade proportion is as of December 2024 and reflects the rating distribution during a period dominated by cloud leasing transactions.
• Inconsistencies Between Data: Estimates for GPU residual value and economic lifespan differ between sources, presenting conflicting figures in the text. Some institutions that produced residual value data are the same as those calculating the underlying index for CME Compute Futures.
• Detailed Conditions Using Secondary Data: The interest rate, maturity, and collateral composition of DDTL 4.0 were synthesized from multiple secondary sources. Credit ratings, procurement scales, and whether there were initial investment-grade cases were confirmed through the company's IR and the original 8-K documents. NVIDIA's revenue, net income, and ending cash and securities for the fiscal year 2026 were cross-verified through secondary sources citing 10-K figures.

② Basis for Calculation

For all figures marked as 'self-calculated' in the text, formulas and calculation criteria will be presented. If the calculation methods differ by source or specific formulas cannot be confirmed, the associated limitations will also be specified.

③ Methodology

• Data Priority: Official data from companies and regulatory agencies → Raw data and aggregation agency databases → Specialized media → Syndication media → Community data in that order.
• Data Classification: Only classify as primary data if the original text has been directly verified. If the original text is not secured, indicate that it is a secondary citation and disclose the fact of not securing the original text.
• Reference Date and Inquiry Date: Values confirmed at a specific point in time will indicate the reference date, while real-time fluctuating dashboard figures will indicate the inquiry date.
• Handling Conflicting Figures: If different figures are confirmed, both will be presented. When selecting a reference value, prioritize the public disclosure of the calculation range, clarity of the inquiry timing, and the possibility of separating detailed items over the media's recognition, and specify the basis for selection in the text.
• Self-Calculation: Self-calculated figures will be individually presented in Appendix ② with their respective formulas, reference dates, and limitations.
• Legal and Regulatory Interpretation: Regulatory agency data will only be cited within the scope of explaining general structural classifications and applicability, and will not be broadly interpreted as official classifications or approvals for specific products. Explanations of product structures and regulatory interpretations will be separated into distinct paragraphs.
• Disclosure of Conflicts of Interest: All reports will include the same disclosure statement in a fixed position at the top of the document.

Published by Bitplanet Research Lab, Written by Kim Tae-won, Reviewed by Kim Soo-young

Disclaimer: This material is an industrial analysis for informational purposes and is not an investment solicitation. It does not recommend trading specific stocks and does not present price forecasts or target prices. The author is not a lawyer or investment advisor, and all figures are based on publicly available sources (as of the time of writing, subject to change).