GPU-Backed Debt: CoreWeave, Lambda and Crusoe Disclosures

Executive Summary

GPU-backed debt is a descriptive category for borrowing supported by graphics processing unit (GPU) equipment, associated infrastructure, contracted payments, or a combination of these. It does not establish a uniform security package, borrower, repayment source or level of parent support. The central credit question is which obligations and assets sit inside the financed perimeter, and what must happen before equipment becomes a cash-generating service. Practitioner analysis accordingly treats GPU finance as underwriting a contract, cash flow and service-performance model. ^[1]

The disclosures examined here show materially different structures. CoreWeave announced the closing of $8.5 billion of delayed-draw financing in March 2026, with approximately $7.5 billion of initial capacity and additional capacity after stabilization. ^[2] Its June debt table records $2.837 billion of DDTL 4.0 principal outstanding, rather than the full commitment. ^[3] Lambda confirmed that its $926 million senior secured term loan B closed in August 2026. ^[4] Its October announcement describes a separate delayed-draw facility linked to cluster commissioning milestones. ^[5] Crusoe's March 2025 announcement reports an approximately $225 million closed credit facility. ^[6] These amounts should not be added into a market-size estimate.

Public detail also differs. CoreWeave's DDTL 4.0 filing describes limited parent recourse and a 1.15x debt service coverage ratio covenant with a specified commencement trigger. ^[7] The reviewed Lambda and Crusoe releases leave substantial security, recourse and covenant documentation outside the public comparison. A facility's investment-grade label cannot be transferred to the parent company or to a different loan: issue ratings evaluate particular obligations. ^[8] “Undisclosed” therefore remains a necessary analytical result.

The practical decision is to connect legal rights, commissioning evidence and cash collection. A credit team should request the asset schedule, applicable contract rights, accepted-service evidence and the operative financial model, then reconcile them to the same borrower and reporting date. Engineering can test system readiness; counsel must assess enforceability and consents. ASHRAE's commissioning guidance makes receipt and review of test reports a hold point, a useful basis for designing an evidence package without pretending that an engineering test alone satisfies a loan's draw conditions. ^[9]

Introduction and Background

This report examines GPU backed debt as of October 15, 2026, using company announcements, Securities and Exchange Commission (SEC) filings, rating-originator material and official technical guidance retrieved for this report. Its purpose is to help bank and private-credit analysts, infrastructure investors, equipment lessors and counsel identify what supports repayment and where additional documents are necessary. It is an educational comparison, with no investment recommendation or estimate of private creditworthiness.

The label can refer to several financing objects. A server purchase, a contracted cloud deployment and a data-center building can share a physical site while belonging to different financing arrangements. Clifford Chance's analysis expressly distinguishes campus-level financing from GPU-as-a-Service and compute financing. ^[1] The engineering perspective in this GPU Smith report follows that distinction: equipment condition and operational readiness become useful credit inputs only when mapped to the financed assets and contractual obligation.

Crusoe provides a concrete boundary. Its Upper90 announcement identifies an exclusive use of proceeds for NVIDIA GPUs and supporting cloud infrastructure. ^[6] Separately, Newmark announced a $7.1 billion construction loan for the Blue Owl, Crusoe and Primary Digital Infrastructure development in Abilene. ^[10] Neither headline alone establishes the security or liabilities of the other transaction.

Even an equipment description needs precision. NVIDIA's DGX B200 specification includes eight GPUs alongside the rest of a server system. ^[11] A lender therefore needs an inventory of the financed configuration, rather than a count of accelerator chips alone. The analysis below first defines the financing vocabulary, then compares the disclosed examples, follows the deployment-to-cash sequence and identifies the documents and technical workpapers needed to close the remaining gaps.

Definitions and the Collateral-to-Cash Chain

What the financing labels establish

GPU collateral loans place equipment within the lender's security analysis, but a statement that financing purchases GPUs does not itself establish an equipment lien. The model Uniform Commercial Code (UCC) text distinguishes enforceability requirements, including a security agreement describing collateral, from the general filing rule for perfection. Applicable enacted law and transaction documents require separate confirmation by counsel. ^{[12] [13]}

The following terms should have separate fields in a credit memorandum:

• Commitment: the contractual maximum, subject to the agreement's conditions; it is not a statement of current principal.
• Borrowing base: the agreed collateral base that limits advances under an asset-based structure. ^[14]
• Advance rate: the percentage a lender lends against a collateral category; no private GPU advance rate is assumed here. ^[14]
• Delayed-draw term loan (DDTL): a term facility permitting later borrowing within its draw arrangements. CoreWeave's DDTL 4.0 is a disclosed example. ^[7]
• Special-purpose vehicle (SPV): a designated legal entity used for the transaction. DBRS describes CoreWeave's vehicle as established to own GPUs and borrow. ^[15]
• Offtake support: customer payment obligations that support the financed service. Its value depends on the actual contract and performance conditions.

An SPV label does not answer whether the parent guarantees repayment. CoreWeave's June filing separately states that DDTL 5.0 obligations are unconditionally guaranteed. ^[3] Borrower identity, guarantee scope and accounting presentation therefore need distinct answers; legal separation cannot be inferred solely from a subsidiary's name.

Four separate links

Table 1 separates the legal or operational link from the evidence needed to evaluate it. These are analytical questions, rather than assertions that every highlighted facility contains every mechanism.

The table prevents a common category error: treating equipment ownership, collateral rights and the ability to continue a customer service as interchangeable. Engineering can reconcile a physical inventory and test it. Counsel must determine whether the required rights actually travel with that inventory. Neither workstream replaces the other.

Key Changes in the Disclosed Deal Register

March 2025: Crusoe's equipment-purpose facility

The Crusoe-hosted Upper90 release describes a closed, approximately $225 million syndicated credit facility dedicated to purchasing NVIDIA GPUs and supporting cloud infrastructure. ^[6] It provides a clear equipment-purpose example, but the reviewed release does not specify the borrower legal entity, detailed collateral package, maturity, amortization, guarantees, covenant floors or current principal outstanding. Those fields remain unknown from that release.

The separate construction announcement describes financing for additional campus buildings. ^[10] Construction completion and GPU deployment can interact operationally, but this does not merge their loan obligations. A credit memorandum should record any contractual dependency only after reviewing the applicable agreements.

March 2026: CoreWeave's DDTL 4.0

CoreWeave's March announcement confirms closing and distinguishes initial from stabilized borrowing capacity. ^[2] The related filing names CoreWeave Compute Acquisition Co. VIII, LLC, or CCAC VIII, as borrower and gives a March 31, 2032 maturity. ^[7] The June note says sizing depends on eligible assets, projected debt-service coverage and project conditions. A headline commitment consequently does not establish immediately drawable cash. ^[3]

DBRS's April 3 action identifies a master services agreement with Meta Platforms and describes full amortization by maturity. ^[15] Its discussion of delivery, installation and GPU acceptance makes clear why a funded purchase and an accepted service belong at different points in the analytical chain. ^[15]

August 2026: Lambda's priced-then-closed term loan

Lambda's pricing announcement said the $926 million facility was expected to close subject to customary conditions. ^[17] Its subsequent closing release identifies a large-scale private-cloud GPU asset-backed SPV financing. ^[4] Latham's own transaction announcement corroborates its role advising the arrangers and the transaction sequence. ^[18]

The closing release describes equipment and associated cash-flow security, full amortization and a December 31, 2030 maturity. ^[4] It does not name the SPV borrower or publish the complete guarantee and covenant package. “SPV financing” should therefore be retained as the disclosed classification, with the legal-entity and recourse fields still requiring documents.

October 2026: Lambda's commissioning-linked delayed draws

Lambda's October release describes $1.008 billion of delayed-draw financing for three deployments involving two investment-grade offtakers, with a May 30, 2033 final maturity and full amortization. ^[5] The release specifies financed equipment and contracted cash flows as security support. It does not establish that the October borrower is the August SPV, or publish each draw amount and the complete commissioning definition.

Table 2 records the four examples at their disclosed level. A closing announcement is not treated as a dated balance confirmation, and a use-of-proceeds statement is not promoted into a complete security description.

This register is deliberately narrower than a transaction tracker. Its useful unit is a facility with a date, legal perimeter and evidence boundary. Empty cells should become document requests. They should not become assumptions about missing protections or a reason to substitute another borrower's terms.

Implementation Considerations and Process Changes

Follow the obligation from procurement to collection

The relevant timeline is a sequence of evidence states. ASHRAE distinguishes factory acceptance, installation checks, functional testing and integrated systems testing; an earlier stage does not by itself establish completion of a later stage. ^[9]

• Procurement: identify ordered equipment, seller, configuration and the entity obligated to pay.
• Delivery and title: reconcile serial numbers, location, receipt records and ownership evidence.
• Installation: document available power, cooling, network paths, firmware and storage configuration.
• Commissioning: retain the approved test scope, raw results, exceptions and remediation evidence.
• Accepted service: obtain the contractual acceptance record and the applicable service-start date.
• Cash collection and debt service: reconcile invoices, credits, receipts, controlled accounts and scheduled payments.

This is an analytical workflow, not an assertion that every facility permits drawing at the same stage. A deployment can pass an equipment test before customer acceptance, and a draw condition can require documents beyond the technical report. The operative agreement determines the legal gate.

CoreWeave's June filing describes committed contracts whose terms generally begin on a fixed date or when contracted capacity is delivered. ^[3] DBRS separately describes availability-linked payment deductions in the DDTL 4.0 customer arrangement. ^[15] Take-or-pay language therefore requires examination alongside commencement, availability and remedy provisions. It does not justify treating all future contracted revenue as unconditional cash already available for debt service.

Map shared dependencies without merging obligations

For each financed deployment, the analyst should build a dependency map covering the campus, electrical supply, cooling plant, server configuration, network and storage. The purpose is to identify which condition an engineer can measure, which party controls it and which contract allocates its consequences. The physical dependency can be common even when the borrower and security are different.

Power requires particular care. The Department of Energy defines power usage effectiveness (PUE) using annual facility and information technology (IT) energy. ^[19] An annual efficiency ratio does not, by itself, prove sufficient instantaneous electrical capacity or the date that capacity becomes available. A calculation must state whether its input represents measured demand, design load or maximum equipment input.

The GPU power calculator can be used to organize an illustrative sizing exercise. The resulting assumptions should be reconciled to vendor specifications and the site's electrical design before they enter a financing model. That internal link is a planning aid, not evidence about any disclosed facility.

Allocate work to the responsible discipline

Engineering should report what was inspected, measured or reproduced, the conditions of the test and unresolved exceptions. Counsel should assess the security grant, applicable law, consent requirements, guarantee scope and the enforceability of the proposed transfer or operating transition. Credit underwriting should translate those findings into the repayment model and approved conditions.

The separation is consequential. The UCC model provision on collection and enforcement expressly does not determine whether an account debtor owes a duty to the secured party. ^[20] A technical statement that a cluster can operate cannot establish that a customer must pay a successor operator. Likewise, a lawyer's conclusion about a security interest cannot establish that a relocated system will achieve the required workload performance.

GPU-Backed Debt Covenants and Public Information Gaps

Read the actual test and its commencement

Debt service coverage ratio (DSCR) compares the agreement's defined cash-flow measure with its defined debt-service requirement. The exact definitions, test period and commencement provisions matter as much as the ratio itself. CoreWeave's DDTL 4.0 summary discloses a 1.15x minimum beginning in the first full calendar month after the earlier of zero remaining commitments or June 30, 2027. ^[7] The agreement separately uses 1.20:1.00 in top-up debt sizing. These are different tests, not competing descriptions of one covenant. ^[21]

The agreement's maintenance test refers to the most recently ended three full fiscal months. Its DSCR definition includes an initial exception: for calculations performed before the date three full fiscal months after the Commitment Termination Date, the measurement period runs from the Commitment Termination Date until the date of measurement. ^[21] A model using a different measurement period could produce a different result even with the same nominal threshold. Readers should therefore request the definitions and calculation certificate, rather than copying the ratio into a generic template.

Later facilities also require separate records. CoreWeave's August DDTL 5.5 filing discloses an unconditional parent guarantee and a 1.35x DSCR covenant beginning in the first full calendar month after the earlier of zero commitments or December 31, 2026. ^[22] Its announcement describes approximately five-year debt alongside customer contracts averaging approximately three years. ^[23] These disclosures are relevant to the later facility; they do not silently amend DDTL 4.0.

Treat disclosure gaps as a controlled worklist

The March CoreWeave filing identifies omitted schedules and exhibits. ^[7] Public availability of a credit agreement is therefore not equivalent to possession of every customer agreement, collateral schedule and consent. The reviewed Lambda and Crusoe releases are less comprehensive than the CoreWeave filing set. Their undisclosed provisions should remain unknown, rather than being described as absent.

The lender's document request should cover:

• Legal perimeter: borrower, guarantors, pledged entities and excluded assets.
• Availability: draw conditions, eligibility rules, reserves and any borrowing-base certificate.
• Cash control: collection accounts, operating payments, reserve funding and distribution restrictions.
• Financial tests: defined numerator, denominator, measurement period, commencement and cure rights.
• Contract support: payment commencement, credits, termination, assignment, step-in and replacement rights.
• Operating access: premises access, third-party dependencies and rights needed to continue service.
• Asset transfers: disposal permissions, replacement rules, relocation conditions and consent requirements.
• Reporting: current principal, subsequent amendments, compliance certificates and the reporting date.

This list is a request framework, not a representation that each item appears in every private agreement. Practitioner analysis identifies availability, service credits, termination and upstream dependencies as bankability questions. ^[1] The executed CoreWeave agreement also restricts certain disposals or assignments involving material project contracts and GPUs, subject to its permissions. ^[21] An engineer should supply the operating facts needed to assess such provisions, while counsel interprets their actual effect.

Technical Report Inputs for Lenders and Lessors

A useful independent engineer (IE) report should connect its conclusions to the financed inventory and the specific milestone under review. Vendor diagnostics can inform the test plan, but no vendor test establishes a lender's private contractual threshold. NVIDIA's H100/H200 guide, for example, recommends a pre-flight test before production or after servicing; its platform scope should remain explicit. ^[24]

Table 3 maps proposed workpapers to credit questions. It describes recommended evidence, not work performed on any named borrower's deployment.

The matrix links a technical observation to a decision without claiming to determine the credit decision itself. A report should identify its exclusions just as carefully as its measurements. A system-health result and a customer-workload result can answer different questions.

For a reviewable submission, the workpapers should also identify:

• Configuration: the hardware and software versions actually tested.
• Population: which assets were tested, sampled or excluded.
• Workload: the test type and its relationship to the contracted service.
• Duration: the sustained test window and any interruptions.
• Instrumentation: the source, timing and retention of telemetry.
• Exceptions: unresolved items and evidence needed for closure.
• Reproducibility: commands, settings and records needed to repeat the test.
• Approval boundary: who accepts technical evidence and who confirms contractual completion.

For network-dependent services, tests should follow the relevant topology. NVIDIA's documentation specifically identifies inter-node InfiniBand or remote direct memory access (RDMA) over Converged Ethernet (RoCE) bandwidth, including rack-to-rack paths. ^[26] A single-node diagnostic should not be presented as measurement of those paths. ASHRAE also emphasizes early instrumentation for operational baselines, making retained logs useful beyond the initial completion decision. ^[9]

Data Analysis and Evidence

Commitment utilization is a dated calculation

Observation: CoreWeave's June debt table reports $2.837 billion of DDTL 4.0 principal against the $8.5 billion facility. ^[3] Illustrative arithmetic using disclosed amounts: dividing principal by commitment gives 33.4%, rounded. This is a reporting-date ratio, not a forecast of eventual utilization, an advance rate or evidence that all remaining commitment was immediately available. The note expressly includes project-level sizing conditions. ^[3]

The calculation uses the table's principal amount. It does not reconstruct principal by adding rounded narrative components. Nor does it subtract original capital expenditure from a headline facility to estimate equity funding. Such a calculation would require a reconciled uses-of-funds schedule, permitted-cost definitions, actual drawings and the applicable reporting date.

Observation about double counting: Lambda's separate May 2026 announcement says its $1 billion credit facility upsized an existing $275 million facility. Those amounts are not independent additions to available financing. ^[28] This is why a register needs amendment and refinancing relationships before any aggregation.

Disclosure boundary: Lambda's October release provides deployment and offtaker counts, but no revenue weights by offtaker. ^[5] A count is therefore insufficient to calculate concentration by receipts or exposure. Similarly, a closed facility amount cannot be treated as current principal without a matching balance disclosure. Comparing those fields requires acknowledging the unequal detail of the public documents.

Depreciation is not a universal GPU recovery curve

Observed accounting examples: Microsoft's fiscal 2025 report gives a two-to-six-year estimated useful-life range for computer equipment. ^[29] Meta's January 2025 announcement increased the estimated useful life of certain servers and network assets to 5.5 years. ^[30] These are explicitly dated, broader asset-category accounting disclosures. They do not supply a GPU-specific liquidation curve, and this report does not present them as current policies for every borrower or every reporting period.

International Accounting Standard (IAS) 16's official overview addresses carrying amounts and depreciation within its accounting framework. ^[31] Separately, the Office of the Comptroller of the Currency (OCC) asset-based lending handbook discusses collateral appraisal using net orderly liquidation value and obtaining updates through a loan's life. ^[14] The analytical implication is to keep book value, contract-supported cash generation and estimated net realization in separate model fields. Each needs its own evidence and assumptions.

Quantify the engineering input before applying a credit assumption

Observed specification: NVIDIA specifies maximum DGX B200 system input power of 14.3 kilowatts (kW). ^[11] Illustrative arithmetic (Hypothetical Example): 100 such systems at that maximum imply 1.43 megawatts of system input before other equipment and facility overhead. The system count is an assumption, not a disclosed deployment or benchmark. Annual PUE, whose denominator is annual IT-equipment energy, should not be inserted as a peak-capacity multiplier without a separately stated design assumption. ^[19]

Evidence limitation: no private spread, advance rate, resale price or residual-value percentage is estimated here. Equipment sale proceeds are a different repayment question from operating cash generation, a distinction also reflected in the Federal Deposit Insurance Corporation (FDIC) discussion of leasing. ^[32] The purpose of the quantitative work is to make the input, date and inference visible, rather than manufacture comparability from different disclosures.

Implications and Future Directions

Compare structures rather than labels

For banks, the immediate implication is to organize the credit record around the obligation chain: borrower, equipment, operating arrangement, customer receipts and support parties. The OCC's general asset-based guidance calls for agreements to define collateral, controls, availability and proceeds. ^[14] Those categories provide a starting structure; the specific GPU documents must supply the actual terms.

For infrastructure and private-equity funds, the implication is to map asset and contract duration separately. CoreWeave's later facility announcement explicitly distinguishes debt tenor from average underlying contract length. ^[23] That is a disclosed structural feature to model, with renewals and redeployment assumptions made visible. It is not a basis for imputing identical duration exposure to another facility.

For lessors and equipment-backed lenders, transferability deserves an operating workstream alongside valuation. The NIST sanitization guidance defines the process in terms of making access to target data infeasible for a specified effort level. ^[33] A proposed transfer plan should identify affected storage media, responsibility, evidence and timing. These are recommended diligence inputs; no repossession cost or resale deduction is estimated.

For counsel, the most useful engineering output is a documented description of what a replacement operator would need to continue the service. Contract rights and operating capability should be reviewed together, with legal conclusions reserved to the applicable agreements and law. UCC model text on identifiable proceeds helps frame one part of that analysis, but it does not supply a complete transaction-specific opinion. ^[16]

Dated update policy

This comparison should be updated when public evidence changes, while preserving the date and status of each prior record. The update worklist is:

• New filings: add amendments and exhibits under the existing facility identifier.
• New balances: replace current-principal fields only with dated balance evidence.
• New announcements: distinguish pricing, commitments, closing and subsequent funding.
• New technical evidence: identify the milestone, tested configuration and acceptance authority.

Ratings also require an entity and action date. S&P distinguishes issuer assessments from ratings on specific instruments and their terms. ^[8] This report dates the rating-originator evidence it uses; it does not turn an earlier action into an assertion of an unchanged current rating. A future update should retain that discipline.

Conclusion

GPU-backed debt becomes understandable when the review starts with the legal obligation and follows the equipment through deployment, acceptance and cash collection. Hardware can provide collateral, customer contracts can support repayment, and a parent or project entity can bear additional obligations. Their respective roles must be established from the relevant documents.

The disclosed examples support a structured comparison, while leaving different questions open. A commitment, a closing and a principal balance are different data fields. An SPV label, a security description and a guarantee are different legal facts. A delivered GPU, a completed installation and an accepted service are different operating states. The review should preserve each distinction rather than compressing them into one financing headline.

The practical output is a documented register and an evidence-based request list. Analysts can then identify which assumptions belong in a cash-flow model, which need an appraisal, which require technical testing and which require counsel's interpretation. Unknown private provisions should remain unknown until the required documents arrive.

For independent pre-close technical due diligence, see GPU financing due diligence. Its relevant role in this process is to help organize the technical questions that a credit team and its counsel need answered.

References

All sources below were fetched for this report on October 15, 2026. Dates in the discussion identify the source event or reporting period; undated technical pages are treated as documentation accessed on that date.

1. CoreWeave, March financing closing announcement. ^[2]
2. CoreWeave, March transaction filing and guarantee/covenant summary. ^[7]
3. CoreWeave, executed DDTL 4.0 credit agreement. ^[21]
4. CoreWeave, June quarterly report and debt table. ^[3]
5. Morningstar DBRS, April CCAC VIII rating action. ^[15]
6. CoreWeave, August DDTL 5.5 transaction filing. ^[22]
7. CoreWeave, August financing announcement. ^[23]
8. Lambda, August term loan closing announcement. ^[4]
9. Lambda, October delayed-draw financing announcement. ^[5]
10. Crusoe and Upper90, GPU financing announcement. ^[6]
11. Newmark, separate Abilene construction financing announcement. ^[10]
12. Lambda, original pricing announcement distributed by Business Wire. ^[17]
13. Latham & Watkins, arranger-counsel transaction announcement. ^[18]
14. OCC, Asset-Based Lending handbook. ^[14]
15. Cornell Legal Information Institute, model UCC security attachment provision. ^[12]
16. Cornell Legal Information Institute, model UCC proceeds provision. ^[16]
17. Cornell Legal Information Institute, model UCC perfection provision. ^[13]
18. Cornell Legal Information Institute, model UCC collection provision. ^[20]
19. International Financial Reporting Standards (IFRS) Foundation, IAS 16 overview. ^[31]
20. S&P Global Ratings, issuer and issue ratings overview. ^[8]
21. FDIC, Loans and Lease Financing examination manual. ^[32]
22. Clifford Chance, GPU infrastructure financing and contracting analysis. ^[1]
23. NVIDIA, DGX B200 hardware and power specifications. ^[11]
24. Department of Energy, cooling and PUE guidance. ^[19]
25. NVIDIA, H100/H200 pre-flight testing documentation. ^[24]
26. NVIDIA, system-management stress-test documentation. ^[25]
27. NVIDIA, node/rack network-validation documentation. ^[26]
28. ASHRAE, commissioning and performance-validation framework. ^[9]
29. NIST, media-sanitization guidance overview. ^[33]
30. NIST, full media-sanitization guidance and records. ^[27]
31. Microsoft, explicitly dated fiscal 2025 accounting disclosure. ^[29]
32. Meta, explicitly dated January 2025 accounting-estimate announcement. ^[30]
33. Lambda, original May credit-facility upsizing announcement. ^[28]