GPUs as Loan Collateral
GPUs score well as collateral on most traditional measures: identifiable by serial number, physically durable, and traded in a growing secondary market. The main weakness is value stability. Current GPUs are in high demand, but this could quickly change with technology shifts. Lenders should structure around this with short terms and underwriting focused on realistic project timelines,residual value insurance (RVI), and the revenue contracts ("compute offtake").
What makes good collateral
Good collateral means the lender can easily seize and sell the asset to recover their money if the borrower defaults. Six criteria makes this possible:
How collateral has worked in other industries
Aircraft.
Airlines have raised over $35B by issuing bonds with their planes as collateral. [1]Acuity Knowledge Partners, "Aircraft Securitisation: Past, Present, and Future" (2025)https://www.acuitykp.com/market-guide/aircraft-securitisation-past-present-and-future/ Every aircraft has a unique, government-registered tail number. A trade body called ISTAT [2]ISTAT, "International Society of Transport Aircraft Trading" (accessed March 2026)https://istat.org/ publishes what each model is worth every month, and an international treaty called the Cape Town Convention [3]UNIDROIT, "Convention on International Interests in Mobile Equipment (Cape Town Convention)" (2001)https://www.unidroit.org/instruments/security-interests/cape-town-convention/ lets lenders seize planes even across borders. Insurers like AFIC [4]AFIC, "Aircraft Finance Insurance Consortium" (accessed March 2026)https://www.afic.aero/ will even guarantee a minimum resale price.
Heavy equipment.
Used Caterpillar, Deere, and Komatsu machines sell through Ritchie Brothers, the world's largest equipment auction house, [5]Ritchie Brothers Auctioneers, "rbauction.com" (accessed March 2026)https://www.rbauction.com/ with published price guides. Railcars trade through leasing companies like GATX and Trinity Industries. These assets last decades and lose value slowly.
Shipping containers.
International standards fix the dimensions, so any container fits any ship, truck, or rail car. The two largest leasing companies, Triton and Textainer, own millions of units. Containers are close to a perfect commodity: fully interchangeable, long-lived, and simple to price.
Commercial vehicle fleets.
Every vehicle has a Vehicle Identification Number registered in a government title system. The two largest wholesale auction houses, Manheim and ADESA, sell millions of vehicles a year with published price guides.
Taxi medallions.
NYC capped supply at 13,587 medallions, the licenses required to operate a taxi. Because of the scarcity, prices peaked at $1.05M each in 2013, [6]NCUA, "Timeline of the NYC Taxi Medallion Crisis"https://ncua.gov/news/responding-collapse-new-york-city-taxi-medallion-market/timeline-nyc-taxi-medallion-crisis and banks lent up to 90% of that value. Then Uber arrived. [6]NCUA, "Timeline of the NYC Taxi Medallion Crisis"https://ncua.gov/news/responding-collapse-new-york-city-taxi-medallion-market/timeline-nyc-taxi-medallion-crisis By 2021 the average medallion sold for $79K, a 92% collapse. [6]NCUA, "Timeline of the NYC Taxi Medallion Crisis"https://ncua.gov/news/responding-collapse-new-york-city-taxi-medallion-market/timeline-nyc-taxi-medallion-crisis The four NYC credit unions holding $2.5B in medallion loans were seized by regulators. [6]NCUA, "Timeline of the NYC Taxi Medallion Crisis"https://ncua.gov/news/responding-collapse-new-york-city-taxi-medallion-market/timeline-nyc-taxi-medallion-crisis
How GPUs score as collateral
Over $20 billion in GPU-backed debt exists as of early 2026. [7]Bird & Bird / Lexology, "GPU-Based Financing in the Global Data Center Market" (2025)https://www.lexology.com/library/detail.aspx?g=71bf28ab-ce78-46ba-be78-9c9a09464767 CoreWeave's $2.3 billion facility in August 2023 [8]Reuters, "CoreWeave raises $2.3 billion in debt collateralized by Nvidia chips" (August 2023)https://www.reuters.com/technology/coreweave-raises-23-billion-debt-collateralized-by-nvidia-chips-2023-08-03/ set the template: NVIDIA H100 GPUs as collateral, Magnetar Capital and Blackstone as lenders. Lambda, Nscale, Crusoe, and xAI followed with facilities ranging from $500 million to $7.5 billion. [9]Reuters, "Lambda secures $500 mln loan with Nvidia chips as collateral" (April 2024)https://www.reuters.com/technology/lambda-secures-500-mln-loan-with-nvidia-chips-collateral-2024-04-04/ [10]Nscale, "$1.4bn Delayed Draw Term Loan Backed by GPUs" press release (February 2026)https://www.nscale.com/press-releases/nscale-signs-1-4bn-delayed-draw-term-loan Lenders are already treating GPUs as collateral.
Identifiability 5 / 5
Every NVIDIA datacenter GPU has a serial number and unique board ID, queryable via code. NVIDIA tracks every chip from fabrication through first sale. These identifiers are burned into firmware and readable remotely, so a lender can verify every GPU in a fleet through software alone.
Liquidation market 3.5 / 5
Brokers, ITADs (IT Asset Disposition companies), and specialized marketplaces trade H100s and A100s regularly. But the market isn't standardized. There is nobody widely publishing monthly valuations and no centralized auction house. Price discovery depends on broker relationships and direct negotiations. A lender liquidating 500 GPUs would likely need to work through multiple channels over several weeks, accepting a discount for speed.
Value stability 3 / 5
GPU depreciation is steep, driven by technology cycles. When NVIDIA launches a new architecture, the previous generation drops in resale value within 12-18 months. [11]AltStreet, "GPU Depreciation & Obsolescence Risk" (January 2026)https://altstreet.investments/reference/risk/gpu-depreciation The Hopper GPU prices held well through 2024 because demand exceeded supply, but as the new Blackwell production ramps, Hopper residual values can and will compress.
Regardless, old GPUs don't go to zero, and are still economically useful.
Physical durability 4 / 5
GPU silicon is durable. Transistors don't wear out under normal temperatures, and Epoch AI estimated a mean time between failures of roughly 50,000 hours, about 6 years. [12]Epoch AI, "Hardware Failures Won't Limit AI Scaling" (2024)https://epoch.ai/blog/hardware-failures-wont-limit-ai-scaling
What can fail more often is the supporting hardware: fan bearings, PSU capacitors, and thermal interface material that dries out over 2-3 years. These are all replaceable parts that a good maintenance program handles. [13]Based on industry experience and conversations with GPU hardware buyers, sellers, brokers, and lenders (2025-2026)
Legal enforceability 5 / 5
Standard UCC-1 filings, the form lenders file for collateral in the US, work for GPU collateral, and existing deals use them. The security interest is filed against the borrower or their special purpose vehicle (SPV) and covers the specified hardware by serial number. [14]Uniform Commercial Code, Article 9: Secured Transactionshttps://www.law.cornell.edu/ucc/9
Redeployability 5 / 5
Any datacenter operator can use H100s, since the hardware runs the same CUDA software stack regardless of who owns it.
Coverage creates a minimum value for what your GPUs are worth at a future date. If they sell below the floor, the policy pays you the difference.
Learn how it works →What to evaluate before lending against GPUs
Generation and age
Lenders should know the exact generation of GPUs and where it sits in its lifecycle. Current generation hardware with 2-3 years of market relevance carries a different risk profile than previous generation hardware entering its depreciation curve.
Warranty status
OEM warranties, typically 3 years from Dell, HPE, or Supermicro, cover manufacturer defects and include replacement services. A server with 2 years of warranty remaining is stronger collateral than one with an expired warranty.
Form factor
SXM GPUs offer higher effective performance than PCIe GPUs. They have started to represent most datacenter GPU deployments for AI, and are now easier to resell in the enterprise market.
Operating & Maintenance history
GPUs with clean air last much longer than those that are contaminated with dust/particulate matter. Thermals also matter. If temperatures were too high, or spiked too often, the GPUs can experience accelerated wear that decreases resale value.
Revenue contracts
Most GPU-backed loans don't just use the GPUs as collateral, but also the customer revenue ("compute offtake"). If the borrower defaults, the contracts may be assignable to a new operator, giving the lender a going-concern recovery path. [7]Bird & Bird / Lexology, "GPU-Based Financing in the Global Data Center Market" (2025)https://www.lexology.com/library/detail.aspx?g=71bf28ab-ce78-46ba-be78-9c9a09464767 A fleet of 1,000 H100s backing a 3-year compute contract with Microsoft is much safer than having no committed buyers (or low creditworthy buyers). Lambda's $500 million facility was harder to underwrite specifically because Lambda sells on-demand compute without long-term customer commitments. [9]Reuters, "Lambda secures $500 mln loan with Nvidia chips as collateral" (April 2024)https://www.reuters.com/technology/lambda-secures-500-mln-loan-with-nvidia-chips-collateral-2024-04-04/
| Factor | Stronger collateral | Weaker collateral |
|---|---|---|
| Generation | Current-gen (H100, H200, B200) | Previous-gen (A100, V100) |
| Warranty | Active OEM, 2+ years remaining | Expired or non-transferable |
| Form factor | Complete SXM server with baseboard | Loose GPUs without matching chassis |
| Environment | Documented temperature and maintenance | Unknown facility, no records |
| Revenue | Multi-year contracts, creditworthy counterparty | On-demand only, no committed revenue |
References
- Acuity Knowledge Partners, "Aircraft Securitisation: Past, Present, and Future" (2025)
- ISTAT, "International Society of Transport Aircraft Trading" (accessed March 2026)
- UNIDROIT, "Convention on International Interests in Mobile Equipment (Cape Town Convention)" (2001)
- AFIC, "Aircraft Finance Insurance Consortium" (accessed March 2026)
- Ritchie Brothers Auctioneers, "rbauction.com" (accessed March 2026)
- NCUA, "Timeline of the NYC Taxi Medallion Crisis"
- Bird & Bird / Lexology, "GPU-Based Financing in the Global Data Center Market" (2025)
- Reuters, "CoreWeave raises $2.3 billion in debt collateralized by Nvidia chips" (August 2023)
- Reuters, "Lambda secures $500 mln loan with Nvidia chips as collateral" (April 2024)
- Nscale, "$1.4bn Delayed Draw Term Loan Backed by GPUs" press release (February 2026)
- AltStreet, "GPU Depreciation & Obsolescence Risk" (January 2026)
- Epoch AI, "Hardware Failures Won't Limit AI Scaling" (2024)
- Based on industry experience and conversations with GPU hardware buyers, sellers, brokers, and lenders (2025-2026)
- Uniform Commercial Code, Article 9: Secured Transactions
Frequently Asked Questions
Can GPUs be used as loan collateral?
Yes. Over $20 billion in GPU-backed debt exists as of early 2026. GPUs score well on most traditional collateral criteria: they have unique serial numbers, are physically durable, and are redeployable across operators. The main weakness is value stability due to rapid technology cycles.
What is GPU residual value insurance?
GPU residual value insurance guarantees what your GPUs are worth at a future date. If the GPUs sell below the guaranteed price, the insurer pays the difference. This transfers depreciation risk from the asset owner to the insurer, enabling better financing terms from lenders.
How do GPUs compare to aircraft as collateral?
Aircraft score higher on value stability (4.5/5 vs 3/5 for GPUs) because planes depreciate on predictable curves over 25+ year lifespans. GPUs depreciate faster and less predictably, driven by technology cycles rather than physical wear. Both asset classes have strong identifiability, legal enforceability, and redeployability.
What factors affect GPU collateral value?
Five key factors: GPU generation and where it sits in its lifecycle, warranty status and remaining coverage, form factor (SXM vs PCIe), operating environment and maintenance history, and the strength of revenue contracts (compute offtake) backing the fleet.
Coverage creates a minimum value for what your GPUs are worth at a future date. If they sell below the floor, the policy pays you the difference.
Learn how it works →