Analysis · Cross-sector lessons · 5 min read
Is there a Cape Town Convention for compute? What aircraft repossession teaches GPU lenders
Aircraft lenders can find, register and repossess their collateral under an international treaty. GPU lenders have nothing comparable. The 2026 deals show how the market is filling the gap with contracts.
- 01Aircraft finance has an international registry, priority rules and set repossession timelines. GPUs have none of these.
- 02GPU debt runs for most of the hardware's useful life, so recovery depends on the next customer rather than a long tail of asset value.
- 03Direct agreements, remarketing rights and contract value tests are the market's workaround.
An aircraft is a mobile, high-value asset that spends its life in other people's jurisdictions. Lenders solved the enforcement problem with the Cape Town Convention and its Aircraft Protocol: an international registry that records interests in airframes and engines, a first-to-register priority rule and, in states that have adopted the relevant insolvency remedy, a fixed waiting period after which the creditor can take the aircraft back. Lenders also commonly take an irrevocable de-registration and export request authorisation, or IDERA, so that the aircraft can be moved.
A GPU cluster is also mobile and valuable. But there is no registry for GPUs, no treaty and no agreed timeline for getting them back.
Why repossession is harder for GPUs
- The hardware sits in someone else's building. Racks, power and cooling usually belong to a colocation provider, which has its own landlord and lenders.
- Title can be hard to trace. Thousands of servers, often without separate title records, make it harder to show which assets secure which loan.
- Moving the hardware can raise export control questions. Advanced AI chips are subject to export controls, so moving them across borders after enforcement may need licences.
- Value falls quickly. A lender that takes months to recover a cluster recovers an older one.
GPU debt runs for most of the hardware's life
Show the data table
| Asset | Useful life (years) | Typical debt tenor (years) | Note |
|---|---|---|---|
| GPU servers | 4 to 6 | 2.5 to 6 | Lenders' models assume six years straight-line (IREN, CoreWeave DDTL 2.0); 2026 loans run from 30 months to about six years |
| Data centre buildings | 20 to 40 | 5 to 23 | ABS anticipated repayment dates of about five years to amortising lease-backed bonds to 2049 |
| Commercial aircraft | 25 to 30 | 7 to 12 | Typical secured loan, EETC and ABS amortisation profiles |
| Wind and solar | 25 to 35 | 15 to 20 | Typical contracted project finance tenors |
| Telecom towers | 30 to 40 | 5 to 10 | Typical tower ABS anticipated repayment dates |
The tenor problem
The chart shows why this matters. An aircraft with a 25 to 30 year life might carry 7 to 12 years of secured debt, so a lender that repossesses still holds an asset with much of its life ahead of it. Lenders' models in the public GPU loans assume a six-year straight-line life, and 2026 GPU loans run from 30 months to about six years. The debt covers most of the hardware's life.
That is why GPU lenders have so far relied on the contract rather than the hardware. Most 2026 loans fully amortise within the customer contract, so the lender is repaid before residual value comes into play.
How the 2026 deals fill the gap
- Direct agreements. IREN's Hardware 3 lenders have direct agreements with its colocation and managed services providers, giving them rights to step in and keep the cluster running.
- Bringing the building inside the security. Nscale's Ward County loan makes the data centre company a guarantor. See Splitting the GPUs from the building.
- Remarketing. IREN has a right to remarket the GPUs, with proceeds applied to prepay the loan, an idea familiar from aircraft leasing.
- Contract value tests. CoreWeave's DDTL 5.5 allows re-leasing but traps cash if the projected value of contracts falls below 2.40x of the loans.
- Depreciated cost tests. IREN caps debt at 65% of cost less depreciation.
Questions for the market
- Would a registry for high-value compute hardware, public or private, make GPU loans cheaper?
- What would an independent appraisal standard for GPUs look like, and who would provide it?
- How should a direct agreement deal with the colocation provider's own lenders?
- As loans start to outlive their contracts, who will price the second customer?
For in-house counsel
Take it to the documents
The filed wording behind this piece, for the lender’s or the borrower’s side. Previews are open; the full text is free with InfraCompute Weekly.
- GPU loan summariesFiled GPU credit agreements side by side: advance rates, covenants and the collateral package.
- Financing clause libraryCash traps, coverage ratios, cures, step-in and GPU security, quoted from credit agreements and indentures.
- Issues list: GPU credit agreementBuild it from the lender’s or the borrower’s side, with where filed deals land.
- The legal desk for in-house counselContract library, issues lists, checklists and English law notes.
Sources
- IREN Hardware 3: SEC filing: sec.gov
- Nscale SEC filing, September 2026: sec.gov
- CoreWeave DDTL 5.5: CoreWeave release: investors.coreweave.com
