Valuing a GPU: invoice, market value and stressed sale
Lending against AI hardware needs three values, not one, and a loan that amortises faster than the equipment loses value.
GPUs are becoming a financeable asset class. But a GPU is not a building: its value falls as newer chips arrive, and its income depends on how much it is used. Lending against it starts with being precise about value.
Three values
| Value | What it is | Use in underwriting |
|---|---|---|
| Invoice | The price paid to the supplier | Sets the maximum advance |
| Market value | What the same hardware sells for today, used | Tracks the collateral cushion over time |
| Stressed sale | A quick sale after a default, net of removal and resale costs | Sets how much can safely be lent |
Depreciation, not just price
Hyperscalers depreciate servers over five to six years for accounting, but economic life for frontier GPUs may be three to four years. Resale evidence from the previous generation shows that values can fall sharply once newer chips ship in volume.
Structuring around it
- Advance a conservative share of cost, stepping down over time
- Amortise the loan monthly, so the balance falls faster than the value
- Lend against installed, working equipment, with serial numbers recorded and a first charge registered
- Take revenue through a controlled account, tested against a minimum debt-service cover
- Insure the hardware and agree access with the data centre for recovery
Handled this way, GPU credit shares the discipline of bridge lending: a conservative advance, a defined repayment source and security that can be enforced.
For information only; not investment advice. See Legal & Risk.