Silicon Lien · Research · Data & trackers
How Long Can a Financed GPU Keep Earning? — July 2026
Dated note · August 11, 2026
A GPU can cover debt today and still fail to recover its deployed capital over its useful economic life. Those are different questions. Lenders generally look first to cash flow for repayment; an owner or lessor also asks whether the asset recovers its cost and required return.
Summary
For the lender, the first test is whether cash flow repays the debt. Residual value becomes relevant if principal remains at refinancing, maturity or default, and should be treated as a separate, conservatively supported recovery assumption. An asset owner or lessor can also test whether operating cash flow and terminal value recover the full deployed cost and required return.
Separate the lender and owner questions
| Perspective | Question | Primary measures |
|---|---|---|
| Lender credit analysis | Will cash flow repay scheduled interest and principal? If not, how much exposure remains and what recovery is realistically available? | Cash available for debt service, coverage, amortization, remaining principal and haircut-adjusted recovery value. |
| Asset-owner or lessor economics | Will the GPU recover its deployed cost and required return over its economic life? | Operating cash contribution, lifecycle expenditure, capital-recovery charge, economic NPV and terminal value. |
The distinction matters. The OCC describes operating cash flow as the primary repayment source in leveraged lending and asset disposal as a secondary source. S&P Global Ratings describes GPU residual value as limited because of rapid depreciation and notes that GPU equipment securitizations may amortize principal toward zero. OCC; S&P Global Ratings.
Extend the same H100 example
Parts 1 and 2 established current-period earnings and the cash required for debt coverage. Part 3 asks whether those economics can persist long enough to recover the deployed system’s capital.
For illustration, suppose the same H100 has the following hypothetical, per-GPU allocations. These are modeling assumptions, not market estimates or borrower observations:
- Deployed-system cost: $40,000, including the GPU server and allocated networking, storage, racks, power and cooling infrastructure, and installation.
- Recovery period: four years.
- Terminal residual value: $8,000 after four years, used to illustrate the asset-owner calculation rather than as a lender base-case recovery assumption.
- Required annual return: 10%.
- Annual cash operating cost: $5,000.
- Annual lifecycle capital expenditure: $1,000.
- Service availability: 100% in the initial levelized test.
- Usage-billed utilization: 70%, or 6,132 hours.
Asset-owner calculation: test a selected recovery period
This is an asset-economics calculation, not a standard debt-service test. It asks what level annual earnings would allow the owner to recover the deployed cost and a required return over a chosen period.
Let C be deployed-system cost, R expected residual value after n years and r the annual required return:
Capital-recovery factor = r(1 + r)n ÷ ((1 + r)n − 1)
Annual capital-recovery charge = [C − R ÷ (1 + r)n] × capital-recovery factor
Under the stated assumptions, the annual capital-recovery charge is approximately $10,900 per GPU. Adding $5,000 of cash operating cost and $1,000 of lifecycle capital expenditure produces a pre-tax economic cost of approximately $16,900, or $2.76 per usage-billed GPU-hour.
Pre-tax economic cost per usage-billed GPU-hour = (capital-recovery charge + cash operating costs + lifecycle capital expenditure) ÷ usage-billed GPU-hours
This is not a debt-service calculation, and it does not by itself determine economic life. It asks whether level annual economics would recover deployed capital over the selected four-year period at a 10% return. Part 2 separately tests whether transaction cash flow covers scheduled interest and principal.
Change one assumption at a time
The apparent precision of one economic-cost figure can obscure how dependent it is on the forecast. Holding the $5,000 operating cost and $1,000 lifecycle expenditure constant:
Swipe to see all columns.
| Illustrative case | Residual | Usage-billed utilization | Annual capital charge | Economic cost per billed hour |
|---|---|---|---|---|
| Illustrative residual | $8,000 | 70% | $10,900 | $2.76 |
| Zero residual | $0 | 70% | $12,600 | $3.04 |
| Lower billed usage | $8,000 | 50% | $10,900 | $3.86 |
The asset-economics analysis can extend the same table with price declines, release downtime, higher power or maintenance cost, a shorter recovery period or additional refresh investment. The purpose is to identify which assumptions cause the cohort to fall below its economic hurdle.
Asset-owner calculation: project economic life year by year
A selected-period test assumes level annual economics. An economic-life forecast instead gives each year its own price, availability, billed utilization, cash operating payments, lifecycle expenditure and residual value:
Cohort pre-tax economic NPV = − deployed-system cost + Σ[(cohort cash receiptst − cash operating paymentst − lifecycle capital expendituret) ÷ (1 + r)t] + residual valuen ÷ (1 + r)n
The declining path below is a mechanical illustration, not a market forecast. It reduces hourly price by approximately 10% each year, availability by two percentage points and usage-billed utilization by five percentage points. Annual cash operating payments follow $5,000, $5,200, $5,500 and $5,800; annual lifecycle expenditure remains $1,000. It assumes full same-period collection:
Swipe to see all columns.
| Year | Hourly rate | Availability | Usage-billed utilization | Cash receipts | Operating payments plus lifecycle expenditure | Pre-financing cash contribution |
|---|---|---|---|---|---|---|
| 1 | $2.97 | 100% | 70% | $18.2K | $6.0K | $12.2K |
| 2 | $2.67 | 98% | 65% | $14.9K | $6.2K | $8.7K |
| 3 | $2.40 | 96% | 60% | $12.1K | $6.5K | $5.6K |
| 4 | $2.16 | 94% | 55% | $9.8K | $6.8K | $3.0K |
With the $8,000 year-four residual value, this declining path has a pre-tax economic NPV of approximately negative $10,000 at the assumed 10% return. Keeping the initial price, availability, utilization and costs level for four years instead produces a pre-tax NPV of approximately positive $4,200. In both cases the GPU still earns positive pre-financing cash in year four; only the level path recovers the assumed capital and return.
The estimated economic life is the period during which continued operation remains preferable to retirement or replacement. The asset-economics forecast should extend each scenario until expected incremental cash contribution becomes nonpositive, technical or contractual constraints prevent continued use, or replacement produces greater value after switching and refresh costs. This may be longer or shorter than the period needed to recover the original capital.
Return the forecast to the lender’s model
The lender can feed each year’s cohort cash contribution into the debt-coverage analysis from Part 2 and compare it with scheduled debt service and remaining principal. A fully amortizing loan may require little or no terminal GPU value to repay. A loan with a balloon balance, refinancing dependence or early default exposure makes recovery value more important. Any assumed recovery should reflect the lender’s lien, possession and sale costs, time to disposition, hardware condition and observable market evidence—not the owner’s unhaircut terminal value.
Technology changes both revenue and cost
More efficient models and software may reduce the GPU-hours required for a task, although lower cost can also make more usage economical. New models may require more memory, bandwidth, interconnect or hardware-specific features and shift demand toward newer GPUs. Those changes can affect future hourly price, billed hours, useful output, refresh spending and residual value.
Cost per unit of useful output can supplement cost per billed hour:
Pre-tax economic cost per unit of useful output = (capital-recovery charge + cash operating costs + lifecycle capital expenditure) ÷ annual useful output
“Useful output” must be workload-specific. For token production, the denominator depends on model, precision, batch size, latency target, prompt-completion mix and software efficiency. Training and inference workloads should not be collapsed into one apparently comparable output measure.
Accounting life is a cross-check
Book depreciation affects reported expense, operating income, net income and book value. It may also affect GAAP-defined covenants. Tax depreciation, which can differ from book depreciation, affects cash taxes. Neither book life nor book value establishes future rental income, economic usefulness or sale proceeds.
The disclosed policies show that management estimates differ and can change. They are not fully comparable: Nebius explicitly describes GPUs as its main assets, while CoreWeave, Meta and Amazon disclose broader equipment categories that can include non-GPU assets.
| Company | Disclosed category | Useful life |
|---|---|---|
| CoreWeave | Technology equipment | 6 years |
| Nebius | Server and network equipment; its 2025 filing describes GPUs as its main assets | 4 → 5 years |
| Meta | Servers and network assets | 5–5.5 years |
| Amazon | Servers and networking equipment | 5–6 years |
Meta extended most server and network-asset lives to 5.5 years in 2025, reducing depreciation expense by $2.92B and increasing net income by $2.59B. CoreWeave extended computing equipment from five to six years in 2023; Nebius moved server and network equipment from four to five years in 2026; and Amazon shortened a subset of servers and networking equipment from six to five years in 2025. These disclosures are management-assumption cross-checks, not GPU recovery curves.
Follow the financed cohort
CoreWeave’s gross technology-equipment balance rose from $9.1B to $20.9B during 2025. The $11.8B increase equaled 56% of the year-end gross balance. Company-wide revenue and utilization can improve as newer GPUs enter the fleet even while an older pledged cohort deteriorates.
For each cohort, the lender should connect identity and age with realized revenue, billed usage, attributable cash cost, useful output, operating condition and relevant market prices. Cohort operating contributions can then be aggregated to the borrowing entity and reconciled with cash available for debt service, scheduled payments and outstanding principal.
Sources
Data and market observations through July 30, 2026; sources verified through August 10, 2026.
- CoreWeave 2025 10-K — useful-life policy and technology-equipment balance.
- Nebius 2025 20-F and Q1 2026 shareholder letter — four-year policy and 2026 change to five years.
- Meta 2025 10-K — useful-life estimate and reported financial effect.
- Amazon 2025 10-K — server and networking-equipment useful lives.
- OCC Comptroller’s Handbook: Leveraged Lending — primary and secondary repayment sources.
- S&P Global Ratings: Equipping Data Centers Through Securitization — GPU residual value and amortization characteristics.
General information from cited public sources, provided without warranty and not investment, legal, accounting, tax, appraisal, valuation, rating, or credit advice. Worked-example costs, returns and residual values are hypothetical and are not market estimates, forecasts or appraisals. Company and product names are used only for identification; no affiliation or endorsement is implied. Corrections: info@siliconlien.com.