StackedAI models GPU residual value at about 15% of cost at the end of a five-year financing term (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026). Accounting schedules in the market run from about two years to about six, but the number that matters is operational: H100 rental rates fell from above $8 per GPU-hour in 2023 to under $3 (StackedAI analysis, Stranded Asset Report, Aug 2026).
This supporting page to the stranded GPU assets pillar separates the two meanings of depreciation and explains why form factor drives resale.
What is the difference between financial and operational depreciation?
Financial depreciation is a bookkeeping choice: the owner picks a useful life, spreads the cost over it, and books the expense. It is set once. Operational depreciation is what the market does to the asset: the decline in the hourly rate a GPU can earn as newer generations arrive. It is set continuously, by other people.
The gap is where neocloud balance sheets get into trouble: an operator booking a long life shows healthy margins while hourly revenue falls, but its lender cares only whether the hourly rate covers debt service. In StackedAI’s model only flat-to-rising $/GPU-hour pricing keeps a GPU-only structure at break-even (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026).
What depreciation schedule does the market use?
There is no single schedule. Useful lives range from about two years at the aggressive end to about six years at the hyperscaler end, where servers are assumed to be redeployed to lower-value workloads as they age (StackedAI analysis, Stranded Asset Report, Aug 2026). Figures vary by operator; StackedAI benchmarks them per engagement.
The schedule changes reported earnings, not cash, so neocloud margins are not comparable without normalizing it. A six-year life holds only if the owner can redeploy the hardware to a paying workload in years four to six: a hyperscaler with internal inference demand can; a neocloud renting at spot rates cannot.
What does the pricing evidence say about secondary value?
Rental pricing is the best public proxy for operational value: a GPU’s resale price is the present value of what it can earn.
| Price point | Figure | Source |
|---|---|---|
| H100 hourly rate, 2023 | Above $8 per GPU-hour | StackedAI analysis, Stranded Asset Report, Aug 2026 |
| H100 hourly rate, open market 2026 | Under $3 per GPU-hour | StackedAI analysis, Stranded Asset Report, Aug 2026 |
| H100 one-year contract, Oct 2025 low | $1.70 per GPU-hour | Akash Network, H100 Rental Price 2026, Aug 2026 |
| H100 one-year contract, Mar 2026 | $2.35 per GPU-hour, roughly +40% | Akash Network, H100 Rental Price 2026, Aug 2026 |
| H100 on-demand, dedicated GPU clouds | Median $4.19 per GPU-hour, +4.3% over 12 months | GetDeploying, GPU Price Index, Aug 2026 |
| B200 hourly rate | $3.70 to $7 per GPU-hour | StackedAI analysis, Stranded Asset Report, Aug 2026 |
The Hopper generation lost more than half its hourly rate within about three years, faster than any five- or six-year schedule. For a lender, a three-year-old H100 fleet is worth a discount to the present value of sub-$3 hourly revenue over its remaining life, less the cost of moving it, while new Blackwell capacity rents at $3.70 to $7. That is why StackedAI’s base-case residual is about 15% rather than the figure a six-year book life implies.
Why do liquid-cooled rack-scale systems resell worse than air-cooled PCIe?
Form factor determines the buyer pool, and the buyer pool determines the discount.
An air-cooled PCIe server fits almost any hall built in the last decade: legacy halls run 5 to 15 kW per rack and air cooling tops out at roughly 30 to 50 kW per rack (StackedAI analysis, Tier-2 conversion thesis, Aug 2026), so its buyer pool is wide.
A GB200 NVL72 rack draws 120 to 130 kW (StackedAI analysis, density roadmap compilation, Aug 2026). It is liquid-cooled as a system, cannot be split without losing the fabric that makes it valuable, and needs facility liquid cooling costing $1.5 to 1.6M per MW, 20 to 25% above air-cooled fit-out (Data Center Knowledge, Apr 2026); the liquid cooling retrofit cost page covers what that involves.
The liquid-cooled share of AI servers rose from 15% in 2024 to 54% in 2025 and is projected at about 76% in 2026 (StackedAI analysis, density roadmap compilation, Aug 2026, from Goldman Sachs estimates), so today’s fleets are the kind that resell worst. Their buyer must own a compatible hall or buy it with the racks, which is why the what happens when a neocloud fails page treats the lease and power position as the prize.
How does StackedAI treat GPU residual value in underwriting?
StackedAI holds GPU exposure as debt or preferred, never as common equity, and owns the building and the power position instead. In its model, the GPU-only operator in leased colocation returns 0.30x MOIC and a negative 28.8% IRR, while the building owner on an NNN lease returns 6.30x MOIC and 26.3% IRR on 28% of the equity with zero technology risk (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026). Three rules follow:
- The residual is a floor, not a forecast. About 15% at year five, tested lower.
- GPU exposure sits senior. Any compute participation is a secured loan or preferred with a take-or-pay floor.
- The lease must survive the tenant. Re-letting at 1% vacancy (JLL, North America Data Center Report Midyear 2026, Aug 2026) is a far better bet than resale of the tenant’s hardware; the NNN lease scarcity pillar explains why.
In short: own the building, not the silicon.
Key terms
- Financial depreciation: the accounting allocation of an asset’s cost over a chosen useful life.
- Operational depreciation: the decline in an asset’s earning capacity as it ages and newer alternatives arrive.
- Residual value: the estimated value of an asset at the end of its financing term.
- Rack-scale system: a GPU deployment sold as a complete rack with its own fabric and cooling, such as GB200 NVL72.
How StackedAI applies this
StackedAI normalizes neocloud financials for depreciation schedule before comparing them, and sets residual assumptions from rental-pricing evidence rather than book life. In sponsor engagements it structures any compute participation as debt or preferred and keeps the sponsor’s equity in the building. In lender engagements it assesses form factor first; that feeds the generation-mix dimension of the neocloud distress screening framework.
Frequently asked questions
What is the difference between financial and operational depreciation of a GPU?
Financial depreciation is the accounting schedule used to expense a GPU over a chosen useful life. Operational depreciation is the decline in what it can actually earn per hour. H100 rates fell from above $8 in 2023 to under $3 (StackedAI analysis, Stranded Asset Report, Aug 2026), faster than most accounting schedules.
How long a depreciation life do GPU operators use?
Schedules range from about two years at the aggressive end to about six at the hyperscaler end (StackedAI analysis, Stranded Asset Report, Aug 2026). A longer life raises reported earnings, not cash. StackedAI’s underwriting uses the five-year financing term and rental pricing instead.
What residual value should a lender assume on a GPU fleet?
StackedAI’s base case models residual at about 15% of cost at the end of a five-year financing term, against an advance of about 60% at roughly 8% (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026). Figures vary by generation and form factor; StackedAI benchmarks them per engagement.
Why is a liquid-cooled B200 rack worth less on resale than an air-cooled H100 server?
Far fewer buyers can accept it. A GB200 NVL72 rack draws 120 to 130 kW (StackedAI analysis, density roadmap compilation, Aug 2026) and needs facility liquid cooling costing $1.5 to 1.6M per MW (Data Center Knowledge, Apr 2026). An air-cooled PCIe server fits almost any hall.
How does StackedAI treat GPU exposure in a deal?
As debt or preferred, never common equity. In StackedAI’s model the GPU-only operator returns 0.30x MOIC while the building owner on an NNN lease returns 6.30x on 28% of the equity (StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026). Own the building, not the silicon.
Sources
- StackedAI analysis, AI Compute Partnership Returns Model, Jul 2026 (internal, illustrative)
- StackedAI analysis, Stranded Asset Report, Aug 2026 (internal)
- StackedAI analysis, density roadmap compilation, Aug 2026 (internal, compiled from Goldman Sachs and Dell’Oro estimates)
- StackedAI analysis, Tier-2 conversion thesis (cooling engineering notes), Aug 2026 (internal)
- GetDeploying, GPU Price Index, Aug 2026, https://getdeploying.com/gpu-price-index
- Akash Network, H100 Rental Price 2026: Cost Per Hour, Aug 2026, https://akash.network/the-bid/h100-rental-price-2026-cost-per-hour/
- Data Center Knowledge, Neocloud Storm Gathers as Data Center Deals Stall Over Credit Risk, Apr 2026, https://www.datacenterknowledge.com/cloud/neocloud-storm-gathers-as-data-center-deals-stall-over-credit-risk
- JLL, North America Data Center Report Midyear 2026, Aug 2026, https://www.jll.com/en-us/insights/market-dynamics/north-america-data-centers