GPU Depreciation Curves
How fast do AI accelerators lose value? Depreciation modeling for H100, A100, H200, and next-gen Blackwell GPUs. Essential for residual projections and LTV covenants.
How fast do H100 GPUs depreciate?
NVIDIA H100 GPUs depreciate approximately 40-50% in the first year, with cumulative depreciation reaching 70-80% by year 3. Key factors include: (1) new generation launches (Blackwell), (2) supply/demand dynamics, (3) workload suitability. For comparison, A100s depreciated faster post-H100 launch (50-60% Y1). Underwriters should model aggressive depreciation scenarios for LTV covenants.
Key Data Points
- H100 Y1 Depreciation: 40-50%
- A100 Y1 Depreciation: 50-60% (post-Hopper)
- Cumulative 3-Year: 70-80%
- Key Driver: Next-gen launches
Depreciation by GPU Generation
NVIDIA H100 (Hopper)
Based on Q4 2025 market data
NVIDIA A100 (Ampere)
Accelerated post-Hopper launch
Understanding GPU Depreciation
GPU depreciation is fundamentally different from traditional IT equipment. Standard servers depreciate linearly over 3-5 years. AI accelerators follow a step-function pattern driven by generational launches.
The "Generation Cliff" Effect
When NVIDIA announces a new architecture, the previous generation experiences a valuation cliff. Examples:
- V100 → A100 (2020): V100 residuals dropped 40% within 6 months
- A100 → H100 (2022): A100 residuals dropped 35% within 9 months
- H100 → Blackwell (2025): Anticipated 30-40% cliff in late 2025
The "Blackwell Cliff" Risk
NVIDIA's B200 and GB200 GPUs are expected to reach volume production in late 2025. H100 residual values may drop 30-40% as buyers shift to the new generation. Underwriters should stress-test portfolios against this scenario.
Depreciation Factors
1. Generation
Current-gen GPUs hold value better. Previous-gen accelerates depreciation as workloads migrate to newer hardware.
2. Supply/Demand
Constrained supply (e.g., H100 in 2023) slows depreciation. Oversupply (e.g., A100 post-Hopper) accelerates it.
3. Workload Fit
Training workloads prefer latest-gen. Inference workloads can use older hardware longer, extending useful life.
Depreciation Schedules by Model
| GPU Model | Year 1 | Year 2 | Year 3 | Notes |
|---|---|---|---|---|
| H100 SXM | 55% | 38% | 25% | Current-gen premium |
| H100 PCIe | 52% | 35% | 22% | Slightly lower than SXM |
| H200 | 60% | 45% | 30% | Memory advantage extends life |
| A100 80GB | 45% | 28% | 15% | Previous-gen discount |
| MI300X | 50% | 35% | 20% | AMD ecosystem risk |
| B200 (Est.) | 65% | 50% | 35% | Projected, supply-constrained |
Implications for LTV Covenants
Given aggressive depreciation, lenders should:
- Start conservative: 50-60% LTV at origination, not 70-80%
- Build in revaluation triggers: Quarterly marks against GLRI index
- Model generation cliff: Stress-test for 30% value drop on new launches
- Shorter tenors: 24-36 months, not 60 months like traditional equipment
The 'GPU Valuation' Playbook
- Full depreciation curves for all major GPU models
- Generation cliff scenario analysis
- LTV covenant structuring recommendations
Need format proof first? View sample output
Explore More
GPU Valuation Tools
GPU Residual/LTV Calculator
Calculate GPU depreciation and residual values
Open Speed-to-Power WatchlistGPU Residual/LTV Calculator
Calculate GPU depreciation using physics-based decay models
Open Speed-to-Power Watchlist