GPU Depreciation & Obsolescence Risk

Historical depreciation curves: V100 (2018) to A100 (2021): 3-year gap, 5x performance improvement (V100 125 TFLOPS, A100 625 TFLOPS). V100 pricing collapsed from $8K-$10K peak (2019-2020) to $2K-$3K (2022) = 70-80% depreciation in 2-3 years. A100 (2021) to H100 (2023): 2-year gap, 3x performance improvement. A100 pricing peak $10K-$15K (2021-2022), current $4K-$7K (2024-2025) = 50-70% depreciation over 3-4 years despite strong AI demand. H100 (2023) to B200 (2026 estimated): 3-year gap, 2.5-3x projected improvement. H100 peak $35K-$40K (2023-2024 shortage pricing), current $25K-$30K (2024-2025 normalization), projected $10K-$15K (2027-2028 post-B200 ramp).

Supply dynamics during transitions: Pre-launch (6-12 months before): Current generation maintains strong pricing, buyers uncertain whether to wait, suppliers clear inventory to enterprise customers (long-term contracts preventing spot price exposure). Launch phase (0-6 months): New generation scarce (allocation to hyperscalers, limited production), current generation pricing stable to slight decline (10-20%), some buyers waiting accepting opportunity cost. Ramp phase (6-18 months): New generation production scaling, current generation pricing declines 30-50% as buyers shift, secondary market emerges (early adopters selling current-gen upgrading). Maturity phase (18+ months): New generation widely available, current generation relegated to commodity status (60-80% cumulative depreciation).

Software and ecosystem effects: Framework optimization: PyTorch, TensorFlow release optimized kernels for new architectures 3-6 months post-launch. Example: Hopper (H100) optimizations in PyTorch 2.0 (2023) improving performance 40-60% versus Ampere (A100) widening performance gap beyond raw hardware specs. Backward compatibility: Older GPUs continue functioning but miss optimizations—A100 running Hopper-optimized code 30-50% slower than potential creating opportunity cost. Developer abandonment: Cutting-edge research and production workloads migrate to latest hardware, community support and troubleshooting knowledge concentrates on current generation, older GPUs face documentation gaps and unoptimized libraries.

Alternative use case transitions: Training to inference cascade: GPUs deprecated for frontier model training (H100 for GPT-5-class models) cascade to inference serving (deploying GPT-4-class models). A100s still viable for inference through 2025-2026 despite training obsolescence. Pricing: Training workload commands $2-$4/hour (cutting-edge, price-insensitive), inference $1-$2/hour (cost-sensitive, commoditizing). Consumer and hobbyist markets: Enterprise GPUs (A100, H100) eventually sold to consumers for local AI/gaming. Pricing floor: $2K-$5K regardless of original enterprise value due to consumer willingness-to-pay limits. Cryptocurrency mining: Low-efficiency fallback use case. Mining profitability: $0.50-$2/day per GPU at 2024-2025 crypto prices and difficulty. Supports $500-$2K residual values (1-3 year payback for miners).

Straight-line depreciation (accounting): Method: Equal annual depreciation over useful life. 4-year life: 25% annually. Example: $40K H100 → $30K Year 1, $20K Year 2, $10K Year 3, $0 Year 4. Limitation: Doesn't reflect actual market value—depreciation frontloaded (35% Year 1) not linear (25% annually). Accelerated depreciation (tax): Method: Modified Accelerated Cost Recovery System (MACRS) for US tax purposes. 5-year property class for computer equipment: Year 1 20%, Year 2 32%, Year 3 19.2%, Year 4 11.52%. Tax benefits: Higher early deductions reduce taxable income during high-revenue years, cash flow benefit from deferred taxes.

Residual value modeling: Conservative approach: Assume 10-15% terminal value at Year 4-5 ($4K-$6K for $40K H100). Used in debt underwriting and investor projections. Moderate approach: 15-25% terminal value assuming secondary markets (inference, mining, consumer) maintain demand ($6K-$10K for $40K H100). Reflects historical outcomes for prior generations. Aggressive approach: 25-35% terminal value assuming extended useful life and strong inference demand ($10K-$14K for $40K H100). Rarely justified—prior generations consistently depreciated 70-85%.

Total cost of ownership analysis: Acquisition cost: $40K H100 GPU, $250K-$300K full server (8 GPUs + infrastructure). Annual operating costs: Power (700W per GPU × $0.08/kWh × 8,760 hours = $490 annually per GPU × 8 = $3,920), cooling (30-50% of power = $1,200-2,000), network/maintenance ($2K-$5K). Total OpEx: $7K-$11K annually per server. Revenue requirements: Must generate $70K-$90K annually (Year 1-3 average) to cover: Depreciation ($60K-$75K based on purchase price and residual value assumption), Operating costs ($7K-$11K), Target margin ($3K-$4K minimum profit). Implies 60-75% utilization at $2.50-$3/hour average blended rate.

Salvage value and disposition strategies: Planned obsolescence timeline: Deploy new GPUs Year 0, operate Years 1-3 for training (premium rates $2.50-$4/hour), Years 3-4 transition to inference (commodity rates $1.50-$2.50/hour), Year 4-5 sell secondary market or retire. Maximize revenue: $150K-$200K cumulative over 4-5 years from $250K-$300K initial investment. Residual sale: Bulk sales to secondary dealers ($8K-$15K for 8-GPU H100 server Year 4), consumer marketplaces (eBay, Craigslist individual GPU sales $2K-$4K each), cryptocurrency mining operations ($500-$1.5K per GPU if bulk sale). Tax loss harvesting: If residual value below book value, recognize tax loss offsetting operating income (sell for $10K when book value $15K = $5K deductible loss).

Fast payback deployment strategies: Target 12-18 month payback: Deploy GPUs immediately upon availability, charge premium rates during shortage ($3-$4/hour H100 in 2023-2024), aggressive capacity utilization (80-85%), minimal idle time. Reduces generation transition risk—if 75% of capital recovered Years 1-2, Year 3-4 depreciation less impactful. Avoid late-cycle purchases: Don't deploy current-generation GPUs 18-24 months after launch (e.g., buying A100s in 2023 when H100 available)—compressed payback window creates unprofitability. Wait for next generation even if forgoing 6-12 months revenue.

Diversified workload mix: Balance training and inference: Training commands premium pricing but vulnerable to obsolescence, inference lower margins but longer-lasting demand. 60% training, 40% inference mix reduces aggregate obsolescence risk. Geographic and customer diversification: Serving 50+ customers across multiple regions prevents single-customer dependency creating pricing leverage and reducing churn risk during transitions. Service tiers: Offer latest-generation (premium pricing), current-generation (standard pricing), prior-generation (discount pricing) serving customer segments with different willingness-to-pay.

Lease and financing structures: Short-term leases (2-3 years): Match loan term to useful life, aggressive amortization (30-40% annual principal reduction), terminal balloon 10-20% of original principal balanced by expected residual. Example: $250K server, $175K loan (70% LTV), 3-year term, $50K annual principal ($4.2K monthly) + interest, $25K balloon. Residual $30K-$50K covers balloon. Lease-to-own for customers: Offer customers 24-36 month leases with purchase option at fair market value (10-30% of original cost). Shifts obsolescence risk to customer if they exercise option, infrastructure provider retains depreciated asset if declined. Operating leases: Account as operating expense not capital asset—GPUs remain on lessor balance sheet, lessee avoids depreciation exposure but pays premium (20-40% higher all-in costs).

Hedging and insurance: Revenue protection insurance: Business interruption policies covering lost revenue if GPUs fail or become obsolete (technology clause). Premiums 2-5% of insured value, claims rare (technology exclusions common). Commodity hedging: Short futures on GPU resale values (if markets existed—currently hypothetical). Would allow locking residual value reducing uncertainty. Alternative: Options on NVIDIA stock (inverse correlation—stock up when new generation strong, GPU resale values down). Manufacturer buyback programs: Negotiate trade-in agreements with OEMs (Dell, HP, Supermicro) guaranteeing minimum buyback 20-30% of purchase price at Year 3-4. Costs 5-10% premium upfront but reduces downside risk.

Moore's Law continuation: Historical pattern: 2-3x performance improvement per generation, 2-3 year cycles. NVIDIA roadmap: Hopper (H100, 2023) → Blackwell (B200, 2026) → Next-gen (2028-2029). Projected continuation: Industry experts expect 2-3x improvements sustainable through 2030 as transistor scaling slows but architectural innovations (sparsity, mixed precision, chiplets) compensate. Implication: Depreciation curves likely persist—60-80% value loss over 4-5 years baseline assumption. Downside risk: Faster improvements (3-5x per generation) accelerate obsolescence. Upside: Slowing improvements (1.5-2x) extend useful life.

Specialized hardware fragmentation: Training-optimized GPUs (H100, B200): Maximum performance for frontier model development, premium pricing, rapid obsolescence (2-3 year useful life). Inference-optimized chips (Groq, Cerebras, AWS Inferentia): Cost-optimized, narrow use case, potentially longer useful life (4-6 years) as inference less performance-sensitive. Hybrid approaches: Future GPUs may target both workloads extending relevance but sacrificing peak performance creating modest depreciation (40-60% over 5 years versus 70-85% current). Impact on depreciation: Fragmentation could stabilize used GPU markets—retired training GPUs valuable for inference longer. Uncertainty: Depends on inference chip success—if Groq/Inferentia capture market, GPU depreciation accelerates.

Energy efficiency and sustainability: Power consumption trends: Each generation improving performance/watt 30-50% (H100 2x performance of A100 at 1.5x power = 33% better efficiency). Older GPUs face rising disadvantage—A100 50% less efficient than H100, V100 70% less efficient. Carbon pricing and regulations: Data centers facing carbon costs ($30-$100 per ton CO2), renewable energy mandates, efficiency requirements. Could force early retirement of inefficient hardware regardless of functional capability—regulatory obsolescence. Implications: Accelerates depreciation for power-hungry older generations (V100, A100), extends value retention for efficient newer chips (H100, B200).

Geopolitical and supply chain risks: Export controls: US restricting advanced GPU exports to China (H100, A100 above performance thresholds). Creates bifurcated markets—restricted GPUs maintain value in unrestricted countries, flood Chinese market at 30-50% discounts. China develops domestic alternatives (Huawei Ascend, Moore Threads) capturing local demand previously served by NVIDIA—if successful, reduces global GPU demand compressing values. Supply chain disruptions: TSMC fab concentration (Taiwan) creating single-point failure. Disruption could extend current-generation useful life (scarcity preserving values) or accelerate development of alternatives (Samsung, Intel foundries scaling GPU production). Tariffs and trade wars: Component tariffs increasing GPU costs 10-30%—newer generations more expensive, older generations relatively cheaper creating slower depreciation. Or import restrictions limiting supply stimulating domestic alternatives.