GPU Depreciation and Useful Life: What Investors Need to Know

Most infrastructure assets depreciate on predictable schedules driven by physical wear. A building depreciates over 25-50 years. A diesel generator over 15-20 years.

An industrial cooling tower over 20-25 years. GPU depreciation is driven by technology obsolescence, not wear. A GPU does not wear out in the conventional sense;the silicon does not degrade meaningfully over 3-5 years of operation.

What degrades is its economic value, as newer, more capable GPUs arrive and reduce demand for older hardware. The A100;released in 2020, priced at approximately $10,000-$15,000 wholesale;was the dominant AI training chip through 2023. The H100;released in 2022, priced at $25,000-$35,000;largely displaced it for frontier training.

The B200;released in 2025;is displacing the H100 for leading-edge workloads. This generational displacement happens every 18-24 months. An operator holding A100 capacity in 2025 competes against H100 and B200 providers who offer 2-4x the performance per dollar. Customers with flexibility migrate to newer hardware. Revenue per GPU falls. The asset's value falls.

Secondary market prices for AI GPUs show rapid depreciation. H100 80GB SXM5 units that traded at $30,000-$35,000 in Q2 2023 (peak scarcity) were available at $12,000-$15,000 on secondary markets by Q1 2026 as B200 supply scaled.

A100 80GB units that were priced at $10,000-$12,000 wholesale in early 2023 traded at $3,500-$5,000 by Q1 2026. This represents 50-65% value decline over 3 years;significantly faster than the 5-7 year useful life used in most financial models.

The economic useful life of a GPU;the period over which it generates acceptable returns;is shorter than its physical useful life. An H100 will function for 7-10 years. But its ability to compete for new customer contracts at margin-positive pricing may be exhausted within 4-5 years as newer generations arrive.

Three depreciation approaches are used for GPU assets in project finance and operating models. Straight-line depreciation amortises the asset equally over its assumed useful life;the simplest approach, and the most likely to overstate asset value if the useful life assumption is too long.

Accelerated depreciation front-loads the cost recognition, matching faster economic obsolescence. A double-declining balance schedule on a 5-year asset depreciates 40% in year 1, 24% in year 2, 14.4% in year 3, and so on.

Economic useful life depreciation ties the depreciation schedule to the expected revenue-generating period rather than physical life. Most UK infrastructure funds use 3-5 year models, with 3 years as the primary revenue period and a residual value of 10-20% for secondary market proceeds. A 3-year schedule versus a 7-year schedule can change reported EBITDA by 15-25% on a GPU-intensive operation. For detailed depreciation modelling, DSCR analysis, and project finance model review, speak to our advisory team at disintermediate.global/services.

Residual value risk;the risk that a GPU is worth less than projected at the end of its primary deployment;is the most underappreciated risk in GPU infrastructure investment. Most financial models assume a 15-25% residual value after 3-5 years. Realised residual values depend on secondary market depth and the pace of technology advancement.

During supply-constrained markets (2023), residual values were artificially high;older GPUs traded at premium prices because buyers had no alternative. As B200 supply normalises in 2026, secondary market values for H100 and A100 assets are falling toward their intrinsic value. Mitigation strategies: structuring customer contracts with terms that match asset useful life, building secondary market relationships in advance, and designing data centre facilities for hardware refresh;standardised mounting, cooling capacity headroom, and network flexibility to accommodate next-generation interconnect standards.

GPU infrastructure investment requires explicit, conservative treatment of depreciation and residual value. Models that use straight-line depreciation over 7 years will show attractive-looking returns that do not survive contact with market reality.

Key parameters for a defensible model: 3-4 year primary revenue period, 15-20% residual value at the end of the primary period, annual performance-adjusted revenue decline of 5-15% in years 3-4 as newer generation competition intensifies, and explicit refresh capex assumption of 40-60% of original asset cost at the end of the primary period. These assumptions produce more conservative returns;typically 15-22% IRR on GPU infrastructure versus 25-35% that aggressive modelling suggests;but they survive institutional due diligence. For financing term sheets, DSCR models, and lender introductions, speak to our advisory team at disintermediate.global/services.

GPU depreciation is driven by technology obsolescence, not physical wear;H100 units lost 55-65% of peak-scarcity value within 3 years

Economic useful life (4-5 years) is shorter than physical useful life (7-10 years);models using 7-year straight-line depreciation overstate asset values

3-year primary revenue period with 15-20% residual value is a defensible base case; 7-year schedules do not survive institutional due diligence

Customer contract terms should match GPU asset lifecycle;3-year contracts on 3-year assets eliminate residual value risk

Refresh capex (40-60% of original cost at end of primary period) must be explicitly modelled;omitting it significantly overstates long-term returns