GPU Useful Life

What is the expected useful life of AI accelerator hardware?

Answer

The expected useful life of AI accelerator hardware is 5 years (central estimate), with an optimistic bound of 6 years and a conservative bound of 4 years. This reflects the economic useful life -- the period over which the hardware generates sufficient value to justify its capital cost -- rather than the physical lifetime, which can exceed 10 years.

The industry is converging toward 5 years as the standard depreciation period. Hyperscalers extended server depreciation from 3-4 years to 6 years during 2020-2023 (Amazon led in January 2020, moving from 3 to 4 years, with all three major hyperscalers reaching 6 years by 2023). AI-native neoclouds, however, use shorter 4-5 year schedules, and industry analysis projects 5 years as the emerging equilibrium. Physical obsolescence is not the binding constraint; economic obsolescence driven by rapid generational performance improvements (3-4x per generation every 2 years) determines useful life.

Analysis

Why 5 years is the central estimate

1. The neocloud range brackets 5 years. CoreWeave (6 years), Lambda (5 years), and Nebius (4 years) center on 5 years. These companies have the most direct exposure to GPU economics and no legacy fleet to subsidize optimistic assumptions. Their revealed-preference choices carry strong weight because their business viability depends on accurate depreciation assumptions.

2. The value cascade supports 5 years. The three-stage model (training -> inference -> batch) maps naturally to a 5-year lifecycle with diminishing returns in years 5-6.

3. Hyperscaler-neocloud convergence points to 5 years. Hyperscalers extended to 6 years (2020-2023) while neoclouds use 4-5 years. Amazon's January 2025 reversal from 6 to 5 years — explicitly citing accelerating AI hardware development hyperscaler-depreciation-sec.1 — confirms this convergence. The current hyperscaler range is 5-6 years (Amazon 5, Meta 5.5, Google/Microsoft 6), and the neocloud range is 4-6 years. SiliconANGLE analysis and Dylan Patel both independently identify 5 years as the standard assumption.

4. NVIDIA's 2-year cadence creates natural breakpoints. With Hopper (2022) -> Blackwell (2024) -> Rubin (2026) -> next-gen (2028), each generation delivers 3-4x performance/watt. After two generations (4 years), older hardware is 9-16x less efficient per watt, making continued operation increasingly uneconomic except for latency-insensitive batch workloads.

Trend direction: slight shortening

The trend is toward slight shortening from the 6-year schedules adopted in 2022-2023:

• Amazon reversed from 6 to 5 years in January 2025, citing AI-driven technology acceleration hyperscaler-depreciation-sec.1
• Meta simultaneously extended to 5.5 years (from 5), booking $2.9B in savings hyperscaler-depreciation-sec.2 — demonstrating that the "correct" answer is company-specific
• SiliconANGLE projects 5 years as the "emerging equilibrium"
• The pace of GPU performance improvement (3-4x per generation every 2 years) has not slowed
• AI-native companies use 4-5 years, and their practices tend to lead hyperscaler policy
• Dylan Patel uses 5 years as his standard assumption and dismisses shorter (3-year) depreciation as overly bearish

However, a countervailing force exists: if chip manufacturing becomes the binding constraint (ASML EUV production limited to ~100 tools/year by 2030), older GPUs could retain economic value longer, potentially stabilizing or even extending useful life assumptions.

Implications for orbital economics

The 5-year useful life creates a hard constraint for orbital data centers:

1. Hardware must generate returns within 5 years. Any time spent on ground testing, launch, and orbital commissioning (estimated 3-6 months by Dylan Patel) reduces the productive window by 5-10%.

2. No mid-life upgrades. Terrestrial data centers can swap individual GPUs (15% RMA rate for Blackwell). Orbital systems must either over-provision for failures or accept degrading capacity.

3. No second-life cascade. Terrestrial GPUs can be redeployed from training to inference to batch workloads. Orbital GPUs are locked into their initial deployment configuration.

4. End-of-life disposal. Terrestrial hardware has residual value; orbital hardware must be deorbited, with the cost of the launch amortized over fewer productive years if hardware fails early.

Evidence

Hyperscaler depreciation schedules

All changes below are documented in SEC filings (10-K/10-Q) and earnings call transcripts. The SEC filing references are provided in the source registry under hyperscaler-depreciation-sec.

Amazon reversed server useful life from 6 to 5 years effective January 2025, citing "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." This cost Amazon $700M in operating income plus $600M in accelerated depreciation for early-retired AI-specific servers. — Amazon FY2024 10-K, Q4 2024 earnings call

As of January 2025, hyperscaler server depreciation schedules span 5–6 years: Amazon 5 years (reversed from 6), Google 6 years, Microsoft 6 years, Meta 5.5 years. The combined financial impact of all depreciation extensions since 2020 exceeded $15B in cumulative operating income benefit, demonstrating these are not minor accounting choices. — SEC 10-K/10-Q filings, 2020-2025

AWS/Google/Microsoft: 6-year depreciation. Industry converging toward 5-year via "value cascade" model. AI-native neoclouds use 4-5 year schedules.

Amazon led the industry shift to longer depreciation in January 2020, moving from 3-year to 4-year server schedules; all three major hyperscalers had extended to 6 years by 2023. However, AI-native neoclouds use shorter 4-5 year schedules, and the SiliconANGLE analysis projects 5 years as the emerging equilibrium — a convergence between the hyperscaler 6-year schedules and neocloud 4-5 year practice.

Neocloud depreciation schedules

AI-first clouds "cannot afford stagnant infrastructure; performance/watt gains in successive GPU generations directly determine competitiveness." Predicts 5 years as the emerging equilibrium.

The "value cascade" model

GPUs follow a three-stage lifecycle that supports extended useful lives:

1. Years 1-2: Frontier training. Peak performance required. Hardware is used for training foundation models where latest-generation compute provides the strongest competitive advantage.

2. Years 3-4: Production inference. Previous-generation GPUs move to high-value real-time serving. Performance remains adequate; latency requirements are less stringent than training synchronization demands.

3. Years 5-6: Batch inference and analytics. Final lifecycle stage supports cost-sensitive, latency-tolerant workloads where the hardware still generates positive economic returns.

The "value cascade" framework is the primary justification for 5-6 year depreciation schedules.

Dylan Patel / SemiAnalysis perspective on GPU economics

H100 all-in deployment cost is ~$1.40/hour across 5 years. At $2/hour market rate, yields ~35% gross margin.

Every 2 years, NVIDIA triples/quadruples performance while increasing price by 50-100%. This compresses the market value of older GPUs.

H100 market rate fell from ~$2/hour (2024) to ~$1/hour (2026) as Blackwell deployed at volume.

"If your argument is that a GPU has a useful life of five years" -- Patel uses 5 years as the standard assumption.

Michael Burry argued for 3-year or shorter depreciation, but Patel notes this is overly bearish.

Counter-argument: if compute demand outstrips chip manufacturing capacity (constrained by ASML EUV tool production at ~100/year by 2030), older GPUs retain value longer. "Maybe the depreciation cycle is even longer than five years."

Physical vs. economic lifetime

11 months of data from 24K A100 GPUs at >80% utilization. Component MTTF data validates that individual GPU physical failure rates are low, but at scale, failures become a daily occurrence.

ChinaTalk analysis

An aggressive lower-bound view used in some cost models is a three-year hardware life for frontier deployments, but this page treats that as a conservative counterpoint rather than the base case because the stronger source-backed industry practice still clusters around 4-6 years.

Inference vs. training lifecycle differences

Training hardware: Frontier training demands latest-generation hardware for competitive advantage. Economic useful life for training-only is effectively 2-3 years before next-gen hardware provides compelling cost-per-FLOP improvements.

Inference hardware: Inference workloads are less sensitive to generational upgrades. Older hardware remains competitive for inference longer because:

• Inference is often memory-bandwidth-bound, not compute-bound
• Latency requirements are application-dependent and often tolerant
• Optimized inference software (TensorRT, vLLM) continues improving on older hardware
• The value cascade model naturally extends GPU life through inference workloads