Server Useful Lives
This post discusses different views on useful lives of server equipment, a topic of certain accounting controversy in recent years.
Server useful-life estimates have received increased attention as entities expand their investment in GPU and data-center infrastructure. Rapid improvements in accelerator performance have prompted questions about whether five- or six-year depreciation periods remain supportable. Installed equipment may nevertheless continue to perform inference, testing, development, batch processing, and other productive workloads after a newer hardware generation becomes available.
Useful life is the period during which management expects an asset to contribute economic benefit to the reporting entity. The estimate depends on the expected use of the asset population within the entity’s operating environment.
Some useful resources that can inform this estimate include industry studies, peer company disclosures, and analyst research. They provide evidence about asset capabilities and the range of observed useful lives. However, management must evaluate that evidence against the characteristics and expected use of its own assets.
This article explains the depreciation principle governing useful life, evaluates the relevance of external evidence, provides an entity-specific assessment framework, and identifies the circumstances that require reassessment.
1. What principles govern depreciation?
Depreciation systematically allocates the depreciable amount of a long-lived asset over its useful life. PwC characterizes depreciation as “a process of allocation, not of valuation.”
The allocation period reflects the reporting entity’s expected use of the asset. Physical capability beyond that period does not extend useful life when management expects to retire, replace, or otherwise cease using the equipment earlier.
Useful life is the estimate of the period over which the reporting entity expects to derive economic benefit from the asset. The estimate should therefore reflect the expected retirement date of the relevant asset population rather than the maximum period during which the physical item of equipment could remain operational.
For servers, depreciation begins when the equipment has been delivered, installed, configured, connected to the required infrastructure, and made available for its intended use (when the asset is “placed in service”). Depreciation ceases when the asset is disposed of or otherwise reaches the point at which the applicable requirements end depreciation.
Because useful life reflects the reporting entity's expected use, an industry-wide estimate cannot determine the appropriate period without evidence regarding the entity’s own assets and operating plans.
2. What does the available external evidence indicate?
External evidence provides context regarding the range of potentially supportable useful lives and the factors that cause estimates to differ. In 2024, the American Society of Appraisers Machinery and Technical Specialties Committee estimated a normal useful life of five to eight years for data-processing equipment, including routers, servers, and storage assets [2]. The study observed that the slowing rate of improvement associated with Moore’s law may have modestly extended average server lives. It also found that useful lives differ among hyperscale, colocation, enterprise, and telecommunications data centers because those operating models have different asset-turnover requirements.
Several public companies have adjusted the estimated useful lives of certain categories of server hardware to approximately 5 or 6 years. Some investors and analysts consider these periods too long for GPU-intensive infrastructure and support useful lives of two or three years for specific assets [5], [6], [7], [8], [9].
Other analysts support useful lives closer to six years based on recent industry developments and available empirical evidence [3], [4].
However, none of these views were authoritative. Further, based on a review of a broader sample of public companies filing their financial statements with the SEC, we observed that useful lives beyond three years may be supportable for certain server categories as the appropriate period varies by asset type, workload profile, operating model, and replacement strategy. The estimate therefore requires an entity-specific assessment, and none of the positions taken by analysts is appropriate for the industry as a whole.
3. How should management determine the useful life?
Management should estimate useful life by determining how long the relevant asset population is expected to remain productive within the entity’s operating environment.
The assessment should address four categories of evidence.
a. Future operating plans
Long-term empirical evidence for modern AI accelerators remains limited because the market is relatively young. Earlier accelerator generations, including NVIDIA V100-class GPUs and early TPU deployments, nevertheless indicate that specialized compute assets can remain operational for multiple years when supported by appropriate architecture, interconnect capacity, cooling, maintenance, and workload allocation.
A server that is no longer efficient for frontier model training may remain suitable for:
inference;
training smaller models;
testing;
internal development; or
other workloads that do not require the latest hardware generation.
Workload migration can preserve economic utility after the equipment ceases to represent the highest-performing technology available. As such, management should identify the workloads the equipment is expected to perform throughout the proposed depreciation period when determining useful lives.
However, continued technical capability provides limited support when approved operating plans indicate that the asset will be retired earlier. As such, the next category of evidence is operational plans.
b. Operational environment
Management should evaluate whether the equipment is expected to remain operational within the entity’s environment.
Relevant evidence includes:
physical condition and expected reliability;
maintenance requirements and availability of replacement parts;
continued support;
network, storage, and interconnect compatibility;
power and cooling requirements; and
the existing data center infrastructure's capabilities to support continued use.
A loss of vendor support or infrastructure compatibility may shorten useful life even when the equipment remains physically operational. The analysis should therefore consider the full system in which the server operates rather than the processor in isolation.
c. Replacement economics
Retirement decisions depend on the overall economic impact of continued use versus replacement. A newer server may offer higher processing speed, greater memory capacity, improved energy efficiency, or better interconnect performance. Those improvements affect useful life when they change management’s expected replacement timing.
Management should compare the incremental benefit of replacement with:
procurement cost;
installation and configuration cost;
operating-cost differences;
cost of service disruption needed to install and configure servers;
cost of infrastructure modernization (related network, power, and cooling changes); and
available data-center capacity.
An existing server may remain productive when the incremental cost of continued use is lower than the cost and disruption associated with replacement. Power availability and data-center constraints may also affect timing. A technically superior replacement may require infrastructure that is unavailable or uneconomic to install during the proposed period.
The estimate should therefore reflect management’s approved replacement strategy and the economic conditions underlying that strategy. However, another important source of evidence we have not yet covered is the consistency between management’s expectations and past experience and future plans.
d. Past experience
Historical retention and retirement patterns provide evidence about expected use.
Management should consider:
how long the entity has retained comparable equipment;
the age at which similar assets were retired;
the reasons for prior disposals;
current capital-expenditure plans;
approved replacement schedules.
A history of disposing of similar equipment before the end of its estimated useful life may indicate that the current estimate is too long. Significant gains and losses from the disposal of servers strongly suggest that the useful life or salvaged value estimates used in calculating depreciation need recalibration.
Historical experience may become less relevant when the entity’s plans include significant changes in its operations. The best practice is to document the rationale of any known departure from prior experience, explaining why earlier retirement behavior no longer represents the expected use of the current asset population.
These four categories of evidence (future plans, operational environment, replacement economics, and past experience) should provide robust support for an expected retirement date.
Management should document the basis for the estimate to also explain why external studies or peer-company practices are relevant to the entity’s circumstances. A peer-company estimate provides limited support when the companies have materially different workloads, infrastructure, utilization patterns, or replacement strategies. Where management selects a useful life outside an observed industry range, the analysis should explain the entity-specific evidence supporting that result.
5. When should management reassess useful life?
Useful life is an accounting estimate. New information or changes in circumstances may alter management’s expected period of use between acquisition and disposition. Management should reassess the estimate when facts and circumstances indicate that the expected retirement date may have changed.
Relevant indicators include:
Sustained reductions in utilization, workload changes, or removal from planned uses;
Physical deterioration or loss of vendor support;
Expected modernization of infrastructure that is incompatible with assets;
Significant deviations of past experience from related estimates (i.e., actual retirements occurring materially earlier or later than expected).
Technological developments that materially change operating costs, replacement benefits, or the expected use of installed assets
A change in useful life affects depreciation prospectively. The remaining depreciable amount is allocated over the revised remaining useful life.
Management should also evaluate whether the same facts indicate impairment, abandonment, or disposal. A useful-life revision changes the future allocation period. It does not address circumstances in which an asset is no longer recoverable or is expected to be abandoned.
Conclusion
Useful life is primarily estimated based on entity-specific evidence. This is why management should consider many facts and circumstances to appropriately determine useful lives of servers. Management should update the estimate when changes in facts and circumstances alter the expected productive-use period.
Sources
“Depreciation” term definition as per FASB ASC 360-10-35-4
“Estimated Normal Useful Life Study [Ver. 1/24]” by ASA – Machinery & Technical Specialties Committee, 2024 [URL]
“GPU Obsolescence is Complicated” by Dave Friedman [here]
“Resetting GPU Depreciation — Why AI Factories Bend, But Don’t Break, Useful Life Assumptions” By David Vellante, “Breaking Analysis”, Issue 298, November 22, 2025 [URL]
“Big Tech’s Deteriorating Earnings Quality” by MBI Deepdives [here].
“Depreciation of GPUs: between useful lives and useful myths” by Olga Usvyatsky, Deep Quarry [here].
“Amazon’s AI Reality Check” by Stephen Clapham [here].
“CoreWeave and the Never-Ending GPU Depreciation — A Masterclass in Accounting Elasticity” by Kakashii [here].