GPU Depreciation Schedule: Building a Defensible Range for Cluster TCO
How do we determine a defensible GPU depreciation schedule? The short answer is a combination of:
- How a fleet is being used; and
- How it will be used in the future.
Let’s start with some data from hyperscalers. In 2025, Amazon reduced the useful life of a subset of its servers from 6 to 5 years. Meta, by contrast, lengthened the useful life of the majority of its servers and network assets to 5.5 years.
These differences suggest that depreciation schedules should be specific to the fleet being operated. When looking at the useful lives that have been made public for various operators, we see a range of 4-6 years.
What Useful Lives Do Hyperscalers and GPU Clouds Use?
Here is how three operators — Amazon, Meta, and Nebius — approach useful life.
According to Amazon's SEC filing, the company reduced the useful life of a subset of its servers and networking equipment from six years to five years, effective January 1, 2025. Amazon estimated that the change would reduce 2025 operating income by approximately $700M.
Separately, Amazon took a ~$920M charge in Q4 2024 for accelerated depreciation and related charges associated with the earlier retirement of certain servers and networking equipment. Aethir's GPU depreciation analysis also discusses Amazon's changing useful-life assumptions in the context of GPU depreciation.
Meta moved in the opposite direction. According to Meta's SEC filing, the company extended the useful life of the majority of its servers and network assets to 5.5 years, effective in 2025. Meta reported that the change reduced depreciation expense by approximately $2.92B in 2025.
For Nebius, Aravolta's GPU depreciation analysis reports an approximately 4-year useful-life assumption.
This data reflects accounting estimates for each operator’s hardware portfolio, not the actual economic lifespan of a given GPU.
Using a Value Cascade to Maximize the Economic Life of GPUs
The idea that GPUs can have long economic lives is related to the value cascade described in Introl's GPU depreciation and asset lifecycle analysis. There are three stages of the value cascade:
- Years 1-2 (Training): The GPU is used for training frontier models, which requires maximum performance from the latest generation.
- Years 3-4 (Production Inference): As the next generation of GPUs is used for frontier model training, the previous generation can be used for inference. Inference does not require the latest generation, so the performance of the previous generation can remain competitive.
- Years 5-6 (Batch/Analytics): As the cycle continues, older GPUs can still be used for batch processing, analytics, and other inference use cases where economics are the primary driver.
This value cascade depends on demand for older generations of GPUs and on those GPUs remaining economical to operate.
One way to reduce the risk of there not being demand in years 4-6 of a GPU’s life is to contract that demand through GPU offtake agreements.
How Operating Conditions Impact the Economic Life of GPUs
Importantly, the “depreciation curve” for GPUs can vary even between identical GPU models operated by different customers.
Aravolta looked at data from multiple customers, ranging in size from 0-50 MW, and found a 30-45% difference in effective depreciation curves across deployments of identical GPU models.
One of the cohorts it examined was tracking toward a useful economic life of 3.7 years, compared with the approximately 5.5 years used in its underwriting.
Some of the reasons for the difference included utilization, thermal stress, and workload. As a rough rule of thumb, Aravolta notes that every 10°C increase in operating temperature can halve the lifespan of electronic components.
This does not mean that relationship should be applied as a formula to GPU useful life. Actual hardware life depends on factors including component temperatures, whether those temperatures approach the hardware's operating limits, cooling design, workload intensity, and other operating conditions.
What Is the Terminal Residual Value of GPUs?
Looking at the secondary market for GPUs in 2026, Aethir's GPU depreciation analysis reports used H100 servers available for approximately $150k-$180k.
A new B300 server is approximately $500k. These are not the same generations of GPUs, so the comparison reflects the economics of buying used H100s versus new B300s rather than the depreciation of a single GPU over time.
Aethir estimates that terminal residual value could reach approximately 10-20% of original cost.
The actual resale value of a GPU will vary based on the generation of the GPU, the configuration of the server, the condition of the hardware, market demand, and when the hardware is being disposed of.
The residual-value risk will generally belong to either:
- The buyer in a direct ownership structure; or
- The lessor in a lease-based financing structure, depending on the terms of the contract.
What Are NVIDIA’s Release Dates and Prior-Tier Pricing?
One of the unique aspects of NVIDIA GPUs is the pace at which NVIDIA is releasing new flagship GPUs for the data center market.
NVIDIA announced the H100 in March 2022, the B200 in March 2024, and the B300 in March 2025.
This means the introduction of a newer GPU generation can occur before the end of the second year of the training portion of the value cascade.
The introduction of a new GPU does not automatically render the previous tier economically obsolete. That depends on factors including availability, performance for the specific workload, power efficiency, rental demand, and the overall operating cost of the hardware.
As of early 2026, on-demand H100 pricing has been as low as $1.25 per GPU-hour on a spot basis and averaged approximately $3.11 per GPU-hour, according to our rent vs. buy H100 GPUs analysis.
H100 GPUs have also seen contract renewals around pricing signed 2-3 years earlier, whereas B200 GPUs have reportedly seen spot pricing as high as $14 per GPU-hour amid shortages.
These pricing trends are important when estimating how long a GPU can continue generating economic value. They do not, however, independently determine its accounting useful life, economic life, or resale value.
How to Determine a Defensible Useful Life for TCO Modeling of a Specific Fleet
Operators should:
- Monitor the duty cycle and utilization of the hardware.
- Monitor the thermal distribution and excursions of the hardware to identify cohorts with potentially different reliability profiles.
- Understand the workloads being run on the GPUs — including training, inference, and batch processing — and what demand may exist for that hardware as newer GPU generations enter the market.
For TCO modeling, we would recommend testing a range between approximately 3.7 years — the cohort that Aravolta reported measuring — and 6 years, the highest useful life we have seen among the operators discussed above.
This should not be construed as a hard upper or lower limit for accounting standards, but rather as a reference range for modeling.
Operators running fleets at high utilization and thermal profiles, and that do not believe they will have demand for their GPUs outside of training, may want to assume a shorter economic useful life.
On the other hand, operators with reliable fleets that believe they will be able to lease GPUs for inference and/or batch-processing workloads at competitive rates during years 5 and 6 may want to assume a longer economic useful life.
Operators should also model the residual value of the GPUs based on the prices they believe they will obtain in the secondary market when they dispose of the hardware.
These factors, in addition to maintenance cost, power efficiency, replacement cost, and expected revenue from renting out the GPUs — whether by leasing the GPUs themselves or leasing out the compute — should all be included in an operator’s total cost of ownership model.
Depreciating GPUs for Tax Purposes vs. Their Useful Life
For operators in the US looking to depreciate their GPUs for tax purposes, there are a few things to keep in mind.
According to IRS Publication 946, for qualifying property acquired and placed in service after January 19, 2025, bonus depreciation was restored to 100%.
The Section 179 deduction limit for the 2025 tax year is $2.5 million, with the phase-out beginning when qualifying property placed in service during the tax year exceeds $4 million.
Operators should note that there are many other considerations that go into determining the appropriate depreciation method for their GPUs, and the above information is not intended to be comprehensive tax guidance.
Tax depreciation affects the timing of deductions and cash flow. It does not determine how long a GPU remains economically useful.
When determining the useful life of a GPU for total cost of ownership modeling purposes, operators should instead consider the fleet's utilization, operating conditions, workload demand, expected residual value, and other factors specific to their deployment.
For operators looking to finance their GPUs, Zettabyte offers a variety of financing options that can be tailored to the operator's requirements.
Depending on the structure of the financing, residual-value risk may be allocated to different parties, but the useful economic life of the GPUs remains an important factor under either structure.