The Depreciation Wall: Why GPU Useful Life Is the Wrong Argument

Effective 1 January 2025, Amazon and Meta moved important estimates of server and network useful life in opposite directions.

Amazon shortened the useful life of a subset of servers and networking equipment from six years to five, increasing 2025 depreciation and amortization expense by approximately $1.4 billion. Meta extended most server and network assets to 5.5 years, reducing 2025 depreciation expense by approximately $2.92 billion.

The same effective date. The same broad hardware market. Opposite accounting effects.

The easy response is to ask which company is right. The Depreciation Wall argues that the better question is what operating system would have to exist for each estimate to be realized.

Useful life is not a silicon fact

A server's physical ability to turn on does not determine its economic useful life. The useful-life estimate depends on whether the organization can continue to place the asset into productive workloads at acceptable economics.

A displaced frontier GPU may move to inference, fine-tuning, simulation, analytics, batch work, development, or another internal service. But the next workload does not appear automatically. It requires qualified demand, software compatibility, security acceptance, power, cooling, migration engineering, facilities capacity, and a contribution profile that remains preferable to the alternatives.

That makes useful life an output of the utilization system.

Cash arrives before depreciation

AI infrastructure capital is paid for or financed before most of its depreciation reaches the income statement. Five- and six-year schedules distribute a large installed cohort across future reporting periods. Changing the estimated life moves the timing. It does not remove the volume.

That is the depreciation wall: the point where the capital already installed begins arriving in earnings at scale while new investment, demand growth, migration throughput, and the retirement wave continue moving at different speeds.

The public debate tends to focus on the slope of the depreciation curve. The book focuses on the volume beneath it and the operating mechanism that must keep the asset productive through the delay.

The cascade is a capacity system

The lifecycle can be mapped as a value stream:

1. Ordered - committed capital not yet installed.
2. Installed - physically available but not yet accepted for production.
3. Frontier - serving the highest-value or most performance-sensitive workload.
4. Cascade - displaced from frontier service and being prepared for another workload.
5. Secondary - productively serving inference, batch, analytics, development, or another approved use.
6. Exit - sold, returned, retired, impaired, or recycled.

The critical constraint usually sits between frontier displacement and accepted secondary production. If hardware leaves the frontier faster than the next tier can absorb it, queues grow. The organization may still report a high average utilization rate while older cohorts wait for migration, qualification, or receiving demand.

Two ratios make the issue visible:

Absorption ratio = productive secondary capacity activated / capacity displaced during the period

Coverage ratio = qualified next-tier demand scheduled within the conversion window / capacity expected to be displaced

An absorption ratio below 1.0 creates backlog. A weak coverage ratio signals that the next retirement wave may arrive before a productive destination is ready.

Put lifecycle evidence inside the capital gate

Every new accelerator purchase creates two projects: deployment now and conversion or removal later.

A complete capital request should identify the constraint being removed, the unit of output being produced, the placed-in-service path, the expected frontier period, the planned secondary workloads, conversion capacity, support economics, exit route, and the evidence that would trigger a review of the useful-life assumption.

That is not an argument for slowing every investment. It is a method for preventing capital speed from outrunning the organization's ability to convert assets into productive capacity.

A thesis designed to be tested

The Depreciation Wall does not present a fixed forecast as a fact. Every substantive claim is graded as verified, reported, projected, or assumed. The source ledger identifies the principal filings and institutional materials. The companion quarterly scorecard records what changes.

Evidence that would strengthen the concern includes depreciation rising above prior expectations, accelerated retirement charges, longer migration queues, weakening previous-generation utilization, or secondary prices falling through a large retirement wave.

Evidence that would weaken it includes revenue and gross profit outrunning depreciation, stable cohort economics, productive absorption of displaced assets, and backlog converting to revenue and cash on schedule.

The central concern should be retired if sustained demand conversion, stable cohort economics, successful cascade absorption, and normalizing depreciation all appear together without material estimate shortening or accelerated charges.

That is the purpose of the operating model: to replace a static argument with evidence, governance, and decisions.

Read and watch

• Find The Depreciation Wall on Amazon
• Watch the 44-second overview on YouTube
• Download the quarterly scorecard

Apply the model

OpX Advisory Group works with leaders to connect capital decisions, lifecycle capacity, operating constraints, and Finance-validated value. To discuss an executive working session using The Depreciation Wall framework, visit https://www.opxadvisorygroup.com/contact.