The Canary Turned
Three hyperscalers stretched their server lives to defer over $10B of annual depreciation. One turned around — and said why in a filing.
Depreciation schedules are where a capital-intensive industry keeps its honesty. They are also where it keeps its discretion — and in the AI build-out, that discretion is currently worth more than ten billion dollars a year of reported earnings.
The stretch
Between 2023 and 2025, three of the four largest hyperscalers extended the accounting lives of their server fleets. Microsoft moved to six years in fiscal 2023 — roughly $3.7 billion of favorable impact. Google moved servers from four to six years in 2023 — about $3.9 billion of depreciation deferred. Meta moved to five and a half years in January 2025 — about $2.9 billion.
Every one of those changes is legal, disclosed, and defensible in isolation. Each one also moves reported earnings up by moving expense recognition out — during the exact window in which those earnings justify the largest capital expenditure program in technology history. Longer lives, lower depreciation, higher margins, at precisely the moment the market is pricing those margins as evidence that the build-out pays for itself.
An attributed estimate puts the cumulative earnings flattery across the sector at roughly $176 billion through 2028. We treat that figure as what it is — attributed, not audited — but the direction is not in dispute. The filings state the per-company impacts themselves.
The turn
Then one of the four turned around.
Amazon — in its FY2024 10-K, filed February 2025 — shortened the useful life of a subset of its servers and networking equipment from six years to five, effective January 2025, citing the pace of AI and machine-learning development. The actual 2025 impact, reported in the FY2025 10-K: roughly $1.4 billion of additional depreciation, about $1.0 billion off net income. Separately, it recorded approximately $920 million of accelerated depreciation in Q4 2024 for AI-exposed equipment it retired early.
Read that sequence again. The company with the largest server fleet on earth looked at AI hardware and concluded the equipment is wearing out faster, not slower — took the earnings hit, and said why in a filed document. Its three peers, running the same hardware through the same workloads, are carrying that hardware at six years.
They cannot all be right.
Why a canary matters more than a chorus
One reversal is worth more, analytically, than three extensions — because it is the disclosure made against interest. A company that shortens asset lives volunteers lower earnings. It has no promotional reason to do so. That is what makes it a canary rather than a data point: it is the first crack in a consensus accounting posture, from the participant with the most information about how the hardware actually ages.
The question the canary poses to the other three is mechanical, not rhetorical: if AI-driven obsolescence is real enough for Amazon to book a billion-dollar hit, what happens to the sector’s reported earnings when the other three converge on the same conclusion? The attributed math above suggests the answer is measured in tens of billions — recognized late, all at once, in whatever quarter the auditors and the capex cycle force the issue.
We track this as one of six fragility indicators, updated as the filings land. The next test dates are already on the calendar: four hyperscalers report within seventy-two hours of each other at the end of July. We will read the useful-life notes the day they file, and publish what they say — whichever way it cuts.
The indicator, the sources, and the live readings: Teardown · the falsifiers: Falsifier Watch