The question this piece answers: how much of the AI profit story is a depreciation schedule? Capex hits the cash flow statement on the day it is paid; the income statement feels it over the "useful life" management assigns. That assignment is an estimate, it is disclosed in one footnote, and it has been moving in one direction for five years — until one company moved it back. Every number below comes from the companies' own SEC filings and press releases, cross-checked across multiple independent transcriptions on 2026-11-18. Where a figure is our own arithmetic it says so; where it rests on a single transcription, the piece says that too.
What is a change in useful life, mechanically?
Under US GAAP (ASC 250), revising a useful life is a change in accounting estimate: it applies prospectively to the remaining book value and to everything purchased afterward, and prior periods are never restated. The mechanics matter more than they look. If a server cohort has a five-year life and no residual value, annual depreciation is 20% of cost. Extend the total life to six years after two years of use, and the annual charge falls from $12B on a $60B cohort to $9B — a 25% cut in the cost of equipment already half-consumed, with no change to the business underneath. Shorten it to four years and the charge jumps to $18B.
That is why this footnote is the highest-leverage line in a hyperscaler's income statement: at the scale of hundreds of billions in cumulative infrastructure spend, moving the denominator by one year moves net income by billions — permanently, and silently, because nothing in the revenue line changes at all. It is also why the disclosed dollar impact always comes with a scope qualifier: companies state the effect "based on assets placed in service as of December 31" of the prior year, a base that excludes whatever gets bought during the year the change takes effect.
What exactly did each company change?
Four filings, four statements, all in the companies' own words. The stated lives in the current policy notes sit within a year of each other — Alphabet says "generally over a period of six years", Microsoft's policy note reads "two to six years" (a range covering short-lived network kit up to its main server fleet, which its own FY2023 disclosure puts at six), Amazon says "five to six years", Meta says "five to 5.5 years". The direction of travel, not the stated number, is what differs.
| Company | Change | Effective | Disclosed effect |
|---|---|---|---|
| Microsoft | Servers & network equipment 4→6 years | FY2023 (July 2022) | Est. +$3.7B FY23 operating income; net income +$3.0B, EPS +$0.40 |
| Alphabet | Servers 4→6 years; certain network equipment 5→6 | January 2023 | 2023 depreciation −$3.9B; net income +$3.0B, EPS +$0.24 |
| Amazon | Servers 5→6 years | January 2024 | 2024 D&A −$3.2B; net income +$2.5B, EPS +$0.23 |
| Meta | Most servers & network assets → 5.5 years | January 2025 | 2025 depreciation −$2.92B; net income +$2.59B, $1.00 per diluted share |
| Amazon (reversal) | Subset of servers & networking 6→5 years | January 2025 | 2025 D&A +$1.4B; net income −$1.0B, EPS −$0.10, primarily AWS |
Two reading notes before the analysis. First, the disclosure conventions are not identical: Microsoft and Amazon's 2024 change were disclosed as anticipated operating-income effects, Alphabet and Meta as realized depreciation reductions with net-income effects — compare like with like. Second, each effect is stated against the year-end asset base of the year before the change; that scope convention is the whole story of Amazon's guidance miss, which we get to below.
The pattern from 2020 through 2024 is unambiguous: every single disclosed change extended lives, exactly as AI capex accelerated. Microsoft had already gone 3→4 years in July 2020 (+$2.7B FY2021 operating income), Alphabet 3→4 in January 2021 (depreciation −$2.6B in 2021), Amazon 4→5 in January 2022 (D&A −$3.6B in 2022) — each extension deferring billions of depreciation into later years precisely when the depreciation base was compounding (per the compiled change timelines at Metro Colocation Advisory and Zan Digital; the pattern, not every line item, is the multi-source claim).
Why is Amazon's reversal the tell?
Because it is the only data point in the set that was not written by the marketing department of capex. In January 2025, Meta and Amazon completed assessments of essentially the same equipment class and reached opposite conclusions in the same month. Meta's 10-K states the reason implicitly through its effect — most servers and network assets extended to 5.5 years. Amazon's filing states its reason verbatim, in both its 2025 10-Q and its 2025 10-K: "The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning."
Read that sentence against every other company's explanation and the inversion is complete. Microsoft cited investments in software that increased operating efficiency. Alphabet cited the same efficiency logic. Everyone else looked at AI and concluded servers last longer; Amazon looked at AI and concluded servers wear out faster. And Amazon is not a peripheral player hedging a small position — it is the largest cloud provider by revenue, running the hardware itself, with the most to gain from a longer schedule if one were defensible. On this specific question, the bears' strongest witness is the bears' own industry.
The same filing paragraph contains a detail that keeps the story honest: in the very same change package, Amazon extended heavy fulfillment equipment from ten to thirteen years, guiding roughly +$0.9B of operating income (per Cayman Journal's reading of the disclosure — single-source transcription, treat accordingly). Two conclusions follow. One, useful-life changes are portfolio decisions, not confessions — companies net them. Two, Amazon's 2025 package as guided was net positive to operating income; the server reversal cost $0.7B guided, the warehouse equipment paid for it. What makes the server reversal notable is not its size but its direction, from the operator with the most at stake.
Why did Amazon miss its own guidance by 2×?
This is the most instructive number in the whole set, and it is mechanical, not sinister. In its 2024 10-K, Amazon guided the effect of the 6→5 change as "a decrease in 2025 operating income of approximately $0.7 billion" — computed, per the disclosure convention, on the assets that sat in "Property and equipment, net" as of December 31, 2024. The realized 2025 number, in the 2025 10-K: an increase in D&A of $1.4 billion, a reduction in net income of $1.0 billion. Double the guide.
The gap exists because a change in estimate applies prospectively to everything purchased during the year as well. Amazon bought servers at a record pace through 2025 — and every one of them entered service on the shortened five-year schedule. The lesson generalizes far past Amazon, and it cuts against anyone modeling this: any forward estimate of a useful-life reversion built off a year-end asset base will understate it, and will understate it badly when capex is doubling. The cost of honest depreciation scales with the capex ramp, not with the existing fleet.
The same filing period shows the ramp arriving in the segment numbers. AWS's depreciation and amortization rose from $13.3B in 2024 to $21.5B in 2025 — up 61% in a year AWS revenue grew about 20% (as computed from the segment disclosures in the 2025 10-K, per BloFin's and Signals' independent reads). That gap — cost growing three times faster than revenue — is the depreciation wave reaching the margin, before any further policy change.
What does the bear case actually claim?
In November 2025, investor Michael Burry publicly argued that the hyperscalers' accelerator fleets have an economic life closer to two to three years than the five-to-six-year schedules in use, and that industry depreciation is understated by roughly $176 billion over 2026–2028. The argument sharpened a debate that had been running in accounting circles all year (Taiwan's accounting profession press, among others, flagged the asymmetry between Meta's +$2.5B estimate change and its disclosure duty).
Three things about the Burry claim deserve precision, and we flag them because this forum's rule is attribution over amplification. One: the $176B figure is his estimate — the calculation has not been published in auditable form and is not verifiable from public filings; Silicon Analysts, whose filing-by-filing read we used above, states this explicitly. Two: the falsifiable core of the claim is not the number but the direction — nobody, on any side of the debate, argues that GPUs last longer than the schedules assume. Three: the market has already priced one data point of the bear case, because Amazon's own useful-life study — the only intra-year, per-share depreciation trail any operator has published — conceded the point for part of its fleet at a cash cost of $0.10 per share.
What the bear case gets right structurally, even if the number is unverifiable: a five- or six-year accounting life is a disclosed estimate, not a fact. It becomes a problem only if the economic life — how long the hardware earns a competitive return before newer silicon makes it uneconomic — turns out shorter. In that scenario the correction does not arrive as a policy footnote; it arrives as impairments, accelerated replacement capex, or margins that quietly stop covering the asset base. The Cayman Journal analysis adds the sharpest structural observation: the largest cohort of undepreciated AI equipment currently sits in construction-in-progress, where it depreciates not at all — the mix shift toward AI-specific hardware is invisible in the policy footnote by construction.
How should a reader price this footnote?
Three habits transfer to any AI-infra earnings analysis. First, read the estimate-change paragraph before the revenue paragraph. Useful lives are disclosed in the accounting-policies note and the MD&A "Use of Estimates" section; the dollar effect is quantified right there. A quarter's beat can be a schedule, not a market. Second, separate the base. A guided effect computed on a December 31 asset base and a realized effect including in-year purchases are not comparable — Amazon's 2× miss is the canonical example. Third, normalize the effect against the profit line it lands on. Meta's $2.92B depreciation reduction was roughly $1.00 of diluted EPS on an advertising machine with enormous operating income — visible, material, survivable. The same dollar effect at a company whose operating income depends entirely on data center equipment would be a different order of exposure.
And one uncomfortable synthesis for the bull case. From 2020 to 2024, every disclosed change extended lives as capex accelerated — which means the industry has been harvesting a growing accounting tailwind precisely during the years its cash spending exploded: the money left the balance sheet on day one but reached the income statement over years that kept getting longer. The tailwind is not fraud and not even aggressive — each change cited a defensible rationale, and software-defined efficiency gains are real. But it is arithmetic that runs in one direction, and January 2025 was the first month a major operator ran it the other way. When the single most sophisticated operator of the fleet in question files a sentence that says AI makes servers obsolete faster, that sentence is worth more than any model.
Comments