Amazon Cut Server Life, Meta Extended It: SEC Filings Diverge

SEC filings show the ‘Big Tech is inflating AI profits’ story is not one industry-wide maneuver — it is a company-by-company pattern that runs in opposite directions, sometimes in the same month.

The depreciation fight isn’t one story — it’s at least three different ones

Since November 2025, one number has anchored almost every retelling of the "Big Tech is inflating AI profits" debate: $176 billion, the figure investor Michael Burry attached to hyperscalers’ server and GPU depreciation practices for 2026 through 2028. What gets lost in the repetition is that the underlying accounting moves are not uniform across companies, and in at least one case, not even uniform across time at the same company.

Amazon shortened the useful life of its servers in a filing effective January 1, 2025 — reversing a lengthening it had made just one year earlier — and disclosed a real, quantified hit to net income as a result. Meta lengthened its useful-life assumption the same month, adding billions back to reported income. Alphabet’s 2023 useful-life extension shows a third pattern: a forecast that undershot what actually happened once results were reported. None of this fits neatly into the "hyperscalers are gaming earnings" framing that dominates search interest, nor does it fully vindicate the companies. The filings show something more specific and more useful to investors: whether a given company’s current depreciation assumption is aging well depends on that company’s own revision history, not on what its peers are doing.

Amazon: extend in 2024, reverse in 2025

According to Amazon’s 10-Q filings, the company changed its estimate of useful life for servers from five to six years, effective January 1, 2024 — the standard "extend life, lower near-term depreciation" move that critics associate with the whole sector. Effective January 1, 2025, Amazon reversed course on a subset of that same equipment, shortening useful life for a portion of its servers and networking equipment from six years back to five.

That reversal was not cosmetic. Amazon’s Q3 2025 10-Q states that for the nine months ended September 30, 2025, the shorter useful life increased depreciation and amortization expense by $889 million and reduced net income by $677 million, or $0.06 per basic and diluted share, with the impact concentrated in the AWS segment. This is a disclosed, filed, GAAP earnings reduction — the opposite direction from what the "understated depreciation" critique generally describes.

Meta: the same month, the opposite direction

Meta’s 10-K for fiscal year 2025 describes a useful-life assessment completed in January 2025 — the same month Amazon’s reversal took effect — that increased the estimated useful lives of most servers and network assets to 5.5 years, effective January 1, 2025. Meta’s own disclosure ties this to an expected reduction of approximately $2.9 billion in full-year 2025 depreciation expense, based on servers and network assets in service as of December 31, 2024. Unlike Amazon’s move, this change added to reported income rather than subtracting from it.

Two companies, same fiscal month, same category of asset, opposite adjustments and opposite income effects. That divergence is the part of the record that gets skipped when the debate is framed as "hyperscalers" acting as a bloc.

Alphabet’s 2023 case: what happens when a forecast meets the actual result

Alphabet’s January 2023 change offers a third data point, and a different kind of lesson. The company extended server useful life from four to six years and certain network equipment from five to six years, effective for fiscal year 2023, and estimated at the time that the change would reduce depreciation expense by approximately $3.4 billion for the full year. The actual reported effect, once the year closed, was larger: a $3.9 billion reduction in depreciation expense and a $3.0 billion increase in net income.

That half-billion-dollar gap between Alphabet’s original estimate and its actual result is a reminder that these are forward-looking accounting judgments, not fixed facts — a company’s own forecast of the benefit can move once real usage data comes in, in either direction.

What Burry’s number is, and what it is not

Burry’s $176 billion figure, reported by CNBC in November 2025, is his own estimate of the cumulative industry-wide earnings overstatement from understated depreciation across 2026 through 2028. CNBC’s reporting notes he singled out Oracle and Meta specifically, projecting 27% and 21% profit overstatement respectively by 2028. Burry has not published the line-by-line methodology behind those percentages, and no SEC filing from Oracle or Meta corroborates a specific overstatement percentage for 2028 — a year that has not yet been reported. The figure should be read as an analyst’s projected estimate, not as a verified accounting result, and it sits on a different evidentiary footing than the dollar amounts Amazon, Meta, and Alphabet have already disclosed for periods that have already closed.

Catalyst versus durability

The short-term catalyst in this story is mechanical: a useful-life change is an accounting estimate revision that mechanically raises or lowers depreciation expense and, therefore, GAAP net income, in the period it takes effect. Amazon’s 2025 change lowered income; Meta’s 2025 change raised it. Both are disclosed, both are quantified, and both are already in the historical record.

The longer-running, unresolved question is different: whether the useful-life assumptions each company is currently using will continue to match how AI-era servers and GPUs actually perform and get retired. Nvidia’s GPU generations have shortened in release cadence, and if actual hardware replacement cycles run faster than a company’s assumed useful life, that mismatch would eventually surface as either an impairment charge or another estimate revision — not necessarily as the kind of coordinated understatement Burry describes, but as a company-specific catch-up. The filings reviewed here do not show which companies are more exposed to that catch-up risk; they only show that the direction and size of past revisions differ enough that a single industry-wide label does not fit the record.

Competing read

A reasonable competing interpretation is that Amazon’s 2025 shortening is itself evidence the system is working as designed: when a company’s own operating data suggests equipment wears out faster than previously assumed, GAAP requires an estimate change, and Amazon made one that reduced its own reported income. Under this reading, Meta’s longer 5.5-year assumption is not necessarily wrong — it may reflect a genuinely different hardware mix or maintenance practice — but it has not yet been tested by a reversal the way Amazon’s assumption has.

What remains unclear

None of the filings cited here disclose Oracle’s own useful-life methodology or a company-specific dollar impact matching Burry’s 27% figure, so that portion of his claim cannot be checked against a primary source in this article. It is also not yet known whether Amazon’s 2025 shortening reflects a one-time correction or the start of a multi-year pattern of further downward revisions as AI-specific hardware ages faster than earlier general-purpose server fleets did. Readers should treat any claim about 2026-2028 industry-wide dollar impact, including Burry’s, as an estimate rather than a filed result until those years are actually reported.

Next watchpoint

The next concrete test is Amazon’s and Meta’s third-quarter 2026 10-Q filings, expected around late October and early November 2026, which will show whether either company has revised its server or network-equipment useful-life estimate again and, if so, in which direction and by how much. A second point to track is whether Oracle’s next 10-K includes a comparable useful-life disclosure that can be checked against Burry’s specific percentage claim for that company.