I Read 14 Big-Tech 10-Ks on Server Life. Amazon Cut to Five Years. Meta Went to 5.5. Same Day.

TL;DR: Effective January 1, 2025, Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five. Meta extended most of its servers and network assets to 5.5 years the same day. Both filings quantify it: Amazon booked $1.4 billion more depreciation and $1.0 billion less net income ($0.10 per diluted share), Meta $2.92 billion less depreciation and $2.59 billion more net income ($1.00 per diluted share). I read the property-and-equipment note in the most recent 10-K of 14 large AI infrastructure spenders on 2026-08-15. Five disclosed a change effective 2023 through 2025: three lengthened, one shortened, one gave no direction. Depreciation assumptions move reported earnings by billions, companies disclose it when they do, and they genuinely disagree about how long the hardware lasts.

What did the two filings actually say?

Amazon's FY2025 10-K, filed 2026-02-06:

Effective January 1, 2025 we changed our estimate of the useful lives of a subset of our servers and networking equipment from six years to five years. The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.

Meta's FY2025 10-K, filed 2026-01-29:

In January 2025, we completed an assessment of the useful lives of property and equipment, which resulted in an increase in the estimated useful lives of most servers and network assets to 5.5 years, effective January 1, 2025.

Both quantify it. Amazon: "an increase in depreciation and amortization expense of $1.4 billion and a reduction in net income of $1.0 billion, or $0.10 per basic share and $0.10 per diluted share, which primarily impacted our AWS segment." Meta: "a reduction in depreciation expense of $2.92 billion and an increase in net income of $2.59 billion, or $1.00 per diluted share."

What turns this from a disagreement into a round trip is Amazon's property-and-equipment footnote, which puts both moves in one sentence: "Effective January 1, 2024, we changed our estimate of the useful lives for our servers from five to six years, and effective January 1, 2025, we changed our estimate of the useful lives of a subset of our servers and networking equipment from six to five years." Five to six, then six to five, inside 24 months.

Why does a useful life change earnings at all?

A server costs what it costs the day you buy it. Depreciation only decides which years absorb that cost. Spread $6 billion over six years and you charge $1 billion a year against earnings. Over five years, $1.2 billion. The cash left the building either way, same day, same amount. Only reported profit moves.

A longer life lowers annual depreciation and raises reported earnings; a shorter life does the reverse. Neither touches free cash flow, which is why the capex slope is what the bubble argument actually runs on rather than the earnings line.

What did all 14 filings show?

I pulled each company's most recent 10-K from EDGAR and read the property-and-equipment note. All 14 retrieved; Microsoft, Oracle, Nvidia, Dell and Applied Digital are fiscal 2026 filers, the rest fiscal 2025. Line-item names differ, so each is reported as the filing words it.

Company	Line item as named in the filing	Stated life	Change effective 2023-2025	$ effect
Amazon	Servers and networking equipment	Five to six years	2024: 5 to 6. 2025: 6 to 5 (subset)	Yes, both
Meta	Servers and network assets	Five to 5.5 years	2025: up to 5.5 (most)	Yes
Alphabet	Servers and network equipment	Generally six years	Jan 2023: up to six	Yes, in its FY2023 10-K
Microsoft	Servers and network equipment	Two to six years	None	n/a
Oracle	Computer, network, machinery and equipment	Six years (footnote)	None	n/a
CoreWeave	Technology equipment	Six years	Jan 2023: 5 to 6	Yes
Equinix	Core systems	3 to 40 years	"certain revisions" in 2024 and 2025	No
IBM	Information technology equipment	1.5 to 6 years	None	n/a
Tesla	AI infrastructure (owned data centers)	5 to 30 years	None	n/a
Dell	Computer and other equipment	3 to 5 years	None	n/a
Applied Digital	Networking equipment, electrical equipment, software	3 to 5 years	None	n/a
Digital Realty	Machinery and equipment	7 to 15 years	None	n/a
Nvidia	No server line (general P&E)	Two to seven years	None	n/a
Apple	None	Not disclosed	None	n/a

Five of 14 use the word "servers" in the disclosure: Amazon, Meta, Alphabet, Microsoft, Oracle. Across those five, the stated life on hardware from the same procurement cycle runs from a low bound of two years to six. Five of 14 disclosed any change: three lengthened, one shortened, and Equinix said only that it "made certain revisions to these estimates" in 2024 and 2025, naming neither direction nor asset class. Three quantified a per-share effect. Apple discloses no useful lives at all, only that depreciation is "recognized on a straight-line basis."

Are Amazon and Meta actually comparable?

No, and the gaps run deeper than the headline invites. Amazon's change covers "a subset of our servers and networking equipment." Meta's covers "most servers and network assets." Neither sizes the subset or the share "most" represents. The bases differ too: Meta computed its effect on assets "placed in service as of December 31, 2024," while Amazon's includes equipment acquired during 2025. And "servers" means the whole fleet, not just accelerators, so this is no clean read on GPU life.

What survives is narrow and worth having: two companies buying from the same vendors in the same quarter told their auditors different things about how long the gear lasts, and both were signed off.

How often does anyone even discuss this?

I ran EDGAR full-text search across 10-Qs filed January 1 to August 15, 2026. "Change in useful life" returns 12 hits. "Changes in useful lives" returns 9, "change in the estimated useful lives" 7, "extended the useful lives" 3, "useful lives of our servers" 2. Phrase search undercounts, since wording varies and annual changes land in 10-Ks. Even so: a disclosure almost nobody makes and almost nobody reads.

What does this not prove?

It does not prove anyone did anything wrong. Changing a useful life is a change in accounting estimate: permitted, applied prospectively, disclosed, audited. Amazon moved its number in the direction that hurt its own earnings, the opposite of what an earnings-flattering story predicts. Nor does it prove any estimate wrong. Nobody has six years of operating history on this generation of accelerators, so every figure in that table is a forecast, and firms with different refresh cadences and resale markets should forecast differently.

The aggregate figure circulating under Michael Burry's name, an estimate of understated hyperscaler depreciation, is a model behind his paywall. I could not read it, so it is not used here and I am not repeating the number. The per-company figures quoted alongside it need no estimate: Meta's $2.9 billion is Meta's own filed number, and the $677 million Amazon figure is Amazon's nine-month disclosure through September 30, 2025, superseded by the annual $1.0 billion.

What should you watch next?

One line, in one place. In Meta's next 10-K, Note 1 under "Use of Estimates," watch whether "5.5 years" becomes six. Meta's servers and network assets depreciation was $13.36 billion in 2025 against $7.32 billion in 2023, so another half-year extension is worth billions. In Amazon's, watch whether "a subset" grows or five years spreads to the whole line.

An extension arriving in the same year capex accelerates is the signal, and the revenue side of that trade is just as loosely specified: Google's 3.2 quadrillion monthly tokens are worth anywhere from $0.96 billion to $115.2 billion a quarter at Google's own list prices. Not because it is improper, but because it would mean the depreciation line stopped tracking the hardware and started tracking the earnings target, visible only in the footnote. The price of running this stuff keeps going one direction; the accounting for it does not have to.