Server Useful Life in Four 10-Ks: A $38 Billion Profit Question

Meta's decision to stretch most of its servers to a 5.5-year life cut its 2025 depreciation by $2.92 billion, according to its own 10-K. Amazon went the other way and added $1.4 billion. Server useful life is the single estimate that moves hyperscaler profit most quietly, and the four latest annual filings put it at five to six years. Rerun at four years, combined operating income falls by about $38 billion.

What the four latest 10-Ks say about server useful life

The short answer: Alphabet depreciates servers and network equipment over six years. Microsoft uses two to six, Amazon five to six and Meta five to 5.5. If all four used four years instead, their combined operating income would be about $38 billion lower, roughly 8.5% of the total.

Useful life is the number of years a company expects an asset to earn its keep. Under straight-line depreciation, the purchase price is divided by that number and charged against profit each year. Double the life and the annual charge halves, with no change to the business underneath.

Every figure in the table below comes from the latest annual report each company has filed with the SEC. For Alphabet, Amazon and Meta that is the 10-K for calendar 2025. For Microsoft it is the 10-K for the fiscal year ended 30 June 2026, filed on 29 July 2026.

Wording is quoted from each company's Note 1 or equivalent accounting policy note. Microsoft's range covers everything from short-lived network kit to its main server fleet. Its own fiscal 2023 disclosure puts the server fleet at six years.

Read the third column, not the second. The stated lives look similar, sitting within a year of each other. The direction of travel does not. Three companies lengthened. One shortened, and it is the one with the largest cloud business by revenue.

This post does not reargue whether any of these lives is right. That debate, including Michael Burry's $176 billion estimate, is covered in the analysis of GPU depreciation and AI earnings. The job here is narrower. Put the four filings side by side, date every change, and compute what each company would report under a shorter life.

Every server useful life change, dated

Useful life changes are disclosed as a change in accounting estimate. They apply going forward, never retroactively, so the prior years in a 10-K are not restated. That makes the date of each change essential context for any year-on-year comparison of operating margin.

Microsoft: two extensions, then silence

In July 2022, Microsoft "completed an assessment of the useful lives of our server and network equipment." It moved both from four years to six, according to its fiscal 2023 10-K. It credited software that made the fleet run more efficiently, plus advances in technology. It had already extended these lives once, in July 2020.

The company put the effect at about $3.7 billion of added fiscal 2023 operating income, based on the equipment it held on 30 June 2022. Its fiscal 2026 10-K still states "two to six years" and discloses no further change. Microsoft's chief executive has spoken publicly about not wanting to carry years of depreciation on one hardware generation. The accounting estimate has not moved.

Alphabet: the longest stated life, unchanged since 2023

In January 2023 Alphabet lengthened server lives from four years to six and certain network equipment from five years to six. Its 2023 10-K said the change cut 2023 depreciation by $3.9 billion and lifted net income by $3.0 billion. An earlier change, in 2021, had moved servers from three years to four, according to coverage of that year's results.

The 2025 10-K states that Alphabet depreciates servers and network equipment "generally over a period of six years." That is the most aggressive single number among the four. Alphabet also has the second-largest server base behind Microsoft.

Amazon: four changes in six years, and the only reversal

Amazon's 2025 10-K footnote records two changes directly. Servers moved from five to six years on 1 January 2024. Then, on 1 January 2025, a subset of servers and networking equipment moved back from six years to five. Earlier steps, from three to four years in 2020 and four to five in 2022, are documented in Hudson Labs' review of the company's filings.

Amazon cites "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning" as the reason. The effect was $1.4 billion of extra depreciation and amortization in 2025 and a $1.0 billion cut to net income, falling mainly on AWS. No other hyperscaler has put that sentence in a filing.

Meta: the latest extension, worth $2.92 billion

Meta's policy table listed servers and network assets at four to five years through its 2024 10-K. Coverage of its 2022 results reported a move first to 4.5 years and then to five within that year. In January 2025 it extended most servers and network assets to 5.5 years, effective immediately.

The 2025 10-K quantifies the effect precisely: depreciation down $2.92 billion, net income up $2.59 billion, or $1.00 per diluted share. That was roughly 3.5% of Meta's $83.3 billion of 2025 operating income, delivered by one change in estimate.

How the shorter-life math works for server depreciation

None of the four companies discloses what it would report at a different life. So the question in the title needs a calculation, and the calculation needs to be simple enough that you can check it.

The method has three steps. First, estimate the annual depreciation charge on servers and network equipment at the company's stated life. Second, scale it to a shorter life: a charge at life L becomes roughly the charge multiplied by L divided by the new life. Third, subtract the extra charge from reported operating income.

For step one, Meta is the only company that discloses its server depreciation directly: $13.36 billion in 2025. For the other three, the estimate is the average gross cost of servers and network equipment across the year, divided by the stated life. Those gross costs come from each property and equipment note.

Alphabet's 10-K states that about 60% of technical infrastructure is servers and network equipment. Microsoft's line includes software, which it depreciates over three years. Amazon's 5.5-year blend is an assumption, because its 10-K gives a five-to-six range without a split. Total P&E depreciation is each company's reported figure, shown as a sense check.

Each estimate sits below the company's total depreciation, as it must, since buildings and other equipment depreciate too. The server share ranges from about 62% at Amazon, which carries large warehouse and logistics assets, to about 85% at Microsoft.

The calibration check against Meta's own disclosure

The scaling rule can be tested against a real number. Meta's prior policy was four to five years, so take 4.5 as the old midpoint. Scaling its $13.36 billion charge from 5.5 years back to 4.5 gives about $2.97 billion of extra depreciation. Meta's 10-K reports the actual effect as $2.92 billion. The rule lands within 2%.

The gross-cost estimate is less precise. Applied to Meta, it gives $15.1 billion against the disclosed $13.36 billion, an overstatement of about 13%. That is why the Meta row uses the disclosed figure. The other three rows should be read as the upper end of a range.

Hyperscaler operating income at 5, 4 and 3-year server lives

Five years is the life Amazon chose for part of its fleet in 2025. Four years is where Microsoft and Alphabet stood before 2022 and 2023. Three years is close to the life sceptics argue matches the AI hardware cycle. It is shown as a ceiling, not a forecast.

Working: extra charge equals estimated server depreciation multiplied by (stated life divided by new life, minus 1). Microsoft's figures are fiscal year to June 2026, the others calendar 2025, so the combined row mixes periods. The 3-year column overstates most, because assets older than three years would already be fully depreciated.

The four-year column is the one worth carrying around. It reverses most of the extensions made since 2022 and is not an extreme assumption. Under it, about $37.9 billion of combined operating income disappears. That is more than twice Meta's entire annual server depreciation charge.

Why Amazon is most exposed to a shorter server life

Amazon has the thinnest operating income relative to its server base. Its $80.0 billion of 2025 operating income sits against $172.5 billion of servers and networking equipment at gross cost. Meta's ratio is closer to one to one, and Microsoft and Alphabet are well above it.

The AWS segment carries most of that exposure. Amazon's segment note shows AWS depreciation and amortization of $21.45 billion in 2025, up from $13.32 billion in 2024, against AWS operating income of $45.6 billion. A 61% jump in one year is what a capex surge looks like on the way through the income statement.

My read is that Amazon's 2025 reversal was the most credible accounting decision any of the four made that year. It cost real reported profit, it was explained in plain language, and it pointed at the asset class that matters. A company that shortens a life when lengthening was the industry norm is telling you what its utilisation data says.

There is a second reading, and it is fair. Amazon has the most to lose if the estimate later proves too long, because its margin is thinnest. Taking a smaller hit early is cheaper than a large impairment later. Prudence and self-interest point the same way here, which makes the decision easier to trust, not harder.

Where this server useful life estimate is weakest

The calculation is deliberately simple, which is both its value and its main weakness. Four problems are worth stating before anyone quotes the $38 billion.

The inputs are partly estimated

Only Meta discloses server depreciation directly. The other three rows rest on gross cost divided by life, a method that overstated Meta by about 13%. Microsoft's line includes software. Alphabet's server share is the company's own approximate 60%. Amazon's 5.5-year blend is my assumption. Treat the Alphabet, Microsoft and Amazon figures as the upper end of a band, perhaps 10% to 15% high.

The scaling rule assumes a young fleet

Multiplying the charge by the ratio of lives works when most assets are younger than the new life. That holds today, because server bases grew by 45% to 63% in a single year at all four companies. It breaks down as the life shortens. At three years, any asset already older than three would be fully depreciated, so the ceiling column overstates by more than the others.

A shorter life is not obviously the right answer

The cascade argument is real. Accelerators that age out of frontier training still run inference, then smaller models, then general compute. If that chain holds, six years can be an honest estimate. Nothing in this post proves the stated lives wrong. It measures how much rides on them, which is a different claim. The distinction matters to anyone tempted to call the gap "overstated profit."

Depreciation is also non-cash. Operating cash flow at all four companies is untouched by any of these scenarios. So are cash taxes, which follow separate tax schedules. An investor valuing these businesses on cash flow can ignore most of this. The question of how the build is funded in cash is the subject of the comparison of hyperscaler capex against operating cash flow.

What would move the server useful life number

The estimate will change, and the change will appear in a note most readers skip. Four signals would tell you which way, and in what order to read them.

The last row deserves attention now. Meta's construction in progress nearly doubled to $50.5 billion during 2025, and Alphabet's assets not yet in service reached $78.6 billion. None of that is depreciating yet. When it enters service, it does so under whatever life is in force. That is why the estimate matters more each year, not less. The full arithmetic of what that spend has to earn is laid out in the 2026 capex payback analysis.

I would also watch the gap between the reported operating margin and a margin restated at four years. It is a one-line calculation from public filings, and it shows the reported margin rests partly on an estimate. The cross-company margin comparison in the AI operating margin scorecard is more useful with that adjustment beside it.

Frequently asked questions

Where to start this week

Open one 10-K and find three numbers: the stated server life, the gross server cost, and operating income. Divide the average gross cost by the life, scale it to four years, and subtract the difference from operating income. It takes about 20 minutes, and you will never read a hyperscaler margin the same way again.

If you buy cloud capacity, ask your account team one question at your next renewal: what useful life does your pricing assume for the hardware I am renting? Most will not answer. The ones who do are telling you how long they expect today's price to hold.

Related reading