Server Useful Life in Four 10-Ks: $33B to $41B of Profit Rests on the Estimate

Microsoft, Alphabet, Amazon and Meta all depreciated servers over 4 years as recently as 2021. Today their 10-Ks say 5 to 6. Put all four back on a 4-year server useful life and their combined operating income falls by roughly $33 billion to $41 billion, or 7% to 9%. That range is computed below from the filings themselves, with the working shown.

Last verified 25 October 2026 against the latest annual filing for each company: Microsoft fiscal 2026 (year ended 30 June 2026), and Alphabet, Amazon and Meta fiscal 2025 (years ended 31 December 2025).

What the four 10-Ks say about server useful life

A useful life is management's estimate of how long an asset will generate benefit. Straight-line depreciation divides the cost by that number. A $60,000 server costs $15,000 a year over 4 years and $10,000 a year over 6 years. Nothing about the server changes, only the line on the income statement.

Each of the four companies prints its estimate in Note 1 of its annual report. The wording differs, and the differences are informative. Here is what each latest 10-K actually says.

Microsoft: a range of 2 to 6 years

Microsoft's fiscal 2026 10-K, filed on 29 July 2026, lists servers and network equipment at "two to six years". It does not split the range by asset type. The last change came in July 2022, when Microsoft moved server and network equipment from 4 years to 6 years, effective for fiscal 2023.

That change was large. Microsoft's own note says it raised fiscal 2023 operating income by $3.7 billion, about 4.2% of the $88.5 billion Microsoft reported that year. Microsoft's depreciation expense has since reached $34.3 billion in fiscal 2026, up from $22.0 billion a year earlier.

Alphabet: 6 years, stated plainly

Alphabet's fiscal 2025 10-K says it depreciates "servers and network equipment generally over a period of six years." That wording is the cleanest of the four. Alphabet got there in two steps, in January 2021 and January 2023, starting from 3 years for servers.

The 2023 step was the bigger one. Alphabet's fiscal 2023 note says moving servers from 4 to 6 years cut that year's depreciation by $3.9 billion. Depreciation of property and equipment then rose to $21.1 billion in 2025, from $15.3 billion in 2024.

Amazon: 5 to 6 years, after a reversal

Amazon's fiscal 2025 10-K lists servers and networking equipment at "five to six years". The range exists because Amazon moved servers out to 6 years in January 2024, then pulled a subset back to 5 years in January 2025. It is the only company in this group to shorten a server life.

Amazon gave a reason in plain words. The shorter lives were "due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning". The change added $1.4 billion of depreciation in 2025 and mostly hit the AWS segment.

Meta: 5.5 years for most servers

Meta's fiscal 2025 10-K says it extended "most servers and network assets to 5.5 years, effective January 1, 2025". The effect was a $2.92 billion reduction in 2025 depreciation and a $2.59 billion lift to net income, or $1.00 per diluted share.

Meta is also the most useful filer for this analysis. It is the only one of the four that discloses depreciation on servers and network assets as its own line: $13.36 billion in 2025, against $18.0 billion of total depreciation. That single number is what makes the estimates later in this post checkable.

Every server useful life change since 2021, dated

The table below lists every change the four companies have quantified in a 10-K since January 2021. Each row comes from the Note 1 text of the filing named. Dollar figures are the first-year effect each company reported, not a running total.

Microsoft reports the effect on operating income; the other three report it on depreciation. For a straight-line charge the two are close to equal before tax. Amazon also recorded about $920 million of accelerated depreciation in Q4 2024 for servers it retired early, which is a separate item and not in this table.

The $19.4 billion of depreciation relief the filings disclose

Add up the first-year effects in the table and you get a figure none of the four companies publishes. The seven extensions total $20.8 billion of lower depreciation. Netting Amazon's 2025 reversal leaves $19.4 billion.

That number understates the lasting effect, and the reason matters. A first-year figure measures the change against the fleet that existed in that year. Each company has since roughly doubled its server base, so the same assumption is now applied to a far larger pool of cost.

Meta shows the scale-up clearly. Its gross servers and network assets rose from $68.4 billion at the end of 2024 to $98.0 billion at the end of 2025. Microsoft's server, network equipment and software line went from $132.8 billion to $215.9 billion in one fiscal year.

Operating income at a 4-year server useful life

The fairer question is not what each change was worth when it happened. It is what each company's latest operating income would look like if it still used the life it used in 2021. All four used 4 years for servers then, which makes 4 years a neutral benchmark rather than a hostile one.

The table also shows a 3-year case. That is closer to the 2 to 3 year economic life that critics such as Michael Burry have argued for, a debate covered in the analysis of the $176 billion GPU depreciation claim. This post does not take that side. It shows the number so you can see how fast the sensitivity grows.

Operating income from each company's income statement. Reductions are Zan Digital estimates, pre-tax, for one fiscal year, holding the asset base constant. Fiscal years do not align: Microsoft's ends 30 June 2026, the others 31 December 2025.

Amazon is the most exposed in percentage terms, at 10.8% to 12.2% of operating income. The reason is margin, not policy. Amazon carries a large server base against thinner consolidated operating income, so the same accounting change moves a smaller denominator.

Microsoft is the largest in dollars. Its server, network and software cost base is the biggest of the four, and its fiscal 2026 additions to property and equipment reached $115.9 billion. A shift of $12.8 billion to $14.5 billion is roughly a year of growth in its operating income, which rose $26.7 billion in fiscal 2026.

How the shorter-life figures were computed

None of the four companies publishes an "if we used 4 years" number, so this section shows every step. If you disagree with an input, you can swap it and rerun the arithmetic in a spreadsheet in ten minutes.

Meta: the anchor, built on disclosed server depreciation

Meta disclosed $13.36 billion of 2025 depreciation on servers and network assets at a 5.5-year life. Moving to 4 years multiplies that charge by 5.5 divided by 4, giving $18.37 billion. The difference is $5.01 billion, the low end of Meta's range.

The high end uses the method applied to the other companies: average gross server cost for the year, divided by the shorter life, minus the same cost divided by the stated life. For Meta that gives $5.67 billion. On Meta, where the true figure is known, that method runs about 13% high, because some fully depreciated assets stay in gross cost and new assets arrive part way through the year.

Microsoft and Amazon: gross cost, calibrated to Meta

Microsoft's average server, network equipment and software cost was $174.4 billion across fiscal 2026. Re-depreciating it over 4 years instead of 6 adds $14.5 billion. Applying Meta's 13% overstatement factor gives the low end, $12.8 billion. The line includes capitalised software, which pushes the true figure toward the low end.

Amazon's average servers and networking equipment cost was $142.8 billion in 2025. Amazon does not say how much sits at 5 years and how much at 6, so the calculation uses 5.5 as a blended life. That gives $9.7 billion, or $8.6 billion calibrated. A pure 6-year starting point would raise the figure, and a pure 5-year one would lower it.

Alphabet: bounded, because the filing allows nothing tighter

Alphabet reports a "technical infrastructure" category of $203.7 billion that, by its own definition, includes data centre land and buildings as well as servers. Servers cannot be separated out. So the estimate is a range between two bounds.

The low bound scales Alphabet's own 2023 disclosure. The 4-to-6-year change cut 2023 depreciation by $3.9 billion when total depreciation was $11.9 billion. Total depreciation reached $21.1 billion in 2025, so the same change now implies about $6.9 billion. The high bound assumes every dollar of the $21.1 billion is server depreciation at 6 years, which gives $10.6 billion. The truth sits between them.

Why Amazon's 2025 cut is the line to read

I would treat Amazon's 2025 note as the most informative sentence in all four filings. It is the only place where one of these companies wrote down, under audit, that AI hardware is ageing faster than the prior assumption allowed.

The timing makes it sharper. Amazon shortened lives for a subset of servers on 1 January 2025. Meta extended most of its servers on the same date. Two companies running similar fleets, with access to the same Nvidia roadmap, reached opposite conclusions on the same day.

That does not prove Amazon is right. Amazon runs a cloud business where customers rent specific instance types and can leave when a newer one launches. Meta runs its own workloads and can push older hardware down to less demanding jobs. Both reasons are plausible. The point is that useful life is a judgement about business model as much as about silicon.

Amazon also told you how much the honesty cost. The $1.4 billion of extra depreciation cut 2025 net income by $1.0 billion, roughly 0.1 percentage points of a company of that size. Being conservative was cheap for Amazon in 2025. My read is that it will look cheaper still if the other three have to follow later on a larger asset base.

Where this analysis is weakest

This post presents a computed number, so it owes you the reasons that number could be wrong. There are three that matter, and the first one may be decisive.

The longer lives may simply be correct

Servers do not stop earning when a newer chip launches. A training cluster can move to inference, then to smaller models, then to general compute. If that cascade holds, 6 years can be an honest estimate even while 18 months of frontier relevance is also true. Auditors at all four companies have accepted the current estimates, and each company says it relies on historical asset performance.

My view is that the cascade works for a fleet that grows slowly and breaks for one that doubles in a year. A company that adds this much new capacity has to find a lower-value job for every older chip it displaces, at the same time as it fills new halls. That is a demand assumption, and the filings do not test it.

The arithmetic is a one-year snapshot

The calculation holds the asset base constant and applies the shorter life to all of it. In reality a change would apply prospectively to remaining book value, and some older assets would already be fully depreciated under a 4-year life. The method runs high on Meta, which is why it is calibrated, but the calibration is borrowed from one company and applied to three. Treat the range as an order of magnitude with stated working, not a restatement.

It also ignores tax, buildings and leases. The reductions are pre-tax, and they cover owned servers only. Servers held under finance leases or inside third-party capacity contracts sit outside these figures. That gap matters for Microsoft in particular, as covered in the breakdown of hyperscaler lease commitments not yet on the balance sheet.

The next estimate to watch is buildings, not servers

The server debate may already be the old one. On its fiscal 2026 fourth-quarter call, Microsoft said it would extend the useful life of its data centres and office buildings from 15 to 25 years, effective at the start of fiscal 2027. That is according to CFO Dive's report of the 29 July 2026 call. The fiscal 2026 10-K, filed the same day, still lists buildings and improvements at 5 to 15 years.

Microsoft's chief financial officer Amy Hood described the operating income benefit for fiscal 2027 as "a minimal benefit". She said the larger effect is on reported capital expenditure, because more future data centre leases will be classed as operating leases. CFO Dive reported that this lowered Microsoft's calendar 2026 capex expectation to about $175 billion, from about $190 billion.

That second effect deserves more attention than it has had. A useful life estimate now moves not only reported profit but also the headline capex figure that investors use to track the AI build. The reported spend is examined against cash generation in the comparison of hyperscaler capex and operating cash flow.

Microsoft's buildings line was $182.7 billion at cost on 30 June 2026. The 10-K for fiscal 2027 will be the first to show the new life in its Note 1. I would read that note before I read the headline capex number.

Frequently asked questions

Where to start

If you model any of these four companies, add one row to the model this week. Take reported operating income and subtract the 4-year adjustment from the table above. Then see whether your valuation still holds at the adjusted number. If it only works at the reported figure, the estimate in Note 1 is carrying part of your thesis.

If you buy cloud capacity rather than shares, ask your provider one question at renewal. Which instance generations does it plan to retire in the next 24 months? A provider that has shortened its own hardware lives will retire generations sooner, and your pricing on older instances will move with it. The broader margin picture across these providers is set out in the AI profitability and operating margin scorecard, and the spending behind it in the analysis of 2026 capex and its missing payback math.

Related reading