Skip to content

Build1 publisher3 min readPublished

Amazon moved a subset of its servers back to a five-year depreciation life

Amazon's filing blamed the pace of change in artificial intelligence and machine learning, and it priced the decision itself: about $0.7 billion off 2025 operating income, with another $0.6 billion of accelerated depreciation.

The Engineer · Build desk

Illustration accompanying Amazon moved a subset of its servers back to a five-year depreciation life

What happened

  • Amazon had raised the estimated useful life of some servers from five years to six, then in 2025 moved a subset back to five, citing technology moving faster, particularly in artificial intelligence and machine learning.
  • The same filing estimated the shorter life would reduce 2025 operating income by about $0.7 billion, with accelerated depreciation taking another $0.6 billion.
  • AWS reported $42.2 billion of revenue and $16.6 billion of operating income in the second quarter of 2026, and Amazon said its AI business inside AWS had passed a $25 billion annualized run-rate.
  • Amazon's trailing-twelve-month free cash flow was negative $7.6 billion, which the company attributed to a $66.1 billion increase in purchases of property and equipment for AI infrastructure.

Compiled by The EngineerSomething wrong?How this is made

Why it matters

  • cost A model that amortizes GPU and server capex over six years books about a fifth less depreciation per year than a five-year schedule on the same hardware, and whoever priced the workload absorbs the gap.
  • contradiction Amazon's filing ties shorter hardware life to the pace of AI while the lives its peers filed run six years and five and a half, so the number tracks management judgement and cannot be read as a measured failure curve.
  • decision Anyone comparing an owned fleet against rented capacity now has two defensible lives on the public record and has to choose one and show it.

Straight-line depreciation divides cost by life. At six years a server expenses about 16.7 percent of its cost annually; at five years it expenses 20 percent, a fifth more per year for the same box [1]. Shortening the life of equipment already in service does not restate the past. The remaining book value is spread over the shorter remaining life, so the charge rises from the change date forward. Equipment pulled out before its schedule ends leaves book value with nowhere to go, and it lands in one period. Amazon booked roughly $920 million in accelerated depreciation and charges tied to retiring equipment early [2].

Add the two disclosed pieces of the 2025 hit and you get about $1.3 billion of reported operating profit given up by revising an estimate downward [2].

Whether five years belongs in anyone else's model depends on what the fleet holds. Microsoft's fiscal 2026 capital expenditure came to about $145.3 billion, and roughly $91.1 billion of it, about 63 percent, went into short-lived assets led by processors, GPUs and network gear [9]. Run one year of those purchases through both lives: six years gives an annual charge near $15.2 billion, five years near $18.2 billion, a difference of roughly $3.0 billion a year [3]. That calculation assumes straight line, no salvage value, and every asset in service on day one. None of those hold for a real fleet, which is why the figure is a bound and not a forecast.

A useful life is a management estimate, and one extra year changes the profit printed on the page [12]. As the dev.to account has it, at a small company an accountant notes the change and moves on, while at the scale of one of the largest compute fleets a single year moves serious money [14]. The traffic had been going the other way: Alphabet raised server life from four years to six in January 2023 and certain network equipment from five to six, cutting depreciation expense by $3.9 billion and lifting net income by $3.0 billion [5], and Meta extended most of its servers to five and a half years in 2025, reducing depreciation by about $2.9 billion [6].

The dev.to review of the filings says they do not separate the money that adds new computing capacity from the money that buys back capacity a company already had [13]. Both land in the same capital expenditure line. Microsoft's cash outlay for property and equipment went from about $23.9 billion to $115.9 billion over four years, close to five times [10][4].

For five years to be the right input somewhere else, the hardware has to sit in the same class as the servers Amazon moved, and it has to be replaced for the reason Amazon gave in the filing: technology moving faster, "particularly in the area of artificial intelligence and machine learning" [1].

What to watch

  • Whether Amazon's later filings identify which equipment classes sit in the subset moved back to five years.
  • Whether Microsoft, Alphabet or Meta revise their extended lives down, and what each says the change costs.
  • Whether the roughly $700 billion of combined capital spending guidance converts into paid capex or gets revised [c11].
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories