On 29 July 2026 Microsoft filed its annual report. In the accounting-policies note, the company states that its property is depreciated over lives that are βgenerally as follows: software developed or acquired for internal use, three years; servers and network equipment, two to six years; buildings and improvements, five to 15 years β¦β (FY2026 10-K, filed 29 July 2026.)
That same afternoon, on the earnings call, CFO Amy Hood said this:
Now, before I move to outlook, effective at the start of FY27, we are extending the estimated useful lives of our datacenters and office buildings, from 15 to 25 years, reflecting our operating history and expected use of these assets.
(Microsoftβs own FY26 Q4 transcript, 29 July 2026.)
Nothing is wrong here. A change in accounting estimate is applied prospectively, and Microsoftβs fiscal 2027 began the day after the period the 10-K covers. The filing describes the year that ended; the call described the year that started. Both numbers are correct, on the same day, about the same buildings.
The useful life of a building is not a property of the building. It is a forecast, made by management, and it sets the size of one of the largest expenses in corporate accounting. Over the last three years the hyperscalers have revised theirs repeatedly, in both directions, and in one case for two asset classes at once in the same paragraph.
Hood kept going, and the next two sentences matter more than the change itself:
This change affects only the timing of future depreciation and is expected to have a minimal benefit to FY27 operating income. The greater impact is on capital expenditures as more of our future datacenter leases will shift from finance leases to operating leases as a result of this update. Finance leases are included in capital expenditures while operating leases are not.
So the earnings effect is βminimal,β by the companyβs own account. The effect that mattered was on the capex headline: lengthening the assumed life of a building changes the lease-classification test, finance leases sit inside reported capital expenditures and operating leases do not, and Microsoftβs calendar-2026 capex expectation therefore moved to βapproximately $175 billionβ with, as Hood put it, investment expectations otherwise unchanged.
The number the market watches most closely moved because of an assumption about how long concrete lasts. No steel was poured differently.
There is a second thing in those prepared remarks, delivered a few minutes earlier, that deserves to be read next to the first:
Capital expenditures were $41 billion including the impact from higher component pricing as noted in our guide. Roughly two thirds of our capex was for short-lived assets, primarily CPUs and GPUs as customers increasingly build solutions that leverage both AI and non-AI infrastructure. The remaining spend was for long-lived assets.
Two-thirds of the quarterβs spending went into equipment that Microsoft depreciates over two to six years. The life extension applies to the other third. That is defensible on its face, since a datacenter shell plausibly does outlast a 15-year schedule while the accelerators inside it plausibly do not. But it means the headline reading, that Microsoft extended the life of its AI infrastructure, is close to backwards. It extended the life of the buildings, and left the silicon where it was.
Amazonβs disclosures make the discretion harder to miss, because two moves sit in a single paragraph of the 10-K covering 2024 (filed 7 February 2025; dated background, well outside the current window, but the clearest statement of the pattern in the public record):
In Q4 2024, we completed a useful life study for certain types of heavy equipment and are increasing the useful life from ten years to thirteen years for such equipment effective January 1, 2025. β¦ We completed our most recent servers and networking equipment useful life study in Q4 2024, and are changing the useful lives of a subset of our servers and networking equipment, effective January 1, 2025, from six years to five years.
Same study period, same effective date, opposite directions: warehouse machinery up three years, servers down one. Amazon estimated the first would add about $0.9 billion to 2025 operating income and the second would remove about $0.7 billion. The stated reason for the server cut was βan increased pace of technology development, particularly in the area of artificial intelligence and machine learning.β
And the servers had been going the other way twelve months before. From the same filing: βWe had previously increased the useful life of our servers from five years to six years effective January 1, 2024,β an increase worth βa reduction in depreciation and amortization expense of $3.2 billion and a benefit to net income of $2.5 billion, or $0.23 per basic share and $0.23 per diluted share.β The 2025 reversal came in, per the FY2025 10-K (6 February 2026), at βan 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.β
Nothing was learned about servers between January 2024 and January 2025 that would justify a one-year round trip on a physical estimate. What changed was the expected pace of replacement: a forecast about the market for accelerators, expressed as a fact about equipment.
The accounting literature has known for a long time that useful life is where discretion goes to hide, and the mechanism is procedural rather than conspiratorial. Changing the depreciation method is conspicuous. Changing the estimate is not. Keating and Zimmermanβs archival study (Journal of Accounting and Economics, 1999; foundational background, not recent evidence) found that roughly 80% of sampled method changes drew an auditor consistency exception, against roughly 10% of estimate revisions, while the median earnings effect of an estimate revision ran around 13% of net income. The larger lever attracted the lighter scrutiny.
More recent work suggests the timing is not random. Albrecht, Glendening, Kim and Lee (Review of Accounting Studies, volume 29, 2024; published online 2023) examined 2,293 material changes in accounting estimates disclosed between 2005 and 2015, of which depreciation and amortization was the second-largest category at about 17%. They found firms significantly more likely to announce an income-increasing change when pre-change earnings just missed consensus, and investors discounting the resulting beat by about 17%.
Their overall conclusion is more forgiving than that one finding sounds, and it should be carried into the rest of this piece: the same paper reports that material estimate changes on average improve the usefulness of earnings, in the sense that they help predict future cash flows. Estimate revisions carry real information about the business. The timing is what looks opportunistic, not the practice. Nothing in this literature says Microsoft or Amazon did anything improper, and I am not claiming they did. It says the lever exists, that it is large, and that it attracts less scrutiny than its size warrants.
I did not expect the sharpest version of this to come from a government statistical manual.
The Bureau of Economic Analysis maintains the depreciation schedules underlying the US national accounts. Its published table, BEA Rates of Depreciation, Service Lives, Declining-Balance Rates, and Hulten-Wykoff Categories, gives almost every class of private nonresidential equipment four values: a rate of depreciation, a service life in years, a declining-balance rate, and a Hulten-Wykoff category. Communications equipment in rental and leasing gets 0.1500 and 11 years. Nonmedical instruments get 0.1350 and 12 years. Railroad equipment gets 0.0589 and 28 years. Even nuclear fuel, which BEA handles with a straight-line rate and a Winfrey retirement pattern rather than the usual geometric treatment, still carries a service life of four years.
The first row of that section reads:
Computers and peripheral equipment /2/ β¦β¦ β¦β¦ β¦β¦ β¦β¦
Four ellipses. The footnote explains why: βFor computers and peripheral equipment, information on used asset prices is available and empirical profiles are used.β The rates come instead from Stephen Olinerβs work on what used mainframes fetched in the secondary market, beginning in 1993.
Computers are not quite alone in that blankness, and the company they keep makes the point better than solitude would. One other line in the section is empty across all four columns: autos, whose rates footnote 12 says are βderived by BEA from data on new and used auto prices.β The two assets the national accounts decline to describe in years are the two with deep, liquid second-hand markets. Where BEA can watch what a used one sells for, it stops estimating a lifespan and reads the price instead.
Asked how many years a computer lasts, the US statistical system declined to answer in years. It went and looked at what buyers would pay for a used one, and it has done so for more than thirty years. That is the same question Microsoft and Amazon are answering with a single integer in a footnote, and it was decided long ago that the integer was the wrong instrument.
The empirical profile that came out of that decision is steep. Doms, Dunn, Oliner and Sichel (Tax Policy and the Economy, 2004; background again, and about personal computers rather than datacenter hardware) report that BEA adopted a geometric depreciation rate of about 34% a year for PCs in the December 2003 revision of the national accounts.
That figure cannot be laid directly against a five-year straight line, and doing the comparison sloppily reproduces the same category error this essay is about. A 34% geometric rate applies to the balance remaining; a straight-line life applies to original cost. The honest comparison is what is left on the books after a given number of years. Three years in, a five-year straight line has 40% of cost remaining and a six-year straight line has 50%, against about 29% under the geometric profile BEA adopted. Two years later the five-year schedule has hit zero, the six-year schedule still carries 17%, and the geometric profile, which never quite reaches zero, is down to 13%. Microsoftβs twenty-five-year buildings, for scale, are still carrying 88% of cost at year three.
The caveat matters: PCs in the 1990s are not accelerators in a hyperscale fleet, and to my knowledge no one has published a used-price study of hyperscale server or GPU depreciation. A search of the scholarly literature turned up nothing. So the asset class currently absorbing the largest capital programme in corporate history has no public measurement of how fast it loses value.
The obvious reading of all this is that the lives are too long. The reliability evidence points the other way, which is what makes the picture two-sided.
An operational study of a 63-node NVIDIA B200 production cluster, covering 504 GPUs across 55 days of telemetry and 224 training sessions (arXiv:2605.09370, May 2026), recorded ten XID-identified GPU failures over the period, with node exclusions concentrated enough that the top three of 63 nodes accounted for more than half of them. Failures are rare and they cluster. That is not a fleet dying of old age at year five. I wanted a second, longer-run number here on how much working hardware operators keep in service past warranty, and I could not verify one well enough to print it, so I am leaving the gap visible rather than filling it.
So both hyperscaler moves can be right at the same time. The shell probably does outlast fifteen years. The silicon probably is pulled before it breaks, retired by a better performance-per-watt figure rather than by a fault. Amazon said as much in its own filing: the reason given was the pace of technology development, not the failure rate.
What that leaves is a single line called depreciation, growing fast. Microsoftβs went from $29.4 billion in fiscal 2025 to $38.5 billion in fiscal 2026, against capital expenditures that went from $64.6 billion to $115.9 billion over the same two years. And it now aggregates two asset classes moving in opposite directions for opposite reasons. The composite tells you very little. It is the same problem as scoring one model against three definitions of the same condition: the number is stable-looking and the thing underneath it is not one thing.
The disclosure that would help is not a better estimate. It is a split: the depreciation charge and the assumed life for the short-lived compute fleet reported separately from the buildings, quarterly, by every company running one. Microsoft is a third of the way there already, since Hood volunteers the two-thirds/one-third capex split on the call; it just does not appear in the financial statements, where the depreciation it eventually becomes is a single number.
CoreWeave, whose asset base is concentrated in the short-lived half, puts the exposure about as plainly as a risk factor can (Q2 2026 10-Q, filed 12 August 2026): running the business βentails cycling out older components of our infrastructure and replacing them with the latest technology available. This requires us to make certain estimates with respect to the useful life of the components of our infrastructure β¦ We cannot guarantee that our estimates will be accurate.β
That is the correct level of confidence. It belongs in more places than the risk factors. Depreciation is now among the largest expense lines these companies carry, and it is the one whose size is set by a prediction about how fast the technology moves. The prediction is made by the people whose earnings it determines. As with a lab publishing its inference bill as evidence of acceleration, the number is precisely stated and audited, and it is measuring something other than what it appears to measure.
References