Successive extensions to the estimated useful lives of data centre equipment have prompted claims that hyperscaler profits are being flattered. One prominent critic describes the practice as one of the “common frauds of the modern era” and estimates that industry depreciation will be understated by $176bn over three years.
We are not persuaded that the useful-life estimates are unreasonable. In our view, the real challenges for investors are the absence of required fixed asset componentisation under US GAAP and the prospective treatment of changes in estimates. Both can affect the usefulness of depreciation as a component of performance and as a guide to future capital spending.
Major cloud and AI infrastructure providers, notably Amazon, Microsoft, Alphabet, Meta and Oracle, are investing heavily in servers, networking equipment and data centre facilities. Depreciation of this asset base is becoming one of the largest expenses in their income statements, which means that even modest changes to depreciation calculations move earnings materially.
In recent years each of these companies extended the estimated lives of servers and networking equipment, in most cases more than once. Server lives that were commonly three or four years are now five or six, and the resulting reduction in depreciation charges has been material. Not all changes to useful life were increases. In early 2025 Amazon reduced the life applied to a subset of its equipment from six years back to five, citing the “increased pace of technology development, particularly in the area of artificial intelligence and machine learning”. However, in the same reporting season Meta extended the life of most of its server and network assets to five and a half years, significantly reducing its 2025 depreciation.
The earnings benefit of longer useful lives is real, but it will only persist if the actual asset replacement cycle matches the assumption. If useful lives are overstated, the short-term earnings benefit will soon be reversed.
The useful life controversy
The most prominent critic is the investor Michael Burry, who in late 2025 accused the hyperscalers of understating depreciation by extending asset lives, describing the practice as “one of the more common frauds of the modern era” and estimating that depreciation would be understated by some $176bn across the industry between 2026 and 20281See for example ‘Big Short’ Michael Burry exposes big tech’s ‘common frauds of the modern era’.. Other commentary highlights the divergence between Amazon and Meta, as proof that these estimates are subjective choices that management can adjust to manage earnings. The debate has continued in recent months.
Are different useful lives for similar assets evidence of earnings management?
These concerns are plausible. Depreciation driven by useful-life estimates is subjective, the earnings effects are potentially large, and the direction of the changes has, until recently, been almost always favourable to reported profit. However, in our view, the inference of misreporting or earnings management does not follow. That is not to say the critics are wrong about the direction of travel. As we explain below, depreciation of AI infrastructure in the early years may well be lower than a componentised charge would produce – but, in our view, the problem lies in the accounting framework.
Why we are not persuaded
Useful life is an entity-specific estimate. Under both IFRS and US GAAP, the useful life of an asset is the period over which the reporting company itself expects to use it, not a physical property of the hardware or a constant applied throughout an industry. Different business models and planned replacement cycles for essentially the same asset can support genuinely different useful lives, which means that differences between companies are not, of themselves, evidence that one of them must be wrong.
Different replacement plans can justify different useful lives
Residual values also moderate the effect of differences in useful life. Depreciation is charged on cost less estimated residual value, and the residual value estimate itself reflects the planned replacement cycle. A company that expects to dispose of equipment after three years should recognise the value it expects to recover on disposal, reducing the annual charge, while a company holding similar assets for six years to the end of their economic life would apply a lower or nil residual. The depreciation difference between the two approaches may therefore be smaller than the difference in useful lives alone would suggest.
A further self-correcting mechanism gives management an incentive to set realistic lives. If useful lives prove too long, the consequences will eventually appear through a downward revision that increases future depreciation or through a loss when equipment is retired early. Neither is welcome considering that both must be explained to investors (although disposal losses may well be removed from non-GAAP metrics). Amazon illustrates the latter mechanism: alongside its reduction in useful lives, the company recorded approximately $920m of accelerated depreciation and related charges on equipment it decided to retire early. This does not prove that its earlier estimates were incorrect, but it does show that changes in expected asset consumption can become visible in reported results.
None of this means that you should not question useful-life estimates. Investors should be sceptical about changes to significant accounting estimates, and the disclosures deserve close reading. Our point is that treating extended useful lives as presumptive evidence of dubious accounting misunderstands the nature of the estimate and ignores the incentives to make estimates as realistic and unbiased as possible.
In our view, the more important question for investors concerns not the honesty of the estimates, but the required accounting methodology itself.
Depreciation issues that should concern investors
Two features of depreciation accounting deserve more attention than they have received in the current debate. Both affect the pattern of reported expense, and both complicate the use of depreciation in performance measurement and in forecasting.
No required componentisation under US GAAP
Under IFRS, IAS 16 requires each component part of an item of property, plant and equipment, with a cost that is significant relative to the total, to be depreciated separately over its own useful life – so-called componentisation. The classic illustration is an aircraft, in which the airframe may be depreciated over a different period compared with the engines. This also applies to the overhaul condition embedded within the engines2Under IFRS, major engine overhauls are accounted for as part of the componentised depreciation cycle. The part of the initial cost of the engine that equates to the cost of the next overhaul is depreciated over the period to that overhaul, with the rest of the initial cost depreciated over the full estimated asset life. When the overhaul subsequently occurs, the amount paid is added to the fixed asset balance and depreciated over the next overhaul period., which is depreciated over the much shorter period to the next major overhaul. In this example there are therefore three components, each with a component-specific depreciation rate.
IFRS works well when asset components are retired at different times
AI infrastructure (and particularly the servers) is similar. The chips may well have a shorter useful life than the server in which they sit, with the chips therefore swapped, potentially more than once, before the full asset is finally retired. In this situation IFRS works well, with the cost of the chips depreciated over their shorter replacement cycle. Furthermore, if the chips are replaced earlier than expected, a disposal loss or accelerated charge applies.
The problem with US GAAP is that componentisation, while permitted by ASC 360, is not required and we understand is uncommon in practice. There is no single US GAAP alternative to componentisation, and different companies may apply different approaches. The problem for investors is that companies rarely spell out exactly what their approach is in their accounting policy notes – including the hyperscalers.
Meta: Property and equipment accounting policy
Meta 2025 10k
The Meta extract is as notable for what it does not disclose as for what it does: servers and network assets are presented as one asset class with useful lives of five to 5.5 years, with no reference to significant components or their replacement. The disclosure therefore does not tell investors how a significant component replacement would affect the carrying amount or future depreciation. The disclosure is entirely compliant; it is the US GAAP requirements that stop short of what investors need.
If the server is treated as a single unit of account, its aggregate cost is depreciated using the useful life and depreciation method assigned to that unit. If that life effectively represents a weighted average of different component lives, the resulting depreciation may approximate componentised depreciation. However, the asset is still a single unit of account and early retirement of chips will likely not produce the same adjustments as would be required under IFRS.
A single long asset life under-depreciates short-lived components
The problem arises where the depreciation rate for the asset is based on its longer-lived components while significant shorter-lived components are expected to be replaced earlier. In effect, the longer server life is attributed to the chips as well, potentially resulting in lower depreciation before the chips are replaced. There will be a catch-up, but this comes later; exactly when depends on the precise methodology adopted.
One approach is to treat the replacement chips as ‘maintenance expenditure’ and charge the amount paid immediately to profit and loss. This may be fine if the replacement parts are small relative to the whole, but it is unlikely to provide a faithful representation of the consumption of the asset if they are more significant. It can also lead to a volatile result.
Another approach is to capitalise a qualifying replacement and remove the carrying amount attributable to the component being replaced. If the original component was not separately identified, estimating its carrying amount is necessary, with the cost of the replacements potentially used as a proxy. This results in an offset against accumulated depreciation and no recognition of a disposal gain or loss at that point, even if one had effectively occurred.
The important point for investors is that the timing of depreciation can differ materially depending on how the asset and its replacements are defined and tracked. US GAAP gives companies more latitude than IFRS over this unit-of-account decision. However, the hyperscalers do not disclose enough detail for investors to reconstruct the effect.
Backloaded depreciation coupled with high growth gives an earnings uplift
Depreciation in the early years of some AI infrastructure may be lower than it would be under IFRS componentisation. If so, the effect could be significant while the asset base is growing rapidly, because lower depreciation on new assets would coincide with a rapidly expanding population of under-depreciated assets.
In our view, this problem stems from lack of clear requirements in US GAAP and, particularly, the lack of a requirement to apply componentisation when calculating depreciation.
Prospective changes and the missing catch-up
The second issue is common to IFRS and US GAAP. A change in the estimated useful life of an asset, because it is a change in accounting estimate, is recognised prospectively. Nothing is restated and no cumulative catch-up adjustment is made; the remaining book value is simply spread over the revised remaining life. When a life is extended, depreciation already charged in earlier periods reflects the previously estimated shorter life. Relative to having the revised estimate applied from the outset, expense has been front-loaded, and the charges reported after the change are correspondingly lower than the revised life alone would imply.
Backloading and front-loading distortions can offset or compound
This is the opposite distortion to the one we believe is created by the absence of componentisation. A composite life that is too long for the dominant component backloads expense; a prospective life extension front-loads it relative to the revised view of the asset. The two effects can offset or compound in unpredictable proportions, which is precisely what makes reported depreciation trends for the hyperscalers difficult to interpret.
The prospective treatment is also an interesting conceptual issue in financial reporting. One could argue that a better approach would be to recalculate accumulated depreciation on the basis of the revised life and recognise the difference as an immediate catch-up adjustment in profit and loss. The advantage is that depreciation in subsequent periods is then ‘correct’, in the sense of reflecting the current estimate applied to the whole life of the asset. Of course, the disadvantage of this approach is the distorting effect of the catch-up itself in the period of change. Financial reporting generally favours prospective treatment, but not universally, nor consistently between US GAAP and IFRS3For example, a change in the estimate of insurance contract profitability is recognised through a catch-up adjustment under US GAAP but prospectively under IFRS 17, where changes in estimates relating to future service adjust the contractual service margin and affect profit over the remaining coverage period..
A simple illustration shows the difference. Assume an asset costing 120 is initially depreciated over four years. At the start of year three the estimated total life is revised to six years. Under the prospective approach required under IFRS and US GAAP, the remaining book value of 60 is spread over the revised remaining life of four years. Under a hypothetical catch-up approach, accumulated depreciation would be recalculated at 40 (two years at 20), producing a credit of 20 in year three, with subsequent charges at the revised annual rate.
A change in useful life: prospective versus catch-up
Assume: asset cost 120, initial estimated life 4 years, revised to 6 years at the start of year 3. Nil residual value. In year 3 the catch-up approach combines a charge of 20 with a credit of 20 arising from the recalculation of accumulated depreciation.
The Footnotes Analyst
Amazon provides a good summary of this analysis.
Amazon: Changes in the estimated useful lives of servers and networking equipment
Accounting policy extract
Use of estimates extract
Amazon 2024 10k
The disclosure shows the mechanics of the prospective method, with the effect of the change quantified for the following year. It also shows the self-correction mechanism, in the form of accelerated depreciation on equipment retired ahead of schedule. The stated reason, the pace of AI-related technology development, is exactly the consideration critics accuse the industry of ignoring. However, like for Meta above, what the disclosures do not show, because the accounting does not require it, is how the composite life relates to the different replacement cycles of the components within it.
EBITDA is not the solution
A common investor response to concerns about depreciation is to exclude it. If useful-life estimates are subjective, and if IFRS and US GAAP measure depreciation differently, then metrics such as EBITDA, and the valuation multiple EV/EBITDA, would appear to offer comparability by removing the problematic number altogether. The current controversy, and our assertion that reported depreciation may be difficult to compare, might seem to strengthen that argument for EBITDA.
Omitting a real expense does not create comparability
We disagree – excluding a real expense does not create comparability. Depreciation represents the consumption of productive capacity that must ultimately be replaced through capital expenditure. Despite depreciation being labelled as non-cash we regard it as an economic cash cost, albeit one whose cash effect arises in a different period. Ignoring depreciation means ignoring a key driver of value, which flatters capital-intensive businesses relative to asset-light ones.
That bias is particularly relevant now, given the hyperscalers have become among the most capital-intensive large companies in the market. A valuation approach that excludes what is perhaps their fastest-growing expense omits information that investors need for their analysis.
We provide a more comprehensive discussion of the merits of EBITDA in ‘EBITDA-AL: More letters but no more insight’ and illustrate the valuation multiple consequences in ‘Relative valuation conflicts – EV/EBITDA versus P/E’.
The best response to concerns about reported depreciation is scrutiny and adjustment, not exclusion. Compare disclosed lives with observable replacement cycles, note the moderating role of residual values, and allow for the front-loading created by prospective life extensions and the backloading created by composite lives. Where you need a cleaner earnings measure, we prefer EBITA, in which we retain depreciation but exclude the amortisation of most intangibles acquired in a business combination, which otherwise creates a genuine comparability problem.
An alternative approach is to use EBITDA less maintenance capital expenditure. While reported depreciation may often be the best proxy for maintenance capital expenditure, it will not always be so. Price changes, accounting policy distortions, and business combination effects, can mean that separate consideration of the future capital spending needed to maintain the existing operating business produces metrics that are more relevant for equity valuation.
The hyperscaler depreciation debate has been conducted largely as an argument about earnings management, but, in our view, that is too narrow. Investors should certainly scrutinise useful-life assumptions, but they should also consider whether the accounting methodology itself produces a useful representation of asset consumption and replacement.
Insights for investors
- Different useful lives applied to similar assets are not necessarily evidence of misreporting or earnings management. Useful lives are entity-specific estimates, and differences in expected use, replacement policy and residual values can legitimately produce different depreciation charges.
- Check whether significant short-lived components may be embedded within longer-lived assets. The absence of required componentisation under US GAAP can affect the timing of depreciation when asset replacement cycles materially differ, and potentially backload the depreciation expense.
- Prospective treatment of useful life extensions effectively front-loads the depreciation expense relative to the revised estimate, which flatters subsequent charges until the next replacement.
- Excluding depreciation entirely through the use of EBITDA is not the solution to subjective estimates and accounting distortions; we prefer EBITA or, even better, EBITDA less maintenance capex.
Note: Following our usual practice, we do not provide any opinion about the financial results of the companies mentioned in this article, the value of their businesses, or their merits as investments. Nor do we suggest that there is anything wrong with their financial reporting. Our analysis is educational and designed to help investors interpret reported depreciation.