An accounting assumption buried in the notes to the financial statements determines a meaningful share of reported profit at the largest technology companies. It is not a controversial estimate in most industries. In one where the assets cost tens of billions and the technology turns over rapidly, it has become the quiet centre of the AI capital spending debate.
This article is for informational and educational purposes only. It is not financial advice, investment advice, or a recommendation to buy, sell, or hold any security, cryptocurrency, or financial product. Always verify data with official sources before making financial decisions.
Short answer: how long a GPU counts as an asset
Short answer: When a company buys computing hardware it does not expense the cost immediately. It capitalises the asset and depreciates it over an estimated useful life, spreading the cost across that period. Hyperscalers currently use schedules in the range of five to six years for server and networking equipment. Lengthening that estimate reduces annual depreciation and increases reported operating income, with no change in cash.
The schedules currently in use
Major cloud operators extended their assumed server lives over the past several years, moving from roughly three or four years toward five or six. Each extension was disclosed and each produced a quantified benefit to operating income in the year of change.
The stated justification was reasonable: improved hardware reliability, better utilisation through software, and the observation that older servers remain productive for lower-tier workloads rather than being retired. For general-purpose compute this holds up well.
The case that real economic life is shorter
The counterargument focuses specifically on accelerators used for frontier AI training, where the economics differ from general-purpose servers in a way the blended schedule may not capture.
Successive generations of AI accelerators have delivered large improvements in performance per watt and per dollar, arriving on a compressed cadence. In a market where compute is sold competitively, an operator running a two-generation-old accelerator faces a genuine problem: the newer hardware does the same work at materially lower cost, so the older asset either drops in price or sits idle.
The counter-counterargument is that older accelerators cascade to inference, fine-tuning and smaller workloads rather than being scrapped, which is precisely the pattern that justified longer schedules for general servers. Whether that cascade absorbs the volume of hardware now being deployed is the open question, and it has not been tested at this scale.
What a one-year change does to reported income
The arithmetic is simple and the magnitude is large. A $100 billion asset base depreciated over five years produces $20 billion of annual depreciation. Over six years it produces roughly $16.7 billion — a difference of more than $3 billion flowing directly to operating income.
Run that across the combined capital spending of the major hyperscalers and a one-year change in assumed life moves reported operating income by an amount comparable to a meaningful earnings surprise. The change is disclosed when it happens, but it enters the year-over-year comparison as an improvement in profitability rather than as an accounting change, and models built on reported margins absorb it silently.
Where to find the assumption in a 10-K
Useful life estimates appear in the summary of significant accounting policies, usually under property and equipment. Changes in estimate are disclosed separately, typically with a quantified effect on the current period.
The property and equipment note gives gross carrying value, accumulated depreciation and the breakdown by asset category. Comparing gross value against accumulated depreciation over time indicates how far through their assumed lives the assets are — and a fleet whose average age is rising relative to its assumed life is one where the depreciation charge is about to increase regardless of new spending.
Cash flow is unaffected — true and incomplete
The standard rebuttal is that depreciation is non-cash, so the assumption changes accounting profit without changing economics. Free cash flow is unaffected because the cash left when the hardware was purchased.
That is correct and it is not the whole picture. Depreciation is an estimate of a real economic cost — the consumption of an asset that must eventually be replaced. If the assumed life exceeds the true economic life, reported profit overstates sustainable profit, and the gap appears later as higher-than-expected replacement capital spending.
The useful discipline is to check whether capital expenditure is tracking depreciation over a full cycle. A company reporting strong operating income while capital spending persistently exceeds depreciation by a wide margin is one where the depreciation charge is understating the cost of staying in business.
Mini glossary
- Useful life. The period over which an asset is expected to generate economic benefit.
- Change in accounting estimate. A revision applied prospectively, without restating prior periods.
- Straight-line depreciation. Equal expense in each year of the asset’s life. Standard for this equipment.
- Impairment. A write-down when an asset’s recoverable value falls below its carrying value.
Risks, uncertainty, and limits
Specific schedules vary by company and by asset category within a company, and blended disclosures make it difficult to isolate the assumption applied to AI accelerators specifically.
Nothing here asserts that any company’s estimate is wrong. Useful life is a judgement made by management and reviewed by auditors, and reasonable people disagree about hardware whose economic life has not yet been observed at this scale.