How long does an AI server live? The hyperscalers can’t agree, and billions in profit ride on it.
How long a company assumes its servers last decides how fast it depreciates them, and therefore how much profit it reports. Since 2022, Microsoft, Amazon, Alphabet, Meta and Oracle all stretched those assumed lives toward six years, each booking a one-time earnings tailwind of roughly 4 to 5% of net income. Then Amazon did something the others have not: it shortened a subset of servers back to five years, citing the pace of AI hardware turnover. Same hardware, opposite calls. Here is the full record, from the filings.
Four extended toward six years and held. Amazon alone stepped back.
A longer assumed life means less depreciation and higher reported profit. Amazon stretched its servers out with everyone else, reaching six years, then in 2025 cut a subset back to five, citing the pace of AI hardware turnover.
Capex is running nearly 4x depreciation. The assumption sets how fast that closes.
Measured over the same recent twelve months, so the five are comparable. They are capitalizing AI hardware far faster than they expense it. The longer the assumed life, the slower depreciation catches up, and the higher reported profit stays in the meantime.
Across the five, trailing-12-month capex of ~$482B is nearly 4x their ~$125B of PP&E depreciation. Stretch the assumed life and that gap, and reported profit, both widen.
Every change, every disclosed dollar.
Each company’s server useful-life history and the earnings impact it disclosed, with a link to the filing. All figures are from the companies’ own 10-Ks.
Five companies, five accounting teams, one number: ~4 to 5% of profit.
In the year each company extended its server lives, the change lifted net income by a strikingly consistent ~4 to 4.6%: Microsoft +$3.0B (4.1%), Alphabet +$3.0B (4.1%), Amazon +$2.5B (4.2%), Meta +$2.59B (4.3%), Oracle +$0.57B (4.6%). Real money, and a real single-year boost to reported earnings per share.
But notice what that is and is not. It is a one-time, disclosed tailwind in a single year, around 4 to 5% of profit. It is not a claim that earnings are fabricated, and it is an order of magnitude smaller than the cumulative figures circulating in the bubble debate.
Michael Burry has accused the hyperscalers of inflating earnings by roughly $176B over 2026 to 2028, putting Meta at ~21% and Oracle at ~27% overstated. That is his own forward projection, and it does not match what the filings disclose: none of these companies break out a separate GPU asset class, and the disclosed, single-year impacts run around 4 to 5% of profit, not 20-plus. The accusation points at a real lever. It overstates the size of it.
Depreciation is non-cash, so none of this changes free cash flow, and these companies are valued largely on cash. So why care? Because reported earnings and margins drive multiples, headlines, and comparability, and right now the five are not comparable: they are assuming different lifespans for the same hardware. The open question is not whether anyone is lying. It is which assumption is right when one of them, Amazon, is already telling you in its own filings that the hardware ages faster.
We check the AI boom’s reported reality against the filings.
Sourced, structured research on what companies actually disclose about AI, before it makes the headlines.