SERVER LIVES AND PROFITS
How much do longer server lives flatter profits?
Key points
- Between 2020 and 2025 Microsoft, Alphabet, Amazon, Meta and Oracle stretched the assumed life of their servers from 3–4 years to 5–6. A longer life spreads the same cost over more years, so each year’s charge is smaller and profit is larger.
- Had the old lives been kept, depreciation in the latest year would have been about $32bn higher across the five (range $27bn–$37bn), about 7% of their combined operating profit.
- The share is largest at Oracle (17%) and Meta (10%), whose spending has grown fastest relative to their profits.
- Even with no new spending, depreciation on servers already bought rises from about $75bn to $89bn next year, because this year’s purchases have only been charged for half a year.
What they changed
| Company | Fiscal year | Server life | Profit effect in the first year (disclosed) | Source |
|---|---|---|---|---|
| Microsoft | 2021 | 3 → 4 years | +$2.70bn | Microsoft 10-K (FY2021) |
| Microsoft | 2023 | 4 → 6 years | +$3.70bn | Microsoft 10-K (FY2023) |
| Alphabet | 2021 | 3 → 4 years | +$2.60bn | Alphabet 10-K (2021) |
| Alphabet | 2023 | 4 → 6 years | +$3.90bn | Alphabet 10-K (2023) |
| Amazon | 2020 | 3 → 4 years | +$2.70bn | Amazon 10-K (2020) |
| Amazon | 2022 | 4 → 5 years | +$3.10bn | Amazon 10-K (2021) |
| Amazon | 2024 | 5 → 6 years | +$3.10bn | Amazon 10-K (2023) |
| Amazon | 2025 | 6 → 5 years (some servers) | −$0.70bn | Amazon 10-K (2024) |
| Meta | 2021 | 3 → 4 years (some servers) | +$0.62bn | Meta 10-K (2021) |
| Meta | 2022 | 4 → 4.5 years | +$0.86bn | Meta 10-K (2022) |
| Meta | 2023 | 4.5 → 5 years | — | Meta 10-K (2022) |
| Meta | 2025 | 5 → 5.5 years | +$2.90bn | Meta 10-K (2024) |
| Oracle | 2023 | 4 → 5 years | +$0.43bn | Oracle 10-K (FY2023) |
| Oracle | 2025 | 5 → 6 years | +$0.73bn | Oracle 10-K (FY2025) |
“Due to investments in software that increased efficiencies in how we operate our server and network equipment, as well as advances in technology”
Microsoft · Microsoft 10-K (FY2023)
Why it shortened: “due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning”
Amazon · Amazon 10-K (2025)
Changing an accounting estimate like this is allowed and is disclosed in the notes. Whether servers really last six years is the question: they may keep running, but AI chips are replaced by faster ones every year or two. CoreWeave, which rents out GPUs, also uses six years. Amazon moved some servers back from six years to five in 2025, citing the pace of AI.
Source: CoreWeave 10-K (2025)
The estimate for the latest year
| Company (fiscal year end) | Capex | Server depreciation (estimated) | With the old lives | Difference (range) | Operating profit | Difference ÷ profit |
|---|---|---|---|---|---|---|
| MicrosoftJun 2026 | $116bn | $21.6bn | $32.6bn | $11.0bn$10.2bn–$13.8bn | $155bn | 7.1% |
| AlphabetDec 2025 | $91bn | $11.2bn | $16.6bn | $5.4bn$5.4bn–$5.4bn | $129bn | 4.2% |
| AmazonDec 2025 | $135bn | $15.7bn | $19.8bn | $4.0bn$3.8bn–$4.3bn | $80bn | 5.0% |
| MetaDec 2025 | $70bn | $18.5bn | $26.8bn | $8.2bn$4.3bn–$10.3bn | $83bn | 9.9% |
| OracleMay 2026 | $56bn | $7.9bn | $11.4bn | $3.5bn$3.5bn–$3.6bn | $21bn | 17.1% |
| Five companies | $467bn | $75.0bn | — | $32.2bn$27.2bn–$37.4bn | $468bn | 6.9% |
The difference is before tax; after tax the effect on net profit is roughly a fifth smaller.
One company over time: Microsoft
Show the numbers
| Fiscal year | Actual | 3-year life |
|---|---|---|
| 2019 | 5.6 | 5.6 |
| 2020 | 6.7 | 6.7 |
| 2021 | 5.6 | 8.2 |
| 2022 | 8.2 | 9.9 |
| 2023 | 6.0 | 11.9 |
| 2024 | 9.3 | 15.2 |
| 2025 | 14.0 | 21.0 |
| 2026 | 21.6 | 32.6 |
The gap opens in each year a change took effect (fiscal 2021 and 2023) and then grows with spending: capex went from $14bn in fiscal 2019 to $116bn in fiscal 2026.
How the estimate works
Companies do not publish how much of their spending goes on servers. So we treat each year’s capital spending (SEC) as a batch, assume a fixed share of it is servers and network gear, and depreciate each batch straight-line over the life in force, switching to the new life for the remaining book value whenever a change took effect (as the accounting rules require). The share is chosen so that the model reproduces the first-year effect each company disclosed in its 10-K. Then we run the same batches with the original life and compare.
The fitted share of spending that is servers:
- Microsoft: 54% (from each change: 68%, 50%)
- Alphabet: 34% (from each change: 34%, 34%)
- Amazon: 25% (from each change: 25%, 27%, 24%)
- Meta: 70% (from each change: 36%, 87%)
- Oracle: 68% (from each change: 69%, 68%)
When a company changed its lives twice, fitting each change separately gives a similar share for Alphabet, Amazon and Oracle, which suggests the model is reasonable. For Meta the two changes give very different shares, because far more of its recent spending went on AI servers; the range in the table uses the lowest and highest. Amazon’s figure includes equipment acquired under finance leases (large in 2015–21). A batch bought during the year is charged for half a year in that year.
Source: SEC EDGAR
This page organises information; it does not recommend buying or selling anything. Data belongs to each publisher and may contain errors or delays. Make your own investment decisions.
Updated 30 Sep 2026 01:15 JST · company filings to Aug 2026 · industry history to 2024