What a server’s useful life is worth

Why useful life matters

When Microsoft or Meta buys a server, the cost doesn’t hit the income statement at once. It is spread over the server’s estimated useful life as depreciation. So the useful life a company chooses decides how much of its spending shows up as expense each year: a longer life means lower depreciation and higher reported profit today, and more expense left for later years.

Between 2022 and 2025 the large buyers lengthened their server lives, which raised reported earnings. Alphabet, for example, disclosed that moving servers from four to six years reduced depreciation by $3.9B in 2023. Meta disclosed an effect of $2.9B from extending servers to 5.5 years. Amazon went the other way in 2025, shortening part of its fleet to five years because of the pace of change in AI hardware.

The issue matters more now than ever because capex is running far ahead of depreciation. In Q2 2026 the five largest buyers spent 3.9× what they recorded as depreciation and amortization. All of that spending will become expense over the lives the companies choose. If accelerators turn out to have shorter economic lives than the servers of the past, which their makers’ yearly product cycles make possible, today’s lives are too long and future earnings will absorb the difference.

What the companies state

What each company states in its latest 10-K, and the changes it has disclosed since 2022 or earlier. Each stated life links to the filing it comes from; the calculator above quotes the sentence.

How the calculator works

The model is straight-line depreciation applied to reported capex. Each quarter’s capex, from mid-2020 onward, is multiplied by the share you set for servers and network equipment, and that amount is depreciated evenly over the life you choose. The depreciation of every quarter still inside its life is added up to give an annual run-rate.

Two simplifications matter. Companies don’t split capex between equipment and buildings, so the equipment share is your assumption; the calculator starts at 50%. And capex from before mid-2020 isn’t in the data, so for long lives the model leaves out some older spending. Because capex has grown so fast since then, the older spending is a small part of the total, but the figure is a model of the equipment depreciation coming from recent spending, not a reconstruction of the reported number.

The third output projects forward: if capex stayed at the latest quarter’s level for two more years, what would the run-rate be? It is not a forecast. It shows how much depreciation is already locked in by spending that has happened, plus spending at the current pace.

A worked example

Take Microsoft, with half of its capex treated as servers and network equipment. At its stated six-year life the model gives equipment depreciation of about $27.7B a year at the current run-rate. At five years it is about $31.1B, a difference of $3.4B, or 2.1% of Microsoft’s operating income over its latest four quarters.

Change the company and the numbers scale with its spending. The ratio to operating income is the figure to watch: it is larger for companies whose capex is large relative to their profits, which is why Oracle and CoreWeave are more exposed to a shorter life than Microsoft or Alphabet.

One year shorter, company by company

The table runs the same experiment for every company: half of capex treated as equipment, depreciated over the top of the life each company states, then over one year less. The last column is the extra depreciation as a share of operating income over the latest four quarters.

The pattern follows the ratio of spending to profit. Companies whose capex is small next to their operating income can absorb a shorter life with little change in reported earnings. Companies spending close to or more than they earn would see a much larger share of their profit disappear, and for CoreWeave, which has little or no operating income to absorb it, the change would show up directly as a larger loss. This is the same group that relies most on backlog and borrowing, so the depreciation assumption and the funding question point at the same companies.

Depreciation, cash and earnings

Depreciation is an accounting expense, not a payment. The cash went out when the equipment was bought, and it shows up in the cash-flow statement as capex. Depreciation then spreads that same cost over the following years on the income statement. That is why the useful-life choice changes reported earnings without changing cash flow at all, and why investors who focus on EBITDA (earnings before depreciation) can see a very different picture from those who focus on earnings per share.

Both views matter for AI. Cash flow tells you whether a company can afford the spending now; earnings tell you whether the spending is earning a return once its cost is spread over the assets’ lives. A company whose EBITDA is rising while its earnings per share stall is usually one whose depreciation is catching up with past capex. With capex running at several times depreciation across the big buyers, that pattern is likely to become common, and the useful-life assumption decides how quickly it arrives.

This is also why changes in useful life deserve attention when they happen. A longer life doesn’t create any cash, but it raises reported profit and can make returns on new spending look better than they are. A shorter life does the opposite. Either way, the companies have to disclose the change and its effect, and those disclosures are among the most informative sentences in their annual reports.

What to watch

A shortening of useful lives would be one of the clearest signals in the filings that AI hardware is aging faster than planned. It would raise depreciation immediately and lower earnings, with no change in the underlying business. Watch the property and equipment notes in each 10-K, and the accounting-estimate disclosures in the 10-Qs, where changes are announced.

The other signal is impairments: write-downs of equipment that has lost value before it is fully depreciated. None of the big buyers has reported a large one for AI hardware, but an impairment would say the same thing more abruptly. The money page tracks depreciation against capex for all five buyers.