Hyperscaler depreciation schedules and what they hide about the AI bubble

Why do hyperscaler depreciation schedules matter for the AI bubble?

A hyperscaler is a company that runs cloud computing at very large scale. Its servers are its largest class of assets, and AI has raised the spending on them. The schedule that spreads that spending over time sets the yearly charge. Because the charge is large, a small change in the assumed life reaches reported profit directly.

Schedules are also public. Each company states the assumed life of servers and network equipment in the notes to its annual report, and it states the effect of any change. That makes them a better test than opinions, because a reader can compare the notes of different companies side by side.

What did Amazon change in its server depreciation schedule?

Amazon shortened the useful life of a subset of its servers and networking equipment from 6 years to 5. For 2025 the change raised depreciation and amortisation by $1.4 billion and lowered net income by $1.0 billion, according to Hudson Labs, which cites the annual report filed on 5 February 2026. Amazon gave faster technology change in AI as the reason.

A shorter life lowers reported profit today. A company that makes that choice accepts a lower earnings figure, and observers often read the move as conservative. It also supports part of the sceptics’ point: if a leader in cloud computing sees the useful life shrinking, the debate about longer lives gains a data point.

What does Meta’s 5.5 year depreciation schedule do to reported profit?

Meta uses 5.5 years for servers and network assets. Hudson Labs reports that this assumption lowered depreciation by $2.92 billion for 2025 and raised net income by $2.59 billion. Burry estimated that Meta’s profits could be overstated by roughly 21% by 2028 (CNBC, November 2025), and his figure comes from the same kind of assumption.

The Hudson Labs entry for Meta cites a July 2025 earnings call and not a filing, and its date details need checking against Meta’s annual report. Readers who use the figures should confirm them in the filing itself. Meta’s own reasons for the life are for Meta to give, and this page reports only the numbers.

What does the gap between Amazon and Meta schedules reveal?

The two companies moved in opposite directions. Amazon cut a life from 6 years to 5, and Meta uses 5.5. The profit effects point apart: a $1.0 billion fall in net income at Amazon and a $2.59 billion rise at Meta. Hardware, suppliers, and workloads overlap between the two, yet the accounting reads differently.

The contrast does not prove that either company is wrong. Fleets differ in age, mix, and use, and a life is an average. It does prove that the assumption is a choice, and that choices produce visible differences in reported profit. Where choices differ, the reader’s task is to ask which evidence supports each.

What do hyperscaler depreciation schedules hide about the AI bubble?

A schedule hides three things. It hides the mix of chips, since one average life covers old and new hardware. It hides the resale value assumed at the end of life, which also changes the charge. And it hides the link between life and use, because the note rarely says how busy older servers stay.

A change of estimate also applies going forward. When a company lengthens a life, the lower charge appears in the years after the change, and prior years are not restated. A rising profit line can therefore contain a step caused by the assumption and not by sales, which is easy to miss on a quick read. Look for the sentence that states the effect of the change on profit.

How can you read a hyperscaler depreciation note in 4 steps?

Step 1: find the property and equipment note in the annual report. Step 2: write down the assumed life of servers and network equipment, with the year. Step 3: find any change of estimate and its stated effect on profit. Step 4: recompute the charge with a life 1 year shorter and note the gap.

Repeat the four steps for two or three companies and place the results in one table. The comparison shows who is more conservative and who is more generous, and by how many months. Update the table each year when new reports arrive, and the pattern becomes a trend that you can follow from one annual report to the next, with each cell backed by a page number.

Which hyperscaler schedules does this page leave out, and why?

This page reports only schedules that were checked against a source that answered. Other large cloud and chip customers, including Oracle, Alphabet, and Microsoft, also disclose lives, and their notes can be read with the same four steps. No figure is quoted for them here, because none was verified for this page.

This page is general information as of October 2026, not investment advice. If a decision of yours depends on AI hardware assumptions, the EQOS Asia pricing page lists the diagnostic formats that test them with a method you can recompute, and the series on the AI bubble places this debate in context.

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile