What GPU depreciation is — change one number and an AI company's profit changes
Depreciation spreads the purchase cost of equipment such as GPUs across the years it is expected to be used. That expected life is an estimate chosen by the company rather than a measured fact, so assuming six years instead of three materially changes reported profit in any given year
The three lines
- Mechanism — a GPU's cost is not expensed in the year of purchase but spread across its assumed useful life
- The issue — the company chooses that life. Major cloud operators assume six years for server assets; Michael Burry argues the real figure is two to three
- Scale — Burry estimated that shortening the schedule to two or three years could move cumulative 2026–2028 earnings by more than $176bn
Key questions
- What is depreciation?
- Spreading the cost of a long-lived asset across the years it is expected to be used, rather than expensing it all in the year of purchase. Buy a $100m server expected to last five years and you record $20m of expense each year. The purpose is to match costs to the revenue they help generate. Expensing the whole amount up front would produce a large loss in year one and cost-free profits for four years afterwards, so no year's figures would describe the business. Note that the cash left when the asset was bought — depreciation is not a cash event, only an allocation on the books.
- Who decides how many years?
- The company does, and that is the heart of the matter. Accounting standards require assets to be depreciated over their economic useful life but do not specify a number. Management estimates it and auditors assess whether the estimate is reasonable. So two companies can assume different lives for the same class of server, and one company can change its assumption from year to year. Meta extended the useful life of its servers and network assets from four years to five and a half, and that single change reduced its 2025 depreciation expense by $2.9bn. Amazon moved the other way, shortening the life of a subset of servers from six years to five and citing 'the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.'
- Why did this become a public argument?
- Michael Burry took it into the mainstream. His claim is that large cloud providers depreciate Nvidia-based data centre hardware over five or six years while Nvidia's chip cycle implies a real economic life closer to two or three. If a schedule is longer than the true life, annual expense is understated and profit correspondingly overstated. Burry estimated the cumulative effect over 2026–2028 could exceed $176bn. There is a counterargument: older GPUs are not discarded when a new generation arrives — they move down to inference and cheaper workloads. This page does not adjudicate between them.
Changing an AI company's reported profit materially requires no new revenue and no cost cuts. One number will do it.
The number is "how many years will we use this GPU," and the party who picks it is the company.
1. What depreciation does
Suppose a company buys a $100m server. The $100m in cash left when it was bought. The remaining accounting question is which year should carry the expense.
| Method | Year 1 expense | Years 2–5 expense |
|---|---|---|
| Expense in full at purchase | $100m | $0 |
| Spread over five years | $20m | $20m each |
Under the first method, year one shows a large loss and the next four show cost-free profits. No year's numbers describe the business.
So accounting uses the second: match the expense to the period over which the asset earns. That is depreciation.
One frequent confusion is worth clearing. Depreciation is not a cash event. The cash went out at purchase. The annual charge is an allocation on the books, which is why cash flow statements add depreciation back to net income.
2. The years are chosen, not measured
Accounting standards require depreciation over an asset's economic useful life. They do not say how long that is.
Management estimates it; auditors assess whether the estimate is reasonable. The result is that two companies can assume different lives for the same class of server, and one company can change its own assumption between years.
Here is how the estimate moves profit. Same $60bn of equipment.
| Assumed life | Annual depreciation | Effect on that year's profit |
|---|---|---|
| 6 years | $10bn | baseline |
| 5 years | $12bn | $2bn lower |
| 3 years | $20bn | $10bn lower |
Not a single chip changed. Only the assumption did.
3. What companies have actually done
This is not hypothetical. It has happened recently.
| Company | Change | Reason / effect |
|---|---|---|
| Meta | servers and network assets, 4 years → 5.5 years | 2025 depreciation expense reduced by $2.9bn |
| Amazon | a subset of servers, 6 years → 5 years | "the increased pace of technology development, particularly in AI and machine learning" |
The directions are opposite. In the same period one company extended and the other shortened.
Between 2020 and 2024 large technology companies broadly extended useful lives, arguing that servers were lasting longer than before. In 2025 the trend split, and the reason Amazon gave for shortening was AI.
What to notice in the Meta example is $2.9bn. That amount was not earned by running more capacity or saved by cutting costs. It is the difference that a change in an estimate produced on the income statement.
4. The question Michael Burry raised
This topic moved from accounting departments into mainstream news when Michael Burry raised it publicly.
His argument reduces to two sentences.
- Large cloud providers depreciate Nvidia-based data centre hardware over five or six years.
- Nvidia's generational cycle implies a real economic life closer to two or three years.
If both hold, annual expense is understated and profit correspondingly overstated. Burry estimated that shortening schedules to two or three years could move cumulative 2026–2028 earnings by more than $176bn.
5. The counterargument
There is a rebuttal, and its core is that being superseded is not the same as being unusable.
| Burry's case | The counter-case |
|---|---|
| A new generation arrives every two to three years | New generations arrive; old chips are not discarded |
| Older chips cannot compete with the newest | Pushed off frontier training, they move to inference and cheaper workloads |
| Real life = frontier life | Real life = however long it earns anything |
The dispute is unresolved, because there is no established consensus on the real economic life of a GPU. What a three-year-old GPU inside a data centre is actually doing is not information the operators publish.
This page does not adjudicate. The verifiable part ends here: a number with a large effect on reported profit is an estimate rather than a measurement, and the party estimating it is the company being judged by that profit.
6. What you can actually check
A few things are visible in the filings.
First, whether the useful life changed. Annual reports disclose changes in accounting estimates. If profit rose in a year that also contained such a change, the change should be separated from operating improvement.
Second, the growth rate of depreciation against the growth rate of capital expenditure. If capex is climbing sharply while depreciation is not following, the gap is a question worth asking.
Third, comparison within a sector. When companies using similar equipment assume different lives, margin comparisons between them are distorted by exactly that difference.
This page covered a different layer of AI earnings on August 12 in "Why a good earnings report can still sink a stock," and covered where the money to buy this equipment comes from in "Broadcom seeks more than $60bn in debt." Depreciation is the story of how fast that equipment disappears from the books.
7. What is unresolved
- $176bn — follows from Burry's own assumptions, not verified through filings. Different assumptions give different results.
- "Six years at the big three" — from analysis pieces, not checked line by line against each company's latest annual report.
- Meta's $2.9bn — as cited in secondary material, not checked against the original filing.
- Real GPU life — no established consensus. Both arguments are set out; neither is endorsed.
- Korean companies — the useful-life policies of Korea's chipmakers and cloud operators are outside this article and need separate verification.
Sources
- National Law Review — Deep Quarry: Useful Lives of GPUs, Key Considerations
- Deep Quarry — Depreciation of GPUs: between useful lives and useful myths
- theCUBE Research — Resetting GPU Depreciation: Why AI Factories Bend, But Don't Break, Useful Life Assumptions
- Forbes — The Hidden Variable In The AI Rally: A Depreciation Reality Check
- Levelheaded Investing — Are AI Chip "Useful Lives" Creating Useless Earnings?
- Harvard Business Publishing — Meta: Accounting for AI Data Center Depreciation