Michael Burry's depreciation argument about the AI bubble

What is Michael Burry’s depreciation argument about the AI bubble?

On 10 November 2025 Burry wrote on X that extending the useful life of assets understates depreciation and artificially boosts earnings. He pointed out that Nvidia products arrive on a 2 to 3 year product cycle, so computing equipment should not be assumed to last much longer. His claim is about accounting choices, not about demand for AI.

The claim has two layers. The first is a general point: the longer an asset is assumed to last, the lower the yearly cost. The second is a factual allegation that hyperscalers use lives that are longer than the hardware earns. Only the first layer is beyond debate, and the second layer is what the evidence must settle.

How does depreciation make AI profits look larger?

Depreciation spreads the purchase price of an asset over its assumed life. Take an illustrative server that costs $30,000. Written off over 3 years, it costs $10,000 a year. Written off over 6 years, it costs $5,000 a year. Nothing about the server changed, yet reported profit is $5,000 higher per year under the longer life.

Multiply that gap across millions of servers and the profit effect becomes large. The cash left the company on the day of purchase, so cash flow is unaffected. Only the reported earnings change, and share prices usually follow reported earnings. That link explains why an assumption on a spreadsheet can matter for a market. It also explains why investors read the notes to the accounts of AI spenders as closely as they read the headline revenue figure.

Which numbers does Michael Burry give for AI depreciation?

He estimated that the practice would understate depreciation by about $176 billion from 2026 through 2028. He also said Oracle’s profits could be overstated by roughly 27% and Meta’s by roughly 21% by 2028 (CNBC, November 2025). The percentages show how a depreciation choice reaches profit at the level of a single company.

CNBC contacted Oracle and Meta, and Nvidia declined to comment. It could not independently confirm the practice. The numbers therefore carry the status of an outside estimate built on assumptions about chip life, and they move when the assumption moves. Readers can rebuild them only by choosing a lifespan, and two readers who choose differently will reach different totals while using the same filings.

How long does an AI chip last in Burry’s depreciation argument?

A chip has a physical life and an economic life. The first is how long it runs, and the second is how long it earns money at a worthwhile price. Burry’s point rests on the second: newer chips may make older ones less valuable long before they stop working. The page on GPU useful life: 3 years or 6 follows that debate in detail.

The answer drives the size of any understatement. A lifespan of 3 years supports his view, and 6 years supports the companies. Since no single lifespan has been proven, the useful work is to measure rental prices for older chips over time and watch whether they hold or fall.

What do hyperscaler schedules say about the depreciation argument?

Companies disclose their assumed lives in annual reports, and the disclosures differ. Some firms have lengthened lives, and at least one has shortened them. The page on hyperscaler depreciation schedules and what they hide about the AI bubble lists the disclosed figures and shows what each choice did to reported profit.

Differences between firms are informative. If one firm shortens a life while another lengthens it for similar hardware, at least one of them is off, and the profit effect shows by how much. This is the cleanest test of Burry’s claim, because it uses documents that companies sign.

What counter-arguments answer the Michael Burry depreciation claim?

Three counter-arguments appear in the debate. Older chips can move to lighter tasks and keep earning, which supports longer lives. Companies test useful lives against actual use and sale prices. And auditors review the estimates every year. Each answer has merit, and none replaces evidence on resale and rental prices. The honest status of the debate, as of October 2026, is open: the mechanism is agreed and the lifespan is contested.

A balanced reading holds both sides. The mechanism of the claim is sound, the size of the effect depends on an assumption, and the companies have reasons for their choices. A reader who can state both positions in one sentence each has understood the argument better than a supporter or a critic.

How can you test the Michael Burry depreciation argument yourself?

Pick one company and open its annual report. Find the useful life of servers, then recompute depreciation with a life of 3 years and note the change in profit. Compare that change with the reported profit. A gap of a few percent is a footnote, and a gap of 20% is a thesis. Repeat the exercise each year.

This page is general information as of October 2026, not investment advice. If AI hardware spending sits inside a decision of yours, the EQOS Asia pricing page lists the diagnostic formats that test such assumptions with a method you can recompute.

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