GPU useful life: 3 years or 6? The accounting debate in the AI bubble

What does the GPU useful life debate ask about the AI bubble?

The debate asks one number: the years over which a company spreads the price of a GPU. A short life raises the yearly charge and lowers reported profit. A long life does the opposite. Because AI spenders buy hardware in very large quantities, a change of one year in the assumed life moves total profit by a visible amount.

Burry put the short view in public. He wrote that Nvidia hardware follows a 2 to 3 year product cycle, so assuming much longer lives for computing equipment flatters earnings (CNBC, November 2025). The companies hold a longer view. The rest of this page separates what each side must show.

What separates a GPU’s physical life from its economic life?

Physical life is how long the chip runs before it fails. Economic life is how long it earns a worthwhile return at the going price. A GPU can run for 6 years and be worth little after 3, if newer chips do the same work faster and cheaper. Accounting rules ask for the period of useful service, which leans toward economic life.

The two lives can diverge sharply in AI, because each new generation raises performance per dollar. A buyer who can rent a newer chip cheaply pays less for an older one. That is the core of the short-life case. It is an argument about prices in the rental and resale markets, so prices can test it.

How much does a GPU useful life of 3 versus 6 years change reported cost?

Take an illustrative GPU system that costs $40,000. Over 3 years it carries $13,333 of depreciation a year. Over 4 years the charge is $10,000, over 5 years $8,000, and over 6 years $6,667. The jump between 3 and 6 years is $6,666 a year per system, which is a halving of the charge.

The example uses invented prices to show the arithmetic. It shows the pattern: the first extra year removes the most cost, and later years remove less. A company that moved from 3 to 4 years improved yearly profit by $3,333 per system, and a move from 5 to 6 years adds only $1,333.

Can GPU rental prices settle the useful life debate?

Rental prices are the cleanest signal, because they show what buyers pay for a given chip right now. Track the hourly price of an older GPU generation over 12 months. If it falls fast after a new generation launches, economic life is short. If it holds steady, the longer life gains support.

Rental data need care. Prices differ by region, contract length, and supply. Compare like with like, and record the date of each reading. A single low price proves little, while a steady decline across many providers is a pattern. Ask each provider for its rate card, and keep the quotes next to the dates you received them.

What do resale prices and failure rates add to the GPU useful life test?

Resale prices show the value a market assigns to a used chip. A chip carried at 50% of its cost in the books, but selling at 20%, signals a gap between the accounting life and the market view. Companies rarely sell data centre GPUs in volume, so the evidence is thin and each data point counts.

Failure rates add the physical side. High heat and constant load can shorten the running life of hardware. If failure rates climb in year 3 or 4, even the physical life is shorter than the long schedules assume. Published failure data from operators would settle this point quickly, and today it remains scarce.

Does GPU useful life extend through second uses, and how can you check?

The main defence of long lives is cascading use. A chip too slow for frontier training can serve lighter workloads, such as running finished models for users, and keep earning for years. If that holds, economic life stretches toward the physical one, and a 6 year schedule gains weight.

The claim can be checked by utilisation. Ask what share of older chips in a fleet still run at high load, and at what price per hour. A fleet where old chips sit idle supports the short view. A fleet where they stay busy supports the long one. Disclosures on utilisation are a useful request to make of any AI infrastructure company.

Where does the GPU useful life debate stand today?

As of October 2026 no public dataset settles the question. The 4 tests, rental prices, resale prices, failure rates, and release cadence, each point to evidence that exists and is only partly published. The honest position is that 3 years and 6 years both remain defensible, and that the gap between them is worth a vote of caution.

This page is general information, not investment advice. If a decision of yours depends on AI hardware assumptions, the EQOS Asia pricing page lists the diagnostic formats that test such assumptions with a method you can recompute, and the series on the AI bubble shows where this debate fits.

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