GPU lifespans and depreciation: a hidden AI bubble risk

What is GPU depreciation, and why does it matter for the AI bubble?

Depreciation spreads the price of an asset over the years the owner expects to use it. A company that buys AI chips does not record the whole cost in one year. It records a share each year, and that share depends on the life it assumes. Cash leaves on the day of purchase, while profit absorbs the cost slowly.

This matters for the AI bubble debate because AI spending is enormous and chips take a large part of it. If the assumed life is too long, reported profit looks better than the economics. The real cost shows up later, as write-downs or as lower returns. Depreciation turns a spending question into an accounting question that investors can check.

How long do GPU lifespans run in AI data centers?

Ask which life is meant, because there are 3. Physical life is how long a chip keeps running, which can be many years. Economic life is how long it earns a good return, and that is shorter when new chips arrive. Accounting life is the number in the books, chosen by management within rules that auditors review.

The gap between physical and economic life is the heart of the issue. A chip that still runs can be too slow and too power-hungry to rent at a profit. Owners then keep it for lighter work or sell it at a discount. Judge the accounting life against the economic life, since the physical life tells you little about returns.

How does the depreciation period change reported profit on AI chips?

Take an illustration, not a real company: $10 million of chips. Over 6 years the yearly cost is about $1.67 million. Over 4 years it is $2.50 million. Over 3 years it is $3.33 million. The cash spent is the same in all 3 cases, yet the reported profit differs by $1.67 million each year between the longest and shortest period.

Scale that difference to a build-out of thousands of chips and the effect reaches reported earnings, the valuation built on them, and the borrowing that rests on both. A company that extends its assumed life can show higher profit in the same quarter without selling one more unit. Read the annual report notes on useful life, and watch for any change from one year to the next.

What makes GPUs lose value faster than the AI bubble books assume?

New generations deliver more computing for each watt and each dollar, so older chips become less attractive even when they work. Software also changes. A model trained on the latest chips may run slowly or expensively on older ones, and customers pay less for capacity they consider outdated. Each of these forces shortens the economic life of a chip.

Demand can slow the same process. If AI usage grows more slowly than the supply of chips, rental prices fall for every generation, new and old. A glut hits the oldest chips first and hardest. When you hear that chips keep their value for many years, ask for evidence in the form of rental contracts at stable prices.

How do rental prices and resale values test GPU lifespans?

The market gives 2 honest checks. Rental price per hour for older chips shows what customers pay today for yesterday’s technology. Resale value shows what a buyer pays when the owner leaves. Both can be tracked over time, and both tell you more than a depreciation schedule written years earlier by the owner.

Compare them with the book value of the same chips. If rental income on 3-year-old chips still covers their yearly depreciation, the assumed life looks reasonable. If it does not, the books are carrying a value that customers no longer pay for. Record the gap each quarter, because a widening gap is an early sign of future write-downs.

How does GPU depreciation raise risk for AI data center owners?

Chips are bought with money that has a deadline. When the loan or lease runs longer than the useful life of the chips, the owner must replace the equipment while still repaying it. That doubles the cash need in the same period and strains a business whose revenue has not yet caught up with its spending.

The risk grows with the speed of the build-out. A company that buys a new chip generation every year carries several layers of equipment at different ages. The oldest layer earns the least. Ask what share of installed chips is older than the assumed life minus 1 year, since that share is close to needing replacement.

How can you check GPU depreciation risk in an AI company?

Use a short list. Find the useful life in the annual report, note any recent change, compare it with rental prices for older chips, and estimate profit under a shorter life using the arithmetic above. If profit turns negative under a shorter life, the company relies on its accounting choice. Record the result with its date.

The EQOS Asia AI study counts about $200B a year of AI spending and 95% of projects failing in production (April 2026). Spending at that scale makes asset lives matter. The method page explains how a score is built, and a 30-minute session on the contact page applies the same reasoning to your own AI investment decision.

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