A central accounting question has emerged as a new battleground in the heated debate over AI investments: What is the true economic lifespan of GPUs, the cornerstone of computing power? The answer directly impacts the billions in profits for tech giants and the validity of the current AI valuation bubble.

According to a November 17 report by Bernstein, analysts argue that setting GPU depreciation cycles at six years is reasonable. The report highlights that even accounting for technological advancements, the cash cost of operating older GPUs remains significantly lower than their market leasing rates, making extended usage economically viable.

This finding suggests that major cloud service providers like Amazon, Google, and Meta are largely justified in their current depreciation accounting policies, rather than artificially embellishing financial statements. It serves as a direct defense of these tech giants' profitability.

However, this view starkly contrasts with market pessimism. Critics, including Michael Burry—who predicted the 2008 financial crisis—argue that AI hardware like chips has an actual lifespan of just 2–3 years. Burry warns that tech giants are engaging in a dangerous accounting "trick" to artificially inflate short-term profits.

**Bernstein: 6-Year Depreciation Is Economically Feasible** Analyst Stacy A. Rasgon explicitly states in the report that GPUs can operate profitably for about six years, justifying the depreciation accounting practices of most hyperscale data centers.

The conclusion hinges on economic analysis: The cash cost of running an older GPU (primarily electricity and hosting fees) is far lower than market leasing rates. This means that even if hardware performance lags, continued operation of older GPUs can yield substantial profit margins as long as demand for computing power remains strong.

Data shows that even five-year-old Nvidia A100 chips still generate "comfortable profits" for suppliers. Bernstein estimates that only GPUs as old as the seven-year-old Volta architecture approach the break-even point for cash costs. Thus, the report argues that the 5–6-year depreciation cycles adopted by hyperscalers are economically sound.

**Economic Viability: High Leasing Rates vs. Low Operating Costs** Bernstein’s analysis dissects GPU operating economics. GPU leasing rates are an order of magnitude higher than their cash operating costs.

For example, the A100 chip reportedly achieves a contribution margin of up to 70%. Hourly revenue is estimated at $0.93, while cash costs—including power, hosting, and maintenance—total just $0.28. This vast profit margin gives cloud providers strong incentives to extend GPU usage as long as possible.

The report also notes that GPU value depreciation is nonlinear. They typically lose 20–30% of their value in the first year due to "burn-in" issues and market preference for the latest hardware, but their value stabilizes afterward.

Additionally, discussions with industry participants confirm that GPUs can physically last 6–7 years or longer. Early failures are often attributed to configuration errors during initial deployment (the "burn-in" phase) rather than inherent design flaws.

**Strong Demand for Computing Power Keeps Older Chips Relevant** Current market conditions further support the value of older GPUs. The report emphasizes that in a "compute-constrained" world, demand for computing power is overwhelming. Leading AI labs believe more computing power is key to advancing intelligence, making them willing to pay for any available capacity—even older models.

Industry analysts note that A100 computing capacity remains nearly fully booked, proving that less efficient older hardware retains value as long as demand persists.

**Tech Giants’ Depreciation Choices** Corporate filings show Google depreciates servers and networking equipment over six years; Microsoft uses 2–6 years; and Meta plans to extend the lifespan of some server and network assets to 5.5 years starting January 2025.

However, not all companies are extending depreciation periods. Amazon shortened the expected lifespan of some servers and networking equipment from six to five years in Q1 2025, citing accelerated AI advancements. This move lends some credence to bearish arguments and reflects divergent industry views on hardware obsolescence.

**The "Big Short" Warning: Accounting "Tricks" and Inflated Profits** Contrary to Bernstein’s optimism, Michael Burry issued a stark warning on November 11 via social media. He alleges that tech giants are understating depreciation by extending asset "useful lives," artificially boosting earnings.

Burry points out that AI chips and servers typically have product cycles of 2–3 years, yet companies like Meta, Alphabet, Microsoft, Oracle, and Amazon stretch depreciation to six years. He projects this accounting approach will inflate big tech profits by $176 billion from 2026 to 2028, with Oracle’s profits potentially overstated by 26.9% and Meta’s by 20.8% by 2028.

Burry’s concerns are not isolated. In mid-September, Bank of America and Morgan Stanley warned that the market severely underestimates the true scale of current AI investments and is unprepared for a surge in future depreciation costs, which could reveal tech giants’ actual profitability as far weaker than expected. These warnings, combined with Burry’s disclosed Nvidia put options, have heightened market anxiety over AI stock valuations.