Research indicates that AI-focused GPUs have a significantly shorter operational lifespan than the six-year depreciation schedules currently used by major cloud providers like Amazon, Microsoft, and Google. Data from Meta’s Llama 3 training cluster, which utilized 16,384 H100 units, recorded 148 GPU-related failures over 54 days, suggesting a 9% annual failure rate.
Studies from the University of Illinois, based on 11.7 million GPU-hours of observation, show that H100 memory degrades 3.2 times faster than A100 memory. Experts from Alphabet estimate that at 60–70% utilization, data center GPUs typically last between one and three years. Princeton researchers note that the secondary market for used hardware is insufficient to absorb the massive volume of decommissioned chips.
Investor Michael Burry estimates that current accounting practices lead to an understatement of depreciation by $176 billion for the 2026–2028 period. Barclays has consequently reduced profit forecasts for AI-focused companies by up to 10%. While some secondary market transactions for H100 units have reached 70–95% of their original price, the long-term viability of these assets remains uncertain.
Industry analysis suggests that companies should account for a 5% hardware buffer to manage constant failure rates. Furthermore, if the cost of electricity exceeds the revenue generated by inference tasks, the residual value of older GPUs effectively drops to zero. Experts advise businesses to align hardware procurement with real-world lifespans rather than optimistic accounting models.
