How Long Does a GPU Last? The Depreciation Assumption Underneath AI Profitability

When a company buys computing hardware, the cost is not charged against a single year's earnings; it is spread across the asset's estimated useful life. The choice of that estimate is an accounting judgement, disclosed in the notes, and for most of corporate history it was an unremarkable one. In AI infrastructure it has become one of the more consequential numbers in the financial statements, because the assets are enormous relative to the business and the range of defensible estimates is unusually wide. The mechanics are simple and the leverage is large. Spreading a given fleet's cost over six years rather than three halves the annual depreciation charge, which flows directly to reported operating profit. Two companies with identical hardware, identical utilisation and identical revenue can therefore report materially different margins purely through the useful-life assumption — and both can be entirely defensible, because there is genuine uncertainty about how long an accelerator generation remains economically productive. The case for the longer estimate is that hardware does not stop working when a faster generation ships. Older accelerators remain perfectly serviceable for inference on smaller models, for development and testing workloads, and for customers whose price sensitivity exceeds their performance requirements — a real secondary life that the shortest depreciation schedules arguably fail to credit. The case for the shorter estimate is that economic life is set by competitiveness rather than function: if the newest generation delivers substantially more throughput per unit of energy, older hardware becomes unattractive to rent well before it becomes unable to run, and the residual value assumed at the end of a long schedule may not materialise. For anyone reading these companies' results, the practical instruction is to find the useful-life disclosure before forming a view on the margin, and to note whether it has been extended — a lengthened schedule improves reported profitability without any change in operations, which is legitimate when the underlying reasoning holds and flattering when it does not. The disclosure is not hidden. It is simply in the notes, where the interesting assumptions in any capital-intensive business have always lived.