Why AI hardware could lose value whether adoption succeeds or stalls

Overview

This study proposes that AI infrastructure could lose value through two different paths. Flexible, general-purpose graphics processors are useful while algorithms keep changing. If workloads stabilise, cheaper and more energy-efficient custom chips could replace them. If AI fails to gain widespread use, insufficient demand could leave the same equipment underused. Avoiding both risks would require adoption to advance fast enough to generate demand, but slowly enough to keep existing hardware useful, with enough differentiation to prevent price competition from eroding returns. The study does not measure how wide that corridor is or assume it will last for a six-to-ten-year asset life. Its framework combines depreciation over time with losses from low utilisation or obsolescence, which can also affect unused hardware. For a $457.1 billion claim stack, including $288.0 billion in leases that have not begun, a stress test requires annual revenue of $753.9 billion over two years, compared with $136.9 billion over twenty years. These are gross-margin-based bounds, not payment schedules or forecasts; the longer horizon is close to lease terms of fifteen to nineteen years. The paper argues that investors may price early demonstrations as widely adopted products, even though adoption takes time that additional capital cannot simply remove. Supply expansion across four tracks, electricity-grid queues, permits and public consent could extend the wait for final consumer demand.