# Depreciation of Compute Assets An accelerator begins losing value the moment it is purchased, and the rate at which it does so governs the entire economics of a compute facility. The clock is set by successors rather than by wear: the hardware works for a decade and stops being the best use of a square foot of floor and a kilowatt of power long before that. > [!note] The one-sentence version > The machine does not wear out; it is outcompeted, and the floor and the power it occupies are worth more to its successor. ## What determines useful life **Performance per watt of the next generation.** When a new accelerator delivers substantially more work per watt, the older device's claim on power and floor weakens. In a facility where power is the binding constraint, replacement is driven by opportunity cost rather than by failure. **Memory capacity and bandwidth.** Model sizes grow, and a device whose memory cannot hold current models economically becomes unsuitable for frontier work regardless of its arithmetic throughput; see [[wiki/Memory Wall|Memory Wall]]. **Software support.** The practical death of hardware is the end of driver, compiler, and framework support, after which running current software becomes an engineering project; see [[wiki/CUDA|CUDA]] and [[wiki/ROCm|ROCm]]. **Interconnect generation.** Accelerators are bought as clusters, and a device that cannot join the current fabric cannot participate in the current scale of training. **Actual reliability.** Silicon and memory do degrade under sustained thermal and electrical load, and large fleets see failure rates rise with age, which adds operational cost beyond the performance argument. ## The accounting and the argument about it Operators assign a useful life for depreciation purposes, commonly in the range of five to six years for data center computing equipment, and this choice has large effects on reported earnings. It has also become a genuine point of contention: the assumption is defensible if hardware remains productive on a long tail of work, and aggressive if frontier hardware is effectively obsolete for its original purpose in three. The honest reading as of October 2026 is that both can be true at once, because a device retired from frontier training may have years of useful service left in inference, development, or research work; the question is whether the second-hand use is genuinely economic at the power price the facility pays. ## The cascade, and why it is good news Hardware moves down a ladder rather than off a cliff: frontier training, then large-scale inference, then development and research, then academic and smaller commercial use, then specialist and hobbyist markets. Each step broadens access. The same cascade is visible in semiconductor manufacturing, where leading-edge equipment retired from an advanced fab is purchased and run profitably for decades producing foundational chips, which is why a component costing cents can be made on a tool that once printed processors. That is the pattern worth noticing: the short depreciation clock at the frontier is what makes capable compute cheap for everyone else a few years later. The expensive machine's second career is somebody's first opportunity, and the pace of replacement at the top sets the pace at which capability reaches the rest of the country. ## Relationships - **drives:** [[wiki/Capital Intensity|Capital Intensity]] · [[wiki/Utilization|Utilization]] · [[wiki/Construction Sequencing|Construction Sequencing]] - **determined by:** [[wiki/Accelerator|Accelerator]] generations · [[wiki/Memory Wall|Memory Wall]] · [[wiki/CUDA|CUDA]] · [[wiki/ROCm|ROCm]] - **cascades into:** [[wiki/Neocloud|Neocloud]] · [[wiki/Edge Inference|Edge Inference]] · [[wiki/Silicon Wafer|Silicon Wafer]] equipment reuse - **civic dimension:** [[wiki/National Compute Capacity|National Compute Capacity]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** obsolescence driven by successor performance per watt, memory capacity, software support, and interconnect generation rather than by wear; rising failure rates in aging large fleets; the cascade of hardware from frontier to broader use; hand-me-down semiconductor equipment economics, in which tools retired from advanced fabs produce foundational chips profitably for decades. - **Established (datestamped):** as of October 2026, useful lives commonly assigned to data center computing equipment fall in the range of five to six years, and that assumption is actively disputed among analysts. - **Unresolved:** the true economic life of frontier accelerators; the deciding evidence is secondary-market pricing and retirement data across a full generation cycle. ## Sources - [How AI Is Pushing the Semiconductor Supply Chain to the Limit](https://www.bloomberg.com/) — Bloomberg Originals, October 2026; hand-me-down equipment economics in foundational chip manufacture. - Operator financial reporting on data center equipment useful lives and analyst commentary, as of October 2026.