A data center's sticker price depends heavily on one accounting guess: how many years before the chips inside are obsolete?

This explores how much an AI data center's cost depends on one accounting choice: how many years you assume the hardware lasts before it has to be paid off and replaced. The corpus doesn't contain a sensitivity analysis that varies the depreciation period directly. It does show why this assumption matters more than almost anything else. Epoch AI's cost model for a one-gigawatt facility puts the annualized cost at about $8.5B. Servers account for roughly $5B of that, or 60%, while electricity comes to only about $0.6B What actually drives AI data center costs?. Server cost here means capital recovery: the purchase price spread over the hardware's assumed life. So if most of the bill is hardware divided by years of use, the number of years you choose largely sets the bill.

The arithmetic follows directly. Shorten the assumed life of a GPU fleet from five years to three, and the yearly hardware charge rises by about two-thirds. Under Epoch's cost mix, that pushes the whole facility's annual cost up by tens of percent. A similar swing in electricity prices would barely show up. Popular debate about AI infrastructure tends to focus on power, but by this model the more fragile number is the depreciation schedule.

Why might that schedule be too optimistic? Other notes in the collection point to fast obsolescence. Denain argues that the cost of reaching a given capability level falls roughly 5–10x per year, partly because of better hardware Does inference cost remain a persistent economic burden?. Epoch also finds that the price of a given level of AI performance drops fastest right after it becomes state of the art: about 66% per quarter, before slowing Do AI performance prices fall faster when capabilities debut?. When the market price of what your chips produce falls that fast, the chips can lose their economic value well before they physically wear out. A five-year accounting life can hide a much shorter useful life in practice.

The same logic appears at smaller scale. A 54-scenario model of self-hosting LLMs finds payback ranging from a few months for small models to more than five years for large ones When does self hosting an LLM beat commercial API costs?. At the long end of that range, the payback period may be longer than the hardware stays competitive. Whether self-hosting pays off then depends on the same depreciation question. One related point: if dollar-cost figures go stale this quickly, benchmarks and procurement decisions should rely on durable units like token counts rather than prices fixed at one moment Should benchmarks report token counts instead of frozen costs?.

The takeaway is that the collection establishes the structure: hardware dominates the cost, and hardware value erodes quickly. It does not give the exact sensitivity curve. If you want to know whether the economics of AI infrastructure hold up, the useful question is less about the price of power and more about how many years the chips can earn their keep.