How Long Do GPUs Last Anyway? (appliedconjectures.substack.com)

The piece examines a flashpoint in AI infrastructure accounting: whether hyperscalers’ assumed useful lives for datacenter GPUs are realistic. Skeptics argue a 2-year replacement cadence (driven by rapid Nvidia generations) would force much higher depreciation and cut 2024 EBITDA by an incremental ~7–22% for most hyperscalers, materially changing their valuations. That matters because shorter useful lives would make AI capex look far less profitable, reshaping demand forecasts for datacenter GPUs and investor views of Amazon, Google, Microsoft and Meta as bargains. Empirical evidence and unit‑economics analysis weaken the 2‑year claim. Secondary market prices for T4, V100 and A100 parts have dropped from launch but often stabilize rather than collapse; rental markets and IT asset‑disposition channels remain active, with legacy GPUs still rented and resold to smaller firms, researchers and constrained geographies. Simple unit‑economics using conservative assumptions (3‑year useful life, PUE ~1.20, $0.10/kWh, DGX‑1 V100 MSRP ~$150k) show legacy cards can generate attractive gross profits even before considering economies of scale and high utilization. Export controls and niche use cases (edge, inference, non‑hyperscaler training) further sustain value. The bottom line: prior‑gen GPUs are far from worthless, so hyperscalers’ depreciation policies are plausibly defensible and the economic life of GPUs extends beyond a strict product‑cadence timeline.