See useful life.
Useful life The years over which an owner depreciates hardware, writing it off on its books, five to six years on the operators’ own filings. The lenders’ clock, the term over which the hardware is paid for, is a different clock. Chapter 4.
From The AI Compute Race. Read the book
Sources that use it
- Chase Lochmiller (co-founder and chief executive, Crusoe), interviewed by Harry Stebbings, “Crusoe CEO: Why Everyone Gets GPU Depreciation & AI Energy Costs Wrong,” the publisher’s title, though everyone is the host’s word and the tape gives no cost of energy, 20VC, published 3 October 2026 (every figure is his own account of a private firm that files no accounts, and no contract or depreciation schedule is shown; at 35:24: “we are renting capacity on a long term basis”, “call it five year contracts to” “credit quality customers”; at 41:54: “we use a six year depreciation cycle” and “that’s sort of like the standard across the industry”; at 45:38: “earlier stage startups that are being asked to commit to compute in 2028” and “that is an eternity from now”; at 17:54: “we had sort of committed to doing this in one year and the next closest bid was two and a half years”; at 18:34, of a medium-voltage power distribution center: “the lead time for this one component was 100 weeks”, and by his account his own team then made it in 28 weeks; at 25:46 and 26:00: “use pretty close to zero water” and, of one of the first eight buildings at Abilene, “about the same water annually as about 10 single family homes”, with no volume given; at 27:01 and 27:18: “typically energy prices for communities have come down” and “more megawatts being amortized over the same transmission and distribution infrastructure”, with no market or study named; at 28:52 and 29:07: “we’re going to account for more than a third of the tax revenue in Taylor County and Abilene, the town” and “we’re more than doubling the tax receipts that are going to the schools”, with no abatement mentioned on the tape; at 43:14: “will ultimately, we believe, extend the depreciation cycle beyond six years”; at 43:40 and 43:54, of the prices charged for Hopper chips and with no price given: “higher than the rates that were being charged three years ago” and “payback very, very quickly”; decoded at two models). In the Sources
- Azeem Azhar, William Gildea, Hannah Petrovic, Nathan Warren and Marija Gavrilov, The State of the AI Economy, Exponential View, 25 June 2026, 66 pp. (a bottom-up, deduplicated, confidence-graded model of generative-AI revenue across more than 1,000 firms; “$110bn trailing 12-month revenues – now at a $175bn pace”, p. 7, the pace defined in the chart legend as “Annualized run rate (last month x 12)”, global ex-China, excluding chips; the open-weight model layer’s “Value add: $0 (revenue counted in hosting)”; p. 38, headroom after chip depreciation, Q1 2026, “% = (Revenue – Depreciation) ÷ Revenue”, infrastructure revenues only: −12% on a four-year schedule, 7% on five and 19% on six, “At standard 6-year chip life, Q1 2026 infrastructure revenue headroom is 19%”, and “If chip life is shorter, revenues don’t repay CapEx (this would require initial H100 purchases becoming obsolete today)”). In the Sources
Related terms
From the glossary of The AI Compute Race.
How to cite this page
Dryburgh, Lee S. “Depreciation.” In The AI Compute Race. DRYBURGH. https://dryburgh.com/vol1/glossary/depreciation (accessed 2026-11-01).
BibTeX
@misc{dryburgh-depreciation-2026,
author = {Dryburgh, Lee S.},
title = {{Depreciation}},
howpublished = {In \emph{The AI Compute Race}, DRYBURGH},
year = {2026},
url = {https://dryburgh.com/vol1/glossary/depreciation},
note = {Accessed 2026-11-01}
}