A Test That Cannot Fail

Resale prices and rental rates both agree with the six-year life cloud operators give AI accelerators. They are one measurement, not two — and it is being taken inside a shortage that suppresses the only signal that could disprove it.

The Ecliptic · Issue No. 012 · October 1, 2026

45–55%

of original value retained by a three-year-old H100 on the secondary market, against a straight-line book value of 50%

+33%

year-on-year rise in H100 spot rental pricing in August 2026, in the chip's third year, while newer parts shipped

$1.00–$11.07

per GPU-hour: the H100 on-demand range across 53 providers on a single day, September 8, 2026

FY2028

the vendor's guided end of the supply constraint; a 2023 H100 reaches the end of a six-year life in 2029

The apparent verdict

The market agrees with the accountants. That is the problem.

Straight-line six-year book value (no salvage) against reported H100 resale value at two and three years, with the 2029 end of life for a 2023 unit and the vendor's guided shortage through fiscal 2028. Resale figures as cited in PAR-0015; the book line is arithmetic. Not illustrative.

Straight-line six-year book value is **66.7%** at year two and **50%** at year three. Reported resale is about **61%** at two years and **45–55%** at three — a match at the midpoint of the assumed life.

A used accelerator's price is the discounted value of its rental stream, so resale and rental are not two confirmations; they share every input.
Used H100s traded as high as $50,000 in the mid-2024 scarcity, above the $25,000–$40,000 retail range. A used price above new measures availability, not durability.
Capacity rebooking at 95% of original pricing is a contract renewal, not a spot clear — evidence about commercial exposure, not hardware life.

The economics

Depreciation is not the binding term — the datacenter slot is.

Annual gross rental revenue per used H100 by provider tier at the midpoint of each September 2026 tier range, against a six-year straight-line charge on a $30,000 new card. Parameter estimates: 70% utilisation is an assumption, revenue is gross of power, facility and operator margin.

On a **$20,000** used H100 at 70% utilisation, gross rental revenue covers a **$5,000** annual six-year charge **~2.7×** at marketplace rates, **~4.8×** at specialist clouds and **~10.9×** at hyperscaler on-demand (Parameter estimates).

The same part priced from $1.49–$2.99/hr on marketplaces to $6.88–$10.98/hr on hyperscaler on-demand: residual value is set by the owner's channel more than by the chip's age.
A three-year-old chip whose rate is rising while newer parts ship is pricing the power, cooling and place in a building, not the silicon.
Silicon turns on a roughly two-year architecture cycle; a datacenter shell lasts 15–20 years. One blended six-year figure averages two assets.

What would settle it

Three disclosures, none of them a monthly rate print.

The six-year life is **not contradicted** by the market. It is not confirmed either, because the market cannot currently contradict it.

A rental rate that falls while utilisation stays high — the only clean obsolescence signal, and guided to arrive no earlier than after fiscal 2028.
Contracted versus spot revenue, disclosed separately — the life is robust on a contracted book and fragile on a spot-exposed one.
Silicon and facility depreciated separately — Amazon's 2025 move to shorten a subset of servers from six years to five is the disclosure most consistent with this reading.

The Ecliptic, Issue No. 012, cover analysis, drawn from Parameter PAR-0015 (September 8, 2026). Resale values, rental rates, tier ranges, the 33% year-on-year rise and vendor supply guidance are as reported in that report's cited sources; the per-tier revenue multiples and the 70% utilisation are Parameter estimates and assumptions. Neither chart is illustrative.

Content is produced by AI systems with human editorial oversight.