GPU Rental Rates by Generation: Nine Years of Posted Prices
The age-curve note measures one date's prices across generations of different ages. This note measures the other axis: each generation's own posted rental rate through time, reconstructed from dated captures of providers' own pricing pages — 1,430 archived on-demand observations, October 2017 to April 2026, every row carrying the archive URL it was read from. The reconstruction ends at the archive's edge; CCIR's live panel is the continuation from there. Three facts sit on the surface of the data: posted rates decline in steps, set by the minority of providers who reprice; demand episodes have twice interrupted the decline — in 2021–2023, and again from late 2025, when the AI buildout pushed posted rates up across every generation still listed; and no generation's realized path matches the smooth decline a single-date cross-section suggests.
01 The Paths
Fitted annual rates of change on the panel median, months with n≥3 only (the endpoint column is a compound annual growth rate, CAGR):
| Generation | Window (n≥3) | Log-linear fit | Less CPI | R² | Endpoint CAGR | Median, first → last |
|---|---|---|---|---|---|---|
| V100 | Jun 2019 – Apr 2026 (6.8y) | −21.1%/yr | −25.0%/yr | 0.81 | −16.1%/yr | $2.69 → $0.81 |
| A100 | Sep 2021 – Apr 2026 (4.6y) | −7.0%/yr | −11.2%/yr | 0.49 | −2.2%/yr | $2.20 → $1.99 |
| H100 | Sep 2023 – Apr 2026 (2.6y) | −13.0%/yr | −16.1%/yr | 0.48 | +1.5%/yr | $3.50 → $3.64 |
Three generations, three different decline rates, none of them smooth — the R² column is as much of the finding as the slopes. V100, the one generation observable from mid-life to end-of-life, declined fastest and most regularly. A100's posted median is $2.20 in September 2021 and $1.99 in April 2026 — −2.2% per year endpoint to endpoint, across four and a half years that also contained the 2022 peak ($3.05) and the 2025 trough ($1.75). H100's posted median ends the window higher than it began it.
All prices are nominal, as posted. The "less CPI" column subtracts each window's consumer price index (CPI) inflation rate — the CPI for urban consumers (CPI-U) series (3–4%/yr over these years) — from the fit — which matters most for A100: a path that is nearly flat in posted dollars is a decline of roughly 6% per year in real terms, endpoint to endpoint.
02 Two Demand Episodes
The decline has been interrupted twice, visibly, in nine years.
2021–2023. Posted rates on then-current and prior generations firmed through the GPU shortage and the first phase of the AI buildout. V100 — four to six years old — rose 48% at DataCrunch and was repriced upward at OVHcloud; A100 rates at the repricers rose through 2023 into early 2024, and Lambda raised its A100 list prices outright in January 2024. The episode ended in mid-2024: A100 and V100 posted rates stepped down 30–56% from their peaks over the following year.
Late 2025 – the window's edge. The declines bottomed around July–August 2025. From roughly December 2025, posted rates turned up across every generation still listed, at every provider that reprices — driven by demand from the AI buildout: H100 +29% at Verda by December, H200 +70% by March; and in spring 2026 Lambda raised A100 by 56% in seven weeks and repriced V100 — silicon approaching its ninth year — from a list unchanged since June 2022. The sticky tier, as in the first episode, has not moved. This is the sharpest broad repricing of prior-generation silicon in the record.
03 Cohort and Cross-Section
The ≈−26%/yr figure in the age-curve note and the slower rates in §01 measure different things, and the difference is in the data rather than between two opinions. The cross-section compares differently-aged chips at one date: how the market prices age, today. A cohort path is one chip through time, and includes everything that happened to the whole market level along the way — including the two demand episodes above. A generation that ages through a firming market declines more slowly than the cross-sectional age gradient; V100, most of whose life predates the boom, is the one cohort that tracks it closely. Both series are published; which one a reader needs depends on the question being asked of it.
04 Limitations
Posted asks, not transactions. Panel composition changes as providers enter, exit, redesign pages, or stop rendering prices server-side; the carry-forward repair and per-point counts are disclosed, and thin months are excluded from fitted figures. A100 pools memory variants. Archive coverage is not uniform — notable gaps: RunPod from mid-2025, Oracle Cloud Infrastructure (OCI) after 2023, OVHcloud's US-dollar pages in 2025. Step dates are bounded by capture dates, not observed exactly: a raise "in spring 2026" occurred between two dated captures. Methodology per the published documents; the underlying observation file, one archive URL per row, is available on request via /contact.
Methodology & data — sources, filters, repair, basis
Sources. Archived captures of provider pricing pages from the Internet Archive's Wayback Machine and from Common Crawl web-archive (WARC) records. Providers: Amazon Web Services (AWS), CoreWeave, DataCrunch (Verda from December 2025), Genesis Cloud, Lambda, OCI, OVHcloud, Paperspace, RunPod, with Common Crawl recoveries of Crusoe, Hyperstack, Together, and TensorDock. Every archived observation is a price the provider itself posted, on a dated page, with the capture URL recorded. The reconstruction ends at April 2026 — one uniform right edge for every generation, all from the same measurement system. CCIR's live panel, which tracks the same providers daily, is the continuation from there; it is deliberately not spliced into this dated note.
Filters. Guaranteed on-demand rates only — reserved, interruptible (hyperscaler "Spot" products and community tiers alike), and "starting-at" teaser rows excluded; USD only; per-instance prices divided to per-GPU with the derivation recorded per row; known page bugs excluded.
Repair. Posted prices are step functions: a page that read $1.79 in January and $1.79 in June did not move in between. Gaps between captures are carried forward at most six months; providers drop out of the panel beyond that. The monthly panel value is the median across providers present that month, with the count disclosed per point. Months with fewer than three providers are drawn dashed and excluded from every fitted figure; trailing thin months at the end of a series are truncated rather than drawn.
Basis notes. CoreWeave's componentized per-GPU rate card (constant from 2021) was withdrawn in November 2025; its instance-derived prices are a different billing basis and are treated as a separate series. A100 pools 40GB and 80GB memory variants at chip grain, matching the published series convention.