🌐 ▾简中繁中EN日本語한국어

GHPI: GPU Hourly Price Index

Version 1.0 · Issue GHPI-2026-10 · Updated: 2026-10-31 18:45
One-liner: GHPI is a public price index representing the hourly rental cost of one GPU card, benchmarked to the H100 80GB, with base period = 100, calculated usingChained matching methodMethodology — New platform listings are not counted as price reductions; GPU model out-of-stock events are not counted as price increases. Current lowest in-stock price for H100 80GB across the market. ¥8.28 / GPU·hour, index 100.00(2026-10-31)。
Benchmark GPU
H100 80GB
Pay-as-you-go, in-stock, per-GPU-hour price
Sample
29 providers
28 GPU models · 34-day series
License
CC BY 4.0
Attribution required; commercial use permitted

Current index

IndexMarket’s lowest in-stock price ¥/GPU·hourIndex (baseline = 100)Daily ChangeDays Since Base PeriodNumber of in-stock quote pairs
GHPI-H1008.28100.00+0.00%+0.00%10
GHPI-CN5.15100.00+0.00%+0.00%7
GHPI-Domestic1.96100.00—+0.00%1
Two distinct metrics—don’t conflate them:「市场最低In架价」是当日全Sample实测最低值,会随新Platform进入而下降; 「Index」才是可比ofpriceChange指标,用Chained matching method剔除Coverage expansionand缺货影响。

GHPI-H100 Trend

GHPI Trend10-1310-2210-31

Data Quality (including excluded items)

Successfully Captured
144
Total quoted prices currently in-stock this round
Participating Platforms
96
Passed scope filtering
Meeting index scope criteria
96 / 144
Excluded 48 listings failing scope criteria
Exclusion Details:stale(expired) 7 listings · suspect(abnormal volatility) 0 listings · seed(无公开价目页of人工登记Reference Price)2 entries · 非Pay-as-you-go或缺货 39 entries。PlatformCoverage 22/29 platforms successfully scraped in this round.
⚠️ The “sample validity rate” here refers toIndex Scope(calculated based on pay-as-you-go + in-stock + real-time) — differs from the site’s /sources 公示of「Real-time crawl success rate 93.8%」不是同A指标 —— 后者只判断是否本轮抓到,不额外Requirements「Pay-as-you-go」and「In架」。两个数都公开,避免误读。

Methodology (reproducible)

sampleScraped every 45 minutes from platforms’ public pricing pages / public APIs; daily value is the last round of that day
filterOnly pay-as-you-go (ondemand) + in-stock (available) quotes; excludes stale / suspect (expired or questionable) and seed (manually entered reference prices without public pricing pages)—consistent with the site’s lowest-price ranking methodology
normalizationFull-system price ÷ GPU count normalized to per-GPU-hour price
currencyCNY, exchange rate: 1 USD = 7.2 CNY
unit_observation(Platform, GPU model) price pairs; for multiple listings of the same platform and GPU model, only the lowest price is retained
index_methodChained matching method (matched-model chained): For two consecutive days, only quote pairs present on both days are retained; daily price ratios are computed for each pair and their median is taken; daily medians are then chained via multiplication; base period = 100
why_chained① When a new platform launches, it lowers the ‘lowest price across all platforms’, but without proper matching, this coverage expansion may be misread as a price drop (real-world test: Paperspace’s integration on 2026-10-20 caused a 48.9% phantom decline); ② If grouped by GPU model family, a single stockout event gets permanently locked in via chain-linking (real-world test: GHPI-CN was inflated to 140.4 due to this). Switching to quote-pair units eliminates both issues.
min_base_pairsBaseline period requires ≥5 quote pairs; when new series lack sufficient samples, baseline is set to the actual maximum across the full series, with base_pairs disclosed for reader assessment of reliability
min_match_pairsIf the intersection between two consecutive days contains fewer than 3 matched quote pairs, no day-on-day comparison is calculated for that day; the index value is carried forward.
caveatListed price ≠ transaction price—always refer to the platform’s order page; index tracks basket price changes, not average transaction price

Index definition: Let U denote the set of in-stock quote pairs on day tt(per个报价对 = Platform × GPU model), C = Ut ∩ Ut-1 为两日都In架of报价对,则 GHPIt = GHPIt-1 × mediank∈C( pt,k / pt-1,k ), 基period日 GHPI = 100。Index下降 = 同一批报价变便宜。

Two pitfalls to avoid(All verified through real-world testing):
① Coverage expansion:新Platform接入会拉低「全PlatformLowest Price」。实测 2026-10-20 Paperspace (United States)接入, H100 lowest price ¥16.2 → ¥8.28,若不做匹配会算出「21 day跌 48.9%」of假信号 —— 那不是市场降价, 是我们多接了一家便宜Platform。
② Stock-out contamination:若以「GPU model组」为统计单元,某张便宜卡缺货会让整组Lowest Price跳变, 而链式法会Replace这time跳变永久固化。实测 GHPI-CN 曾因此被一time性抬到 140.4 再也回不去。 改用 (Platform, GPU model) 报价对作为单元后,缺货只是剔除那A报价对,不污染整体。

Citation:
Chinese GPU Compute Radar (backed by Chainfir Capital). GHPI: GPU Hourly Price Index [EB/OL]. 2026-10-31. https://rentgpu.link/ghpi
English GPU Price Radar (Invested in by Chainfir Capital). GHPI: GPU Hourly Price Index [EB/OL]. 2026-10-31. https://rentgpu.link/ghpi

Daily Index (Permanent URL)

DateH100 lowest price ¥Lowest price on Chinese platforms ¥Lowest price on domestic GPUs ¥GHPI-H100 Index
2026-10-318.285.151.96100.00
2026-10-308.285.15—100.00
2026-10-298.285.15—100.00
2026-10-288.285.15—100.00
2026-10-278.285.15—100.00
2026-10-268.285.15—100.00
2026-10-258.285.15—100.00
2026-10-248.285.15—100.00
2026-10-238.285.15—100.00
2026-10-228.285.15—100.00
2026-10-218.285.15—100.00
2026-10-208.285.15—100.00
2026-10-1918.005.15—100.00
2026-10-1818.005.15—100.00
2026-10-1718.005.15—100.00
2026-10-1618.005.15—100.00
2026-10-1518.005.15—100.00
2026-10-1418.005.15—100.00
2026-10-1318.005.15—100.00
2026-10-12—5.15——
2026-10-11—5.15——
2026-10-10—5.15——
2026-10-09—5.15——
2026-10-08—5.15——
2026-10-07—5.15——
2026-10-06—5.15——
2026-10-05—5.15——
2026-10-04—5.15——
2026-10-03—5.15——
2026-10-02—5.15——
2026-10-01—5.15——
2026-09-30—5.15——
2026-09-29—5.15——
2026-09-28—6.98——
This site has received Chainfir Capital 投资(Invested in by Chainfir Capital), 由专职团队运营,不是个人站点。Prices are scraped from each platform’s public pricing page.Prices on the order page apply; 已排除 seed / stale / suspect item目。 OpenData见 /data。
🐆 Leopard · AI Customer SupportPowered by Tongyi Qwen
Hello there, I’m Leopard~ Feel free to ask about GPU pricing, cost-effective rental strategies, or what Pro membership offers.