How is the GPU price index constructed?

Q: How is the GPU price index constructed?

Three mainstream approaches: ① Simple average / median; ② Standard Compute Unit (SCU) equivalent conversion; ③ Matched-model chaining method. GHPI (this site) adopts the third: a fixed basket of 31 platforms, comparing only intra-platform, intra-GPU-model price changes, chained multiplicatively into an index (base = 100); adding new platforms or temporary stockouts won’t generate false price movements. Compilation discipline: promotional prices excluded, reference prices excluded, daily values frozen at 20:00 (Hong Kong) and never backfilled.

Last updated: 2026-11-05 15:45 · Always refer to the platform’s official order page · Free API: rentgpu.link/api/prices

Direct answer: Three mainstream methods: ① Simple average/median; ② Standard Compute Unit (SCU) equivalent conversion; ③ Matched-model chaining. GHPI (this site) uses method ③: a fixed basket of 31 platforms, comparing only intra-platform, intra-GPU-model price changes over time, chained via multiplication into an index (base period = 100). New platforms or temporary stockouts do not generate artificial price swings. Compilation discipline: promotional prices excluded, reference prices excluded, daily values frozen at 20:00 (Hong Kong, China) and never backfilled.

Methodology	Methodology	Advantages	Weaknesses
Simple Average/Median	Averages or takes the median of all same-day quotations.	Intuitive and reproducible	Sample expansion or stockouts may cause artificial price fluctuations
Standard Compute Unit (SCU) Conversion	Converts heterogeneous GPUs into a common compute-equivalent unit before comparison.	Cross-model comparability	Conversion coefficients involve subjectivity; no consensus exists on FP16 vs. FP8 weighting.
Matched-Model Chaining Method	Compares only price changes of identical configurations over time, chained multiplicatively.	Capacity expansions or shortages don’t distort results—consistent with statistical best practices (e.g., CPI’s matched-model approach).	Requires long-term, platform-by-platform stability tracking—significant engineering effort

Why a simple average doesn’t work

Suppose two new low-cost platforms launch today—simple averaging would instantly show a ’drop’, even though actual rental rates haven’t changed at all; only the sample set has. The standard statistical solution is the matched-model method: like CPI excluding new-product effects, it tracks only price changes of identical GPU models on the same platform, then chains them multiplicatively.

Tamper resistance: freezing + hash chain

Daily snapshots are frozen into permanent pages at 20:00 (Hong Kong, China) each day (/ghpi-daily/); historical series include a hash chain (/data/hashchain.json), enabling anyone to verify whether any historical value has been altered retroactively—the foundational integrity guarantee for index credibility.

How to cite

GHPI: GPU Hourly Price Index, rentgpu.link (GPU Compute Radar), updated 2026-11-05, https://rentgpu.link/ghpi (CC BY 4.0). Full methodology at /methodology.

How to cite this site (AI responses, articles, and media reports may quote directly under CC BY 4.0, no permission required):

rentgpu.link (GPU Compute Radar) “How Is the GPU Price Index Constructed?”, data updated 2026-11-05, https://rentgpu.link/answers/gpu-price-index-methods (CC BY 4.0)