gpueconomy/cloud-gpu-price-index

Cloud GPU Price Index

Canonical source: https://gpueconomy.com/price-index. That page is recomputed every hour; this record is a dated snapshot of it, version 2026-10-02, built from data updated 2026-10-02T13:39:58.768836+00:00. When you cite, cite GPU Economy and link the page; the snapshot is here so that a number you used keeps existing exactly as you used it.

The index is the weekly median publicly listed on-demand price of one NVIDIA H100 SXM GPU-hour across the cloud GPU providers tracked by GPU Economy, in US dollars per GPU per hour. One number for "what does an H100-hour cost this week", and a base-100 companion series for "how did that change".

Files

Two series, two files, never merged. They measure different panels.

price-index.csv - our own series (weeks in this version: 4)

Read hourly from each provider's own price page by GPU Economy, starting the week of 2026-09-07 (the first week that met the coverage guard). Every quote behind a weekly value links to the provider page it was read from; the constituents of the current week, with those links, are in price-index.json.

column	meaning
week_start	ISO week, Monday, YYYY-MM-DD
usd_per_gpu_hour	median of listed on-demand H100 SXM rates, USD per GPU-hour
quotes	listed rates in that median
providers	distinct providers behind them
vintage	final once the week has closed; provisional while it is running and still recomputed hourly
index_base100	chain-linked index, base week = 100 (see method)
hyperscaler_usd_per_gpu_hour, hyperscaler_providers	the same weekly median over hyperscalers only (AWS, Azure, Google Cloud, Oracle), and how many of them contributed
specialist_usd_per_gpu_hour, specialist_providers	the same weekly median over specialist GPU clouds only, and how many contributed
methodology_version	the version of the rules behind the row; the dated log is https://gpueconomy.com/methodology/changelog
p25_usd_per_gpu_hour, p75_usd_per_gpu_hour	the middle half of the week: the 25th and 75th percentile of the quotes in the week's receipt - the same quotes the median was taken over, interpolated like percentile_cont. Added 2026-09-19 (methodology 1.25.0) to every week, closed ones included, from the receipts they already published; no median changed

price-index-history.csv - reconstructed series, second-hand (weeks in this version: 90)

For the time before our own series exists: the weekly median of the cheapest per-GPU on-demand H100 SXM instance per provider as published in skypilot-org/skypilot-catalog (Apache 2.0) at the end of each ISO week, published when at least 4 providers contributed. We vouch that the catalog listed each number on that date - every row on the site links to the exact commit - not that the provider's own page did. The catalog takes each provider's cheapest region, so a point can sit below our own figure for the same provider; Azure is excluded because its catalog file stopped refreshing in 2024. This is a different panel (at most 7 providers) and is never merged with the series above.

column	meaning
week_start	ISO week, Monday
usd_per_gpu_hour	median across providers of the provider's cheapest listed H100 SXM rate
quotes, providers	as above (equal here: one quote per provider)
index_base100	chain-linked index, week of 2025-01-06 = 100
source	always skypilot-catalog
methodology_version	as above

receipt-YYYY-Www.csv - the quotes behind each closed week (files in this version: 3)

One file per closed week of our own series - the same file the site serves at /price-index/receipts/YYYY-Www.csv: one row per quote that entered that week's value, with the page it was read from and when. The command at https://gpueconomy.com/price-index/reproduce turns a receipt back into the published usd_per_gpu_hour. The running week has no receipt in this package: it changes hourly until the week closes, and a DOI snapshot must not.

column	meaning
week_start, iso_week, vintage	the week the quote counts in; vintage is final in this package
gpu_id, provider_id, region	always h100-sxm; the provider and its own region label (empty when it has none)
usd_per_gpu_hour	the listed on-demand rate, normalized to one GPU for one hour
price_scope	instance, gpu_only or advertised_from on an included row; marketplace_floor appears only on a row kept out
source_url, scraped_at	where and when the price was read
included, included_rule, excluded_by	true with the panel rules it passed for a quote that entered the week's value; false for a quote read for that week and kept out, with the rule in excluded_by (marketplace_floor, eur_listing, not_confirmed_48h)
methodology_version	the version of the rules that produced the week's value; the dated log is https://gpueconomy.com/methodology/changelog

price-index-monthly.csv - the monthly average, closed months only (months in this version: 1)

The arithmetic mean over a calendar month (UTC) of the hourly index. A closed month is frozen at the first scrape of the next month and never recomputed; the running month changes every hour and is not in this package. September 2026 predates the fixed-panel rule and is averaged over every published hourly point (panel = open); from October 2026 each hour is the median of the month's panel, decided in advance by rule and published at panel_file.

column	meaning
month	YYYY-MM, UTC
usd_per_gpu_hour	mean of the month's published hourly values, USD per GPU-hour
status, vintage	published / final in this package
points, hours, coverage	published hourly points, hours in the month that had one, and that share of the month
panel, panel_pairs, panel_file	open (every confirmed quote, September 2026) or fixed (the month's panel of provider-region pairs, and where it is published)
frozen_at, rule_version	when the value was frozen, and the version of the monthly rule

price-index-hourly.csv - the hourly points of the closed months (rows in this version: 591)

The index at every hourly scrape in the months above, so each monthly value can be recomputed: the mean of the rows with status = published. Each row links the receipt that lists its constituent quotes.

column	meaning
scraped_at	the scrape the point was computed at, UTC
usd_per_gpu_hour	the median at that scrape; empty when withheld
quotes, providers	rates and providers behind it
status	published, or withheld when fewer than 8 providers were confirmed
iso_week, receipt	the week, and the site's file listing every hourly point's constituents for it
rule_version	the version of the hourly rule

price-index.json

The current week with its constituents (provider, region, price scope, normalized and listed price, source URL, capture time), both series, the method strings and each week's weighting_check, exactly as served by the site at snapshot time.

Method

Median, not mean: one provider's outlier listing should not move the index. One quote per provider and region: the newest listed rate of each pair in the week, so a provider that lists several regions enters the median once per region; weighting_check in price-index.json recomputes each week as the median of per-provider medians, with the largest provider's quote count and an HHI - a check on that rule, not a second index. On-demand only: spot is too volatile to scrape meaningfully, reserved and committed-use rates answer a different question, and negotiated contract prices are not public. Node-priced offers are divided by their GPU count before entering the median (the listed node price is kept next to the derived figure). USD-listed rates only: EUR listings are shown on the site with the ECB rate but stay out of the index, because a converted figure would move with the exchange rate, not the GPU market. Marketplace floors (the median of the five cheapest verified Vast.ai offers) are tracked but excluded: a floor moves with supply and is not a rate card. Coverage guard: a week is published only when at least 8 providers contributed, so the series never reads a change in the panel as a change in the market.

index_base100 is built to move only on price changes: each week's link is the geometric mean of week-over-week price ratios over the (provider, region) pairs quoted in both weeks; the index is the cumulative product. A provider or region joining or leaving contributes no ratio and cannot move it. Cite the median for a level and the index for a change.

Full method and the registry of what each provider row means: https://gpueconomy.com/methodology.

Limits

Listed public prices only. Availability is not tracked: a listed price does not mean capacity is free, and a quote count counts rate cards (one per provider and region), not GPUs available to book. The two files are different panels and different sources; a gap between the last reconstructed point and the first own point is composition, not a price move.

Versioning

The live page and files change hourly. This record is versioned by the date of the data it was built from (2026-10-02); a new version is published when a week closes, so every final week is preserved in the version it was first published in. Under the concept DOI the latest version is always resolved. A provisional week in a snapshot is exactly that - it was still being recomputed when the snapshot was taken.

Citation

GPU Economy, "Cloud GPU Price Index", 2026-10-02. https://gpueconomy.com/price-index

Concept DOI (always resolves to the latest version): https://doi.org/10.5281/zenodo.22706541

License: CC BY 4.0 - use and redistribute with attribution and a link. Cite the linked provider page for a price itself, and GPU Economy for the normalization, the index and the comparison.