The public dataset

Everything on this site is also a file. 12 files, built with the site from the same data the pages show, so a number here is the number on the page. All of them are published under CC BY 4.0: use, copy and redistribute with attribution to “GPU Economy” and a link to the page or file you used. Cite the linked source for a price itself, and cite us for the normalization, the index and the comparisons. No key, no rate limit.

Files

Settling a contract on the index: reference price · fetching more often than hourly: how often to fetch · from an AI agent: MCP server.

From a script, send a User-Agent header. Any value works: curl, requests, Node and Go clients are served as they are. The hosting platform, though, refuses Python’s built-in urllib when it sends its default agent (Python-urllib/3.x): HTTP 403, “error code: 1010”. That check sits above the settings we control, so with the standard library, name your client:

import json, urllib.request
req = urllib.request.Request('https://gpueconomy.com/prices.json', headers={'User-Agent': 'my-analysis/1.0'})
prices = json.load(urllib.request.urlopen(req))
print(prices['count'], 'quotes, data updated', prices['data_updated_at'])

Two sources are labelled wherever they appear: gpueconomy is our own hourly scrape (from the week of September 7, 2026); skypilot-catalog is a second-hand history reconstructed from a public price catalog (named in the methodology), with a commit permalink per row. They are never merged. How the numbers are made: methodology.

Slices

The price file also ships narrowed, so a reader who wants one card or one provider does not fetch all 1269 rows. Same fields, fewer rows, same license:

The ids are the ones in gpus.json and providers.json, and they are the same ids in the page URLs — so a slice URL can be derived from a page URL without a lookup. There is no query API and no key: the site is static files on a CDN, and a filter parameter would need a runtime that can be down while the pages are up.

MCP server

Since September 28, 2026 the same data answers AI agents directly: https://gpueconomy.com/mcp is a read-only Model Context Protocol server (Streamable HTTP, no key, no account). How to connect it, with every tool: /mcp.

• list_gpus - every GPU model with its id, VRAM, providers and median on-demand rate card.
• gpu_prices - current rates for one GPU, cheapest first; on-demand by default, or any market.
• provider_prices - every current rate at one provider.
• price_index - the H100 Price Index this week and the last eight weeks.
• recent_price_changes - quotes whose listed price changed in the last 30 days, one row each; marketplace floors on request.

It reads the files on this page and nothing else, so it carries their rules: live quotes only, every price with the time it was read and a link to our page for it, where the price stands beside the page it was read from, CC BY 4.0. It is an experiment with an end date: if too few agents use it, it is taken down and this section says so. What it counts, for that verdict: the time, the tool, the GPU or provider asked about and the name the client gives itself - no IP address, no user agent.

How often to fetch

Prices are read every hour, at about 37 minutes past, and these files change only when the build after a read publishes a price that moved — so at most about once an hour, and some hours not at all. Fetching a file every few minutes returns the same bytes. Every file carries an ETag: send it back as If-None-Match and an unchanged file answers 304 Not Modified with no body. The time the data was read is data_updated_at at the top of each JSON file.

Please fetch each file at most every 15 minutes. Once an hour, at about 50 minutes past, gets every change as soon as it is published; nothing is gained by asking more often, since the next change cannot come sooner than the next hourly read. A client that fetches the same files every few minutes around the clock costs this free service bandwidth for no new data, and conditional requests (If-None-Match) are the polite way to poll. The data stays free under CC BY 4.0 whatever the cadence.

Building something on this data? It is free under CC BY 4.0 with the attribution above, and you do not need to ask. Tell us who you are and we will warn you before anything changes: write to data@gpueconomy.com with the files and fields you use, and you hear from us before a field, a file address, a unit or the set of slices changes. There is no key and no limit either way; the same offer is in each JSON file as change_notices. If you need something the files do not give — history per quote, a feed pushed to you, guarantees on freshness — write to data@gpueconomy.com and say what you are building.

What the licence covers

CC BY 4.0 covers what we made: the compilation — which quotes are read and how, their normalization to US dollars per GPU-hour, the Price Index and its receipts, the comparisons — and our own text. It does not cover what belongs to the providers: their names, trademarks and logos, and their own price pages. A provider’s figure is the provider’s: every row links the provider’s own page for it (and, for a figure read through the provider’s API, names the endpoint beside it), and that is what to cite for the figure itself. When you republish a row, keep that link — to the page the figure stands on, not the provider’s home page — as we do (link policy).

Receipts for the Price Index

Every published week of the Price Index has a receipt: one CSV row per quote that entered that week’s value — provider, region, usd_per_gpu_hour, scope, source URL, read time — and one per quote read for the week and kept out of it, with included=false and the rule in excluded_by, at /price-index/receipts/{YYYY-Www}.csv, for example /price-index/receipts/2026-W44.csv (provisional). A closed week’s receipt never changes. How to turn one back into the published number, with the command: reproduce the index. Each week in price-index.json carries its receipt URL.

/prices.json JSON

Every current rental quote: one row per provider, GPU, price type and region, normalized to USD per GPU per hour, with the listed price it came from and its source URL. The same rows and fields as CSV: /prices.csv.

Updated: Hourly, when a price moved. · Size now: 1269 quotes, 116 providers, 81 GPUs.

/price-index.json JSON

The Cloud GPU Price Index: current value with its constituents (each linked), the weekly series with vintage and the base-100 index, and the reconstructed historical series.

Updated: Hourly; the running week is provisional, closed weeks are final. · Size now: 8 weeks of our own series from September 7, 2026, 94 reconstructed weeks from January 6, 2025.

/price-index.csv CSV

The weekly Cloud GPU Price Index from our own hourly scrape only, one row per published ISO week - no reconstructed history.

Updated: Hourly. · Size now: 8 weeks.

/price-index-hourly.csv CSV

The same Cloud GPU Price Index at every hourly scrape, one row per point, as the database computed and stored it when the scrape came in; never rewritten. Receipts: /price-index/hourly/{YYYY-Www}.csv, every quote of every point with its read time and source.

Updated: Hourly. · Size now: 1257 points.

/price-index-monthly.csv CSV

The monthly average of the hourly index over a fixed monthly panel: closed months as the database froze them (never revised), the running month provisional. Panels: /price-index/panel/{YYYY-MM}.csv.

Updated: Hourly for the running month; a closed month never changes. · Size now: 2 months.

/price-index-history.csv CSV

The reconstructed H100 series from a public price catalog (named in the methodology), kept in its own file.

Updated: Weekly, the Monday after a week closes. · Size now: 94 weeks.

/history.json JSON

Weekly history for every model with enough weeks: median and cheapest listed on-demand rate per week, who listed the cheapest and where, from both sources, labelled.

Updated: Hourly for our own series; weekly for the reconstructed one. · Size now: 20 models, 1695 weekly points.

/token-cost.json JSON

The inputs of the one modelled figure on this site, cost per million output tokens: NVIDIA’s published inference throughput per card, each run with its model, framework, precision, sequence, concurrency and date. Cards are not comparable on it - NVIDIA runs each card in its own configuration.

Updated: When the snapshot of NVIDIA’s figures is refreshed; the file is rebuilt with the site. · Size now: 6 cards, 20 runs.

/changelog.json JSON

Every observed move in a listed rental price, newest first - the list /changelog shows, as data. The same rows and fields as CSV: /changelog.csv.

Updated: Hourly, when a price moved. · Size now: 12779 moves: 30 days for every provider, the last 24 hours for the marketplace (Vast.ai).

/changelog.xml Atom

Providers’ own rate changes as an Atom feed, to follow in a feed reader: the newest 100 moves from changelog.json with the marketplace floor (Vast.ai) left out, since it moves almost every hour. One entry per move, its id stable, linked to the card page and to the source.

Updated: Hourly, when a price moved. · Size now: up to 100 entries.

/providers.json JSON

Every provider tracked or listed: type, how its prices are read, what its region label means, and its current contribution.

Updated: With every site build. · Size now: 122 providers, 116 scraped.

/gpus.json JSON

Every GPU model tracked: specs, current coverage, the 3+-seller retail median, and the manufacturer list price where the vendor publishes one.

Updated: With every site build. · Size now: 81 models.

Not a settlement price

These files record the prices providers publish, read as described on /methodology. They are offered as they are, without warranty, as the CC BY 4.0 licence itself states. We do not design, license, run or guarantee any figure here as a settlement, fixing or reference price for a contract, option, token or other instrument, and we are not a party to anything built on them — unless a licence says otherwise (below). A figure can be withheld, marked unconfirmed or stopped, retired or corrected when its source is read again, a quote can be gone from the next build, and the site can be unavailable. Anyone who settles money on these figures carries that risk. Stated on September 27, 2026; nothing about access or the licence changed with it. What the publisher itself does and does not do is on /methodology.

Using the index as a reference price

Since October 1, 2026 the Cloud GPU Price Index has a written specification: what it measures, how each week is computed and from which sources, its calendar, what is final and never recomputed, what happens when a week is withheld, and — stated as plainly — what a settlement reference needs that the free series does not define (a fallback value, a fixed calculation time, a dispute procedure, a notice period, an obligation to continue). The data is free under CC BY 4.0 for any use, settling a contract, option, token or other instrument on it included: the licence is irrevocable, and nobody needs our permission to do that. What the free files do not carry, a settlement licence adds, per licensee and for a fee: those undefined terms, written down and binding on us; the right to name the index as the reference in the instrument's documents, with GPU Economy's agreement; values pushed at the moment they freeze and signed so that a contract can verify them; and indices the free series does not publish (another card, region or averaging window). Write to data@gpueconomy.com with what you settle, on which series and when. Anything built without one runs on the terms above. Changes: October 1, 2026 — first published saying that settling on the index "needs a licence"; corrected the same day, since CC BY 4.0 already permits it.

Reading the numbers

• usd_per_gpu_hour is the only cross-provider comparable figure. A node price is divided by its GPU count; a per-second rate is multiplied by 3,600; a listing in another currency is converted at the official rate stored next to it (ECB; the Bank of Russia for RUB and VND).
• price_scope says what a figure covers: a full instance, the GPU line item alone, a marketplace floor (median of the five cheapest verified asks), or an advertised “from” rate. Compare like with like.
• The index is a level (the dollar median, printed with its panel size) and a change (the base-100 index, which moves only when a listed price changes). Cite the median for “what it costs” and the index for “how it moved”.
• Vintage: the running week is provisional and is recomputed every hour; a week that has closed is final and never recomputed.

Cite this page

Published under CC BY 4.0: use and redistribute with attribution and a link. Cite the linked source for a price itself, and cite us for the normalization, the index and the comparison. For a single figure, cite the page it appears on; for bulk use, cite this page and name the file.

Data updated: October 28, 2026