Methodology

Notes on the series

Every number on this site comes from a data file or a cited public source. This page says which, how each series is built, and where the figures stop.

M1GPU rental price index

Source and unit

The index is a snapshot of a GPU price tracker maintained by the publisher: for each of 10 models it records the cheapest accepted on-demand price, a median, the highest list price and, where any provider publishes one, a 12-month reserved rate. The current snapshot is dated 16 Oct 2026. All values are US dollars per GPU-hour; multi-GPU instances are divided by their GPU count (for example, an eight-GPU instance at $98.32 an hour is $12.29 per GPU-hour).

Floor, median, ceiling

• Floor — the lowest on-demand price the tracker could confirm on a provider page or a live comparison tracker. Spot, preemptible and interruptible tiers are recorded but excluded. Serverless and marketplace prices count, since they are available without commitment.
• Median — the tracker's typical price across providers and comparison trackers. Where a tracker publishes its own median (GetDeploying, ComputePrices) and it agrees with the provider spread, that figure is used; the per-model note on each page says what anchors it.
• Ceiling — the highest published on-demand list price, usually a hyperscaler instance.
• Reserved 12 mo — a published 12-month (1-year) rate. If none is published the cell says so; it is never derived from a discount assumption.

Provider quotes

The provider tables on model pages list each individual price recorded in the tracker's note for that model. Each row is dated by when that note was last written, which can be earlier than the snapshot date when a later refresh confirmed the aggregate figures without rewriting the note. Where the provider's own price page is the source, that page is listed for the model on the Sources page; otherwise the price was read through the comparison trackers listed there for the model. Third-party index rows are shown for context and are not counted as providers.

History

Time series of GPU rental prices. Each point: [date, median, min, max, contract1y] in USD per GPU-hour. Points before 2026-08 are seeded from published market indices (cited in seedSources); from 2026-08 onward the series is collected by Coloprice from provider price pages. Points before 1 Aug 2026 are drawn dashed on every chart. "vs previous snapshot" compares the current median with the preceding point of the series; "last move" is the most recent point where the median changed.

Model pages

A model gets its own page only when at least 3 distinct providers have a recorded price for it. Qualifying now: GB200 (5), B200 (9), MI355X (6), H200 (5), H100 (8), MI300X (7), A100 (5), L40S (7), RTX 4090 (4). Not qualifying: RTX 5090 (2) — shown in the index table only.

Specifications and costs

Specifications on model pages are quoted from vendor pages (or, for the MI355X, from trade press quoting AMD), checked 17 Oct 2026. Monthly figures are arithmetic: hourly rate × 730 hours × GPUs. No purchase prices are shown, because none of our sources gives a dated, citable one.

M2AI data center campus census

Source records

The census is drawn from the publisher's catalogue of 6,018 data center records, in the version of 16 Oct 2026. Each record carries its own source links (operator announcements, regulator and planning documents, exchange filings, trade press and public registries), and the source of every row is listed, campus by campus, on the Sources page.

Definition of an AI campus

A record is included when either (a) it carries the catalogue's AI-ready flag, in any status — an editorial label set when a record's sources present the site as built or marketed for AI or GPU workloads; there is no stricter published test behind it, so every row's source is listed for checking — or (b) it is under construction or announced with a published capacity of at least 50 MW. Rule (b) captures large projects whose AI use is not yet stated; they are labelled "≥50 MW pipeline" in the Basis column so readers can exclude them.

Result: 634 campuses — 557 on the AI flag, 77 on the pipeline rule only — in 55 countries; 467 publish a capacity, totalling 89,035 MW.

Duplicates removed

• tcs-hypervault-hyderabad → kept tcs-hypervault-future-city-hyderabad: same 1,000 MW TCS HyperVault Hyderabad announcement, identical source URLs.
• moro-hub-warsan-green-dc → kept moro-hub-warsan-green-data-centre-dubai: same Moro Hub Warsan (Dubai) build, 100 MW, two records.
• khazna-ajman-ai-data-center → kept khazna-qaj01-ajman: Khazna Ajman AI-optimised site and QAJ01 are both 100 MW in Ajman; treated as one project until shown otherwise.
• stt-gdc-palava-1 → kept stt-gdc-palava-mumbai: same STT Palava 1 building in Mumbai, 50 MW under construction, two records citing different articles (found 17 Oct 2026).

Capacity, cooling, density

• MW is kept exactly as the source publishes it — some quote IT load, others grid connection or a campus master plan. Totals add the published numbers and are lower bounds, because 167 campuses publish none.
• Liquid cooling is marked only where the record says so; absence of the mark is not evidence of air cooling.
• Rack density (kW per rack) appears only where published (24 campuses). Values above 300 kW per rack are shown as published but left out of the median quoted in the text (1 such record).

Capacity statements

14 public statements (approvals, power agreements, financings, allocations, delays), updated 15 Oct 2026. Figures are quoted as stated by the party, from press releases, regulator decisions, exchange filings and operator announcements; private correspondence is never used. They are not added to census totals because some describe campuses already in the table.

M3Updates, corrections, licence

The underlying price tracker and catalogue are refreshed by scheduled jobs; this site republishes when the data is synced, and every page shows the date of the data it uses. Corrections: if a figure is wrong, send the source through the contact form and it will be checked against the original.

Our compilation — selection, structure, normalisation to USD per GPU-hour, and the census itself — may be reused under the Creative Commons Attribution 4.0 licence (CC BY 4.0; licence text linked on Sources) with attribution to Compute Census. Individual prices and statements remain those of the cited providers and publishers. Machine-readable copies: /data/gpu-prices.json and .csv, /data/ai-campuses.json and .csv.