Data & Citation
Downloadable datasets
Sovereign AI programme records and cloud GPU rental prices. Review column definitions and missing values below, or open a dataset for downloads and citation details.
Sovereign AI Program Index
Government AI programs, with separate funding stages and qualitative ratings of dependence on chip suppliers and US-jurisdiction inputs.
Compute Cost Index
Cloud GPU rental prices and cost per petaFLOP-day, calculated from dense BF16 throughput.
Sovereign AI Program Index
30 columns · 36 rows
Funding coverage. Announced funding, commitments and disbursements have values for 8, 9, 0 programs, respectively.
The funding stages distinguish a pledge from an approved commitment and from money disbursed or contracted. Compare the source definitions before combining the amounts.
The records contain gaps and mismatched amounts: four programs have a later funding stage recorded without a value for the preceding stage. These amounts require checking against the individual program disclosures before treating them as successive portions of one funding total.
Column completeness. Each bar counts non-empty cells in a column of the downloadable CSV. Select a column to read its definition.
Full bar length represents all 36 rows. 9 of 30 columns are populated on every row, and 3 on none; 2 of those columns are marked "one value" because the entry is the same on every row.
A blank cell has no value in this file. The accuracy of a populated value requires checking its source and any calculations applied to it.
Symbols identify required columns (◆), source and retrieval information (◈), and estimate flags (◊). Required columns can still contain null values.
Compute Cost Index
15 columns · 508 rows
Cost calculations. The hourly rental price is an input. The three cost-per-compute columns normalize that price by the instance’s dense BF16 throughput.
Normalization requires a price, GPU count and per-chip throughput. Missing inputs leave the calculated cost blank; the source and retrieval date identify each price observation.
Full bar length represents all 508 rows. 10 of 15 columns are populated on every row; one of those columns is marked "one value" because the entry is the same on every row.
Download contents
Each dataset has six files. The CSV uses UTF-8 without a byte-order mark and ISO 8601 dates. JSON Schema specifies column types and required fields; the data dictionary supplies their definitions. The methodology document explains the sources and calculations. A SHA-256 checksum verifies the CSV bytes, and a Croissant JSON-LD descriptor records the file structure for compatible data tools.
Versions use the year followed by the release number. The release date identifies when the file was packaged; retrieval dates, where present, identify the observations within it.
Record the version and checksum with your citation. The citation guide gives formatted references.
Supply-chain downloads
The supplier-customer graph behind Supply Chain and Cascade is not available as a packaged download. It includes subscription data and CSET ETO chip-explorer data licensed under CC BY-NC 4.0. Those source restrictions continue to apply; the CC BY 4.0 license on the downloadable datasets does not extend to the supply-chain graph.