Gringras, D., & Scrutica. (2026). Scrutica Sovereign AI Program Index (Version 2026.3) [Data set]. Zenodo. (DOI pending deposition)

@dataset{scrutica_sovereign_ai_index_2026_3,
  author    = {Gringras, David and {Scrutica}},
  title     = {Scrutica Sovereign AI Program Index},
  year      = {2026},
  version   = {2026.3},
  publisher = {Zenodo},
  note      = {DOI pending Zenodo deposition; cite by version + SHA-256 hash from /datasets/sovereign-ai-index until assigned},
  url       = {https://scrutica.com/datasets/sovereign-ai-index}
}

# Scrutica Sovereign AI Program Index — v2026.3

**Release date:** 2026-07-17
**License:** CC-BY-4.0
**Suggested citation:** see `CITATION` block at the bottom.

## Coverage

34 sovereign AI investment programs across 34 jurisdictions. Each row carries three distinct funding stages — `announced_usd`, `committed_usd`, `disbursed_usd` — because conflating them is the single most common source of the "$X billion sovereign AI" inflation pattern in trade press. `announced_govt_only_usd` separates the government-pledged component from stacked private FDI pledges (Stargate, Cumulative Microsoft, etc.).

## Funding-stage definitions

- **Announced**: total publicly-attributed pledge, including private FDI partners. Inflation-prone.
- **Committed**: subset that has cleared legislative / board sign-off, where verifiable.
- **Disbursed**: subset that has actually been contracted / spent, where verifiable.

`*_is_estimated = true` flags when the figure is analyst-assigned (e.g., a backed-out fraction of a stacked pledge).

## Interdependence scoring

Four qualitative axes (low / medium / high) describe each program's structural exposure:

- `nvidia_dependency`: reliance on NVIDIA accelerators
- `tsmc_dependency`: reliance on TSMC foundry capacity
- `us_dependency`: reliance on U.S.-jurisdiction inputs (chips, tools, IP)
- `bis_jurisdiction_reach`: whether the program sits inside BIS Entity-List / FDPR reach

These are not numeric scores; they are analyst calls anchored on the published primary chips, partners, and supply-chain edges of the program. The qualitative ladder is preserved to surface uncertainty.

## Currency handling

`currency_of_record` records the native announcement currency. USD figures use the FX rate in `fx_rate_used` on `fx_rate_date` (with the source in `fx_rate_source`); we do not re-baseline historical announcements to a single year's USD purchasing-power index — the rate is the rate that turns the literal announcement into USD at announcement time.

## Refresh path

The upstream JSON (`data/processed/sovereign_programs.json`) is curated against the program-page primary-source layer in the Scrutica platform; this dataset is regenerated by `scripts/package-datasets.ts`.

## Known limitations

- Some programs (e.g., Chinese provincial / military-aligned) have `structural_non_disclosure = true` — the announced figures should be read as floors, not actuals.
- Stacked pledges are flagged in `key_notes` and broken out in the platform's `/sovereign-ai` decomposition view; the flat CSV preserves the stacked total in `announced_usd`.
- The qualitative interdependence ratings are intended for triage, not for cross-program quantitative ranking.

## Citation

```bibtex
@dataset{scrutica_sovereign_ai_index_2026_3,
  author       = {Gringras, David and {Scrutica}},
  title        = {Scrutica Sovereign AI Program Index},
  year         = {2026},
  version      = {2026.3},
  publisher    = {Zenodo},
  note         = {DOI pending Zenodo deposition; cite by version + SHA-256 hash from /datasets/sovereign-ai-index until assigned},
  url          = {https://scrutica.com/datasets/sovereign-ai-index}
}
```

The interactive surface that consumes this dataset lives at /sovereign-ai; for the platform-wide data-provenance overview, see /methodology.