# Sovereign AI Index — Methodology (v1.0)

**Product:** `juncture-data-lab/products/sovereign-ai-index/v1.0/`
**Maintainer:** Juncture Policy / Juncture Data Lab
**Version:** 1.0.0
**Methodology last updated:** 2026-08-08
**Status:** Published methodology; scored countries limited to those clearing the evidence-coverage threshold (see §9).

---

## 1. Research question

*How exposed is a country to a loss of strategic autonomy over the infrastructure that produces advanced AI — and how much practical capacity does it currently have to change that exposure?*

The Sovereign AI Index (SAI) is a comparative measurement product. It does not predict outcomes, and it does not measure "AI capability," "AI adoption," or "AI regulation quality" as such. It measures **infrastructural and institutional sovereignty over the AI compute stack**: whether a country can build, control, finance, and keep running the physical and institutional layers advanced AI depends on, under conditions where a small number of firms and one supplier country (the United States, secondarily Taiwan/South Korea/Netherlands for fabrication and lithography, and China as an increasingly separate stack) currently dominate the stack.

## 2. Conceptual definition of "AI sovereignty"

Juncture defines AI sovereignty along four practical axes, each mapped to a pillar below:

1. **Capacity** — does the country have, or credibly will soon have, enough advanced compute and power to run frontier-relevant AI workloads domestically?
2. **Control** — who actually owns and governs that capacity — domestic public/private actors, or foreign capital and foreign-controlled platforms?
3. **Resilience** — if a supplier nation restricted access (export controls, sanctions, supply shock), how exposed is the country, and how diversified is its access?
4. **Institutional and financial capacity to execute** — does the country have the regulatory, data-governance, and financing/delivery capacity to actually convert plans into operating infrastructure, not just announcements?

This is deliberately narrower than "AI readiness" or "AI competitiveness" indices (e.g. Oxford Insights Government AI Readiness Index, Stanford HAI AI Index), which blend talent, research output, adoption, and policy environment. SAI is infrastructure- and control-centered by design, because that is where Juncture's own research base (Fortress AI, When Digitization Becomes Dependency) sits, and because it is the axis most policymakers, journalists, and institutional partners currently lack a rigorous comparative measure for.

## 3. Why these five pillars

| Pillar | Weight | What it answers |
|---|---|---|
| Domestic Compute Capacity | 30% | Can the country actually run frontier-relevant AI workloads today or soon, with power to match? |
| Control & Ownership | 20% | Who owns and governs the infrastructure — domestic actors, or foreign capital/platforms? |
| Compute Access Resilience | 20% | What happens to this country if chip/cloud access were restricted or shocked? |
| Institutional & Data Sovereignty | 15% | Can the country's institutions actually manage, govern, and audit advanced AI infrastructure? |
| Financing & Execution Capacity | 15% | Can the country turn announcements into financed, built, operating infrastructure? |

Domestic Compute Capacity is weighted highest (30%) because it is the precondition the other pillars are about protecting or enabling — a country with no compute and no path to compute cannot meaningfully act on strong scores elsewhere. Control & Ownership and Compute Access Resilience are weighted equally (20% each) because sovereignty is as much about *who controls existing infrastructure* as about *what could be taken away*. Institutional & Data Sovereignty and Financing & Execution Capacity are weighted lowest (15% each) not because they are unimportant, but because — per the "announced ≠ financed ≠ operational" discipline in §6 — execution and institutional maturity are already partially reflected in how Pillar 1 and Pillar 5 indicators are scored (an "announced" data center scores lower than an "operational" one specifically because of execution risk).

**These weights are Juncture's v1.0 analytical judgment, not a universal scientific fact.** A different research team weighting these dimensions differently (e.g., prioritizing institutional capacity over raw compute) could produce a defensible alternative index. Juncture will publish a sensitivity analysis (§12) showing how rankings shift under at least one alternative weighting scheme, and will revisit these weights at each major version.

## 4. Indicators

Each pillar is composed of indicators — the specific, evidence-checkable questions actually scored. Indicator sets are documented per-pillar in `products/sovereign-ai-index/v1.0/config.yaml` and in the per-country data files (`data/countries/<iso3>.json`). Representative indicators per pillar (not exhaustive — see config.yaml for the authoritative list, which may grow as v1.x data availability improves):

**Domestic Compute Capacity (30%)**
- Operational advanced-compute capacity (named clusters, GPU/accelerator counts, EFLOPS)
- Credible committed buildout (financed, under-construction capacity, with target dates)
- Power/grid access adequacy for the country's compute ambitions

**Control & Ownership (20%)**
- Domestic/public ownership share of critical AI infrastructure (data centers, sovereign compute)
- Foreign capital exposure in domestic AI labs/infrastructure (equity stakes, governance rights ceded)
- Bargaining leverage secured over hyperscalers (procurement terms, board seats, security conditions)
- Existence and maturity of domestic model/infrastructure alternatives

**Compute Access Resilience (20%)**
- Exposure to US semiconductor/AI-chip export-control regimes (tier/classification, conditionality)
- Supplier concentration (how many distinct chip/cloud suppliers a country actually depends on)
- Domestic or allied-fab substitution capacity (operational vs. pilot vs. announced)

**Institutional & Data Sovereignty (15%)**
- Institutional capacity to procure and manage advanced AI infrastructure (named body, staffed, funded)
- Data governance/AI regulation status (issued and in force vs. proposed/draft)
- Domestic technical/research base (national labs, research institutes actively operating relevant infrastructure)

**Financing & Execution Capacity (15%)**
- Committed public and private capital (with a citation for "committed," not "potential")
- Demonstrated announcement-to-operational track record (at least one dated example per country, where evidence exists)
- Grid/permitting/site-readiness signal

## 5. Scoring thresholds — the 0/25/50/75/100 rubric

Every indicator is scored on a bounded five-point scale. Continuous "true" measurement is not attempted — SAI deliberately avoids false precision (see §11). Threshold definitions are indicator-specific (documented in `config.yaml`), but follow this general rubric:

| Score | Definition |
|---|---|
| **100** | Fully realized: operational, in force, or verifiably at parity with capacity-leading peers. Evidence: primary source confirms current, active state. |
| **75** | Substantially realized: financed and under construction / issued but not yet fully in force / operating at meaningfully reduced scale. Evidence: primary source confirms committed, in-progress state. |
| **50** | Partially realized: announced and credibly financed, or a real but partial/pilot-scale capability exists. Evidence: primary or credible secondary source; may include some inference. |
| **25** | Nominal/early-stage: announced only, not yet financed; draft regulation; exploratory MOU without binding commitment. Evidence: announcement-stage source only. |
| **0** | No credible evidence of this capability existing or being pursued. **Not the same as NA — see §6.** |

Indicator scores are converted to a 0–1 normalized value before pillar aggregation (`score / 100`). Pillar scores are the indicator-weighted mean of available (non-NA) indicators within that pillar, rescaled to 0–100. The overall country score is the weighted sum of pillar scores using the weights in §3.

## 6. Scoring rules (non-negotiable, enforced by validation gates)

1. **No evidence = NA, not zero.** An indicator with no located evidence is left unscored (`null`) and excluded from that pillar's denominator — it does not silently become a 0. A 0 is a scored judgment ("no credible evidence this exists," positively established); NA is an admission of unknown.
2. **Weak press speculation ≠ confirmed deployment.** A single aggregator or unverified blog claim is never sufficient for a score above 25, and is disqualifying for VERIFIED FACT status (see §7).
3. **Announced ≠ financed.** An announcement without a named financing source/mechanism scores at most 25.
4. **Financed ≠ under construction.** Committed capital without physical construction evidence (site work, procurement, workforce reporting) scores at most 50.
5. **Under construction ≠ operational.** Only confirmed live/operating status scores 75–100, depending on scale relative to stated targets.
6. **Proposed regulation ≠ issued regulation.** Draft/proposed frameworks score at most 25–50 on Institutional & Data Sovereignty indicators; only regulation confirmed in force scores higher.
7. **Selected sites ≠ completed infrastructure.** Site selection (e.g., DOE's 4 selected AI-infrastructure sites) is an announcement-stage fact and is scored accordingly, regardless of how significant the selection itself is.
8. **Analytical inference must be labeled as inference.** Where a score requires interpretation beyond what a primary source states outright (e.g., estimating "supplier concentration" from multiple disclosed contracts), the indicator's evidence-status field is set to `ANALYTICAL_INFERENCE`, never `VERIFIED_FACT`.
9. **A pillar is not scored if evidence coverage is too thin.** A pillar score is only computed if at least 60% of its indicators (by weight) have a non-NA value. Below that, the pillar is marked `insufficient_data` and excluded from the overall score calculation (its weight is redistributed proportionally among scored pillars, and the country's `score_status` is downgraded — see §9).

## 7. Evidence hierarchy and status classification

**Source-quality hierarchy (preferred → last resort):**
1. Government / regulator / official program documents and press releases (e.g. energy.gov, bis.gov, gov.uk, meity.gov.in)
2. Company filings and official hyperscaler/infrastructure-operator announcements (SEC filings, investor-relations releases, company press releases)
3. Multilateral institutions (IMF, World Bank, OECD, BIS, IEA)
4. Credible research institutions (Stanford HAI, Epoch AI, Oxford Insights, CSIS, RAND) — including independent trackers whose methodology is published and whose figures are cross-corroborated
5. High-quality journalism (Reuters, Bloomberg, FT, CNBC, Nikkei, The National, specialist trade press e.g. Datacenter Dynamics, MediaNama) — used for facts not yet available in tiers 1–3, or to corroborate them
6. Low-quality aggregators — **not used where any higher-tier source exists**, and never sufficient alone to justify VERIFIED FACT status

**Evidence status, recorded per indicator:**
- `VERIFIED_FACT` — confirmed by a tier-1/2/3 primary source, unambiguous, ideally corroborated by a second independent source.
- `ANALYTICAL_INFERENCE` — a reasonable, disclosed interpretation of primary evidence (e.g., computing an implied percentage from disclosed dollar figures), or a well-corroborated tier-4/5 synthesis where no primary source states the fact directly.
- `PROVISIONAL` — real but recent/developing, sourced but likely to change (forward-looking dates, in-progress negotiations, single-source figures pending corroboration).
- `UNVERIFIED` — found only in low-quality aggregators, contradicted by other sources, or not independently corroborated. **Not published in scored form**; may appear in narrative notes as an explicitly flagged open question.

Every scored indicator carries: `value_or_score`, `source` (organization + document title), `source_url`, `source_date`, `access_date`, `evidence_status`. This is stored at the **indicator level**, not just the country level (see `products/sovereign-ai-index/v1.0/data/countries/<iso3>.json`).

## 8. Missing-data treatment

- Missing indicators are `null`, excluded from pillar aggregation denominators (§6, rule 9).
- If a pillar falls below the 60%-weighted-coverage threshold, that pillar is not scored (`insufficient_data`), and its weight is redistributed proportionally across the remaining scored pillars for that country only — it is not silently dropped from the visible methodology, and the redistribution is disclosed in the country's `notes` field and on the country card.
- No imputation, interpolation, or peer-average substitution is used anywhere in v1.0.

## 9. Overall score display rules

- An overall score is displayed as a normal numeric score (0–100) only if **at least 80% of total weighted index coverage** is supported by scored (non-NA, non-insufficient) pillars.
- Between 50–80% weighted coverage: the country is marked **Provisional** — a numeric score is shown but visually flagged, with an explicit coverage percentage disclosed.
- Below 50% weighted coverage: the country is marked **Insufficient Data** — no numeric overall score is shown. Only whatever pillar-level or narrative evidence exists is published, with an explicit statement of what is and is not known.
- `score_status` field values: `scored` | `provisional` | `insufficient_data`.

## 10. Confidence methodology

Each country card carries an overall `confidence` rating (`high` | `medium` | `low`), computed as a function of:
- Share of indicators rated `VERIFIED_FACT` vs. `ANALYTICAL_INFERENCE`/`PROVISIONAL`
- Recency of sources (sources older than 18 months from the vintage date are down-weighted)
- Degree of independent corroboration (single-source vs. multi-source facts)

`high` requires ≥70% of scored indicators at `VERIFIED_FACT` with multi-source corroboration on the highest-weighted (Pillar 1) indicators. `medium` is the default for countries with real but partially single-sourced or `ANALYTICAL_INFERENCE`-heavy evidence. `low` applies where most scored indicators rely on `PROVISIONAL` or single-source evidence, even if the coverage threshold for a numeric score is technically met.

## 11. Posture methodology

**Sovereign AI Posture** is a human-readable classification derived from the *pattern* of pillar scores, not a simple re-bucketing of the overall number (two countries with the same overall score can have different postures if the underlying pillar mix differs):

- **Sovereign Builder** — high Domestic Compute Capacity AND high Control & Ownership: the country both has and controls significant AI infrastructure.
- **Strategic Builder** — high/rising Domestic Compute Capacity but moderate Control & Ownership (e.g., substantial foreign capital participation, or execution still mid-flight): building at scale, but on terms it does not fully control.
- **Strategic Buyer** — moderate-to-low Domestic Compute Capacity, but active, well-financed, and often favorable-terms procurement of foreign compute/capital (e.g., negotiated export-control access, sovereign-fund-backed offtake deals): buying its way toward capacity rather than building domestic capacity from a low base.
- **Import-Dependent** — low Domestic Compute Capacity, low Control & Ownership, and no credible near-term financed path to change that.

**Constrained flag** (separate from posture, can co-occur with any posture): applied where geopolitical exposure, export-control conditionality, financing limits, or power/grid constraints are documented as materially limiting the country's options, independent of its current posture. This is intentionally not folded into the headline posture label, because a country can be actively building (Strategic Builder) while still being materially constrained (e.g., grid interconnection delays capping an otherwise well-financed buildout).

**Numeric bands used to classify pillar scores as High / Moderate / Low for posture purposes (added in the v1.0 adversarial-review pass — v1.0.0's initial publication did not publish these, which a reviewer correctly flagged as making posture unfalsifiable):**

| Band | Compute Capacity / Control & Ownership pillar score |
|---|---|
| High | ≥ 70 |
| Moderate | 40–69 |
| Low | < 40 |

Applying these bands: Sovereign Builder requires compute High AND control High. Import-Dependent requires compute Low AND control Low (or Low-and-insufficient-data). Strategic Builder covers compute Moderate-or-High paired with control that is not High, **or** compute High paired with control High where the pathway to that control is primarily domestic construction/ownership rather than negotiated foreign access. Strategic Buyer is reserved for the specific pattern of Moderate-or-Low compute paired with High control achieved primarily through negotiated foreign access (export-control authorizations, sovereign-fund-backed offtake/JV deals) rather than large-scale domestic fabrication or site construction — this last distinction (negotiated access vs. domestic construction) is a qualitative check on top of the numeric bands, disclosed per country in the `notes` field, because the underlying difference (buying capacity vs. building it) is not fully reducible to the two pillar scores alone.

Where `control_ownership` is itself `insufficient_data` (did not clear the 60% indicator-coverage floor), posture is based on `compute_capacity` plus whatever partial control evidence exists below the scoring threshold, and this is disclosed explicitly in the country's `notes` field rather than presented as if a full control pillar score existed.

Posture and constraint classification are `ANALYTICAL_INFERENCE` by definition — they are Juncture's structured reading of the underlying scored evidence, not directly observed facts, and are labeled as such on every country card.

## 12. Weighting sensitivity

v1.0 publishes one alternative weighting scenario alongside the primary weights, re-running the aggregation with Institutional & Data Sovereignty and Financing & Execution Capacity raised to 20%/20% (from 15%/15%) and Domestic Compute Capacity reduced to 25% (from 30%), to show readers how sensitive rankings are to Juncture's specific weighting choice. This is published as a supplementary table in the dataset, not as a second headline ranking.

## 13. Update cadence

- **Quarterly** re-verification pass for all scored/provisional countries (next scheduled: 2026-Q4).
- **Ad hoc corrections** issued immediately when a factual error is identified (logged in the changelog, §14), independent of the quarterly cycle.
- Each update is a new dataset vintage (`last_verified` date at the indicator level, `version` at the dataset level); historical vintages are retained, not overwritten, so trend/direction fields are auditable.

## 14. Versioning policy

- Version numbers follow `MAJOR.MINOR.PATCH`. MAJOR changes = pillar/weight changes or scope changes (e.g., adding/removing a pillar). MINOR = adding countries or indicators, quarterly re-verification passes. PATCH = factual corrections, source-link fixes, typos.
- A changelog is maintained in the dataset (`data/CHANGELOG.md`) and summarized on the public page.
- v1.0.0 (2026-08-08): initial publication. Corrects the prior unsourced page's DOE "11 federal AI sites" error (see indicator-level correction on the United States Domestic Compute Capacity pillar).

## 15. Known limitations

1. **Coverage is uneven by design.** Countries with more developed capital markets, more transparent government disclosure, and more English-language press coverage will systematically have higher evidence coverage — not necessarily higher actual sovereignty. A low `score_status` should be read as "Juncture could not verify enough to score," not "this country has low sovereignty."
2. **The index measures infrastructure and control, not AI capability, safety posture, talent base, or research output.** A country can score low on SAI while having a strong AI research sector, or vice versa.
3. **Currency and figures are nominal and date-stamped**, not adjusted for purchasing power or inflation; cross-country dollar comparisons should be read with that caveat.
4. **Fast-moving facts.** Export-control regimes, financing rounds, and construction status can change within weeks; every indicator carries a `last_verified` date specifically so readers can judge whether a given fact might be stale.
5. **Foreign-capital ownership stakes in AI labs are frequently not disclosed at the cap-table level.** Where exact ownership percentages are not public (a recurring issue for Gulf sovereign-wealth participation in US labs), Control & Ownership indicators are scored on qualitative governance evidence (board seats, security agreements, voting-control filings) rather than false-precision percentages.
6. **This is Juncture's analytical judgment**, not a peer-reviewed academic instrument. The pillar selection, weights, and posture taxonomy in §3 and §11 are original to Juncture and are disclosed as such throughout.

## 16. Suggested citation

> Juncture Policy, *Sovereign AI Index v1.0* (Juncture Data Lab, 2026), https://juncturepolicy.org/sovereign-ai-index/. Methodology: https://juncturepolicy.org/sovereign-ai-index/methodology/ (or this document).

---

*A skeptical external researcher should be able to reproduce any single country's overall score from its published indicator-level data and this document alone. If they cannot, that is a defect in the dataset, not an acceptable ambiguity in the methodology.*