Juncture Data Lab
Sovereign AI Reality Check
An annual audit scorecard that separates sovereign AI announcements from operational capacity, across seven auditable dimensions: compute access, model layer, data and governance, procurement terms, security levels, financing structure, and operational evidence.
Why this scorecard
CNAS’s Sovereign AI Index counted 184 government-backed projects across 67 countries plus the European Union in its latest data, with infrastructure projects rising to 80 percent of new projects in the first half of 2026, and more than three-fifths of projects disclosing a foreign partner (four-fifths of those American). Carnegie’s agency framing argues that sovereignty is the capacity to make consequential choices under interdependence, not autarky. RAND’s Security Levels (SL1-SL5) define what a facility can actually defend against, with SL3 as the practical baseline against organized crime and capable insiders.
This scorecard takes those three anchors and changes the unit of analysis. It does not count programs; it scores countries on seven dimensions that can be audited, with explicit holds where evidence is insufficient. A country that announces ten data centers and energizes none is scored differently from a country that operates one cluster and can prove it.
Key findings
- The announcement-to-energization gap is the story. The UAE ranks second-lowest on the reality score (compute access plus operational evidence) despite the largest announced pipeline in the seed. Saudi Arabia crossed into production at modest scale.
- Legal sovereignty runs ahead of operational sovereignty. Vietnam has four comprehensive instruments in force (AI Law, Personal Data Protection Law, Law on Data, Cybersecurity Law) and one of the seed’s two lowest operational scores (D7=1, tied with the UAE).
- Procurement terms are the least transparent dimension. Five of twelve countries carry a published hold; only France scores 4, with sovereign-provider restrictions, SecNumCloud, version-freeze terms, and EU Data Act switching rights, all offset by NVIDIA hardware dependency.
- Security-level disclosure is the weakest scored dimension. No country scores above 3; only France, Japan, and Singapore publish auditable government-cloud security regimes (SecNumCloud, ISMAP, GDC Air-Gapped), and none publishes RAND SL3-equivalent controls for AI weights.
- Financing structure separates countries more than announced size. Gulf programs use sovereign-fund equity with offtake-disciplined build (4 of 5); India scores 2 because about INR 400 crore of the INR 10,371.92 crore IndiaAI Mission outlay was released over two years.
- Reality leaders are Japan, Singapore, India, and France. Japan scores 9 of 10, Singapore and India 8, France 7.
Tier groups
Tier 1: Operational
Verified operating sovereign compute and live workloads. Japan (9), Singapore (8), India (8), France (7).
Tier 2: Building
Meaningful operational base plus financed construction. Saudi Arabia (6), Indonesia (6), Brazil (6), Kazakhstan (5).
Tier 3: Early
Small operating base or construction only; announcements dominate. Kenya (4), Mexico (4), United Arab Emirates (3), Vietnam (3).
Tiers are set by the reality score (compute access plus operational evidence, 0-10), not by total announced dollars. No seed country falls in Tier 4 (announcement-only). Coverage varies: held dimensions are excluded from totals, so totals carry a coverage column.
Ranked table
| Rank | Country | Tier | Reality (D1+D7) | Total | Coverage | Flags | One-line read |
|---|---|---|---|---|---|---|---|
| 1 | Japan | 1 | 9 | 24 | 7/7 | – | ABCI 3.0 and SoftBank SuperPOD operating; ISMAP security gate; model layer strong. |
| 2 | Singapore | 1 | 8 | 25 | 7/7 | Power-constrained | ASPIRE 2A+ and 2B operating; SEA-LION open weights on a foreign base. |
| 3 | India | 1 | 8 | 22 | 7/7 | Grant-heavy; hardware-import-dependent | 34,333 empanelled GPUs and from-scratch models in production; slow grant release. |
| 4 | France | 1 | 7 | 26 | 7/7 | Hardware-dependency | Most auditable procurement terms; 96 MW secured but not built; Mistral in ministries. |
| 5 | Saudi Arabia | 2 | 6 | 17 | 5/7 | – | Production compute live 2026-09-01; ALLAM on Microsoft Foundry; offtake financing. |
| 6 | Indonesia | 2 | 6 | 16 | 6/7 | – | BRIN HPC and Indosat AI factory operating; Sahabat-AI app live. |
| 7 | Brazil | 2 | 6 | 18 | 7/7 | – | R$23 billion PBIA; Santos Dumont operating; new supercomputers procured, not built. |
| 8 | Kazakhstan | 2 | 5 | 12 | 5/7 | Constrained | Two TOP500 systems; Data Center Valley under construction for June 2027; no model, no terms. |
| 9 | Kenya | 3 | 4 | 14 | 7/7 | Import-dependent; donor-financed; Framework-first | Konza data center live for government; UNDP AI infrastructure agreed 2026-09-11. |
| 10 | Mexico | 3 | 4 | 10 | 6/7 | Import-dependent; announcement-gap | Coatlicue budgeted but not started; live capacity is foreign hyperscaler regions. |
| 11 | United Arab Emirates | 3 | 3 | 15 | 5/7 | Announcement-gap; constrained; Framework-first | 0 GW energized against 1 GW plus campus announced; campus being redesigned. |
| 12 | Vietnam | 3 | 3 | 16 | 7/7 | Framework-first | Strongest in-force legal stack in the seed; no verified sovereign AI workloads. |
Total is the sum of scored dimensions (maximum 35), shown with coverage. A hold is a published finding about transparency, not a zero. Full evidence, confidence, and per-dimension verification date for every cell are in SCORECARD_DATA.json.
Dimension matrix
| Country | Compute | Model | Data/Gov | Procurement | Security | Financing | Operations |
|---|---|---|---|---|---|---|---|
| Japan | 5 | 4 | 3 | 2 | 3 | 3 | 4 |
| France | 3 | 4 | 4 | 4 | 3 | 4 | 4 |
| Singapore | 4 | 3 | 5 | 3 | 3 | 3 | 4 |
| India | 4 | 4 | 4 | 2 | 2 | 2 | 4 |
| Saudi Arabia | 3 | 4 | 3 | H | H | 4 | 3 |
| Indonesia | 3 | 3 | 3 | H | 1 | 3 | 3 |
| Brazil | 3 | 2 | 4 | 2 | 1 | 3 | 3 |
| Kazakhstan | 3 | H | 3 | H | 1 | 3 | 2 |
| Kenya | 2 | 1 | 4 | 2 | 1 | 2 | 2 |
| Mexico | 2 | 1 | 2 | H | 1 | 2 | 2 |
| United Arab Emirates | 2 | 4 | 4 | H | H | 4 | 1 |
| Vietnam | 2 | 3 | 4 | 2 | 2 | 2 | 1 |
H = hold (insufficient evidence). Scores are 0-5 per dimension. Compute + Operations form the reality score (0-10).
Methodology summary
Seven dimensions, each scored 0-5 with published evidence, source date, evidence tier, confidence, and a verification date:
- Compute access: contracted or built capacity versus announced; energized capacity is scored separately from megawatts announced.
- Model layer: domestic from-scratch, adapted open-weight, or licensed foreign; production deployment required above score 3.
- Data and governance: law in force, not drafts; operating authority and national datasets.
- Procurement terms: exit and portability rights, egress, model-version freeze, training and transfer obligations, multi-vendor qualification.
- Security levels: RAND SL1-SL5 as the ladder; auditable national certification and EU SEAL as public proxies.
- Financing structure: grants and budget lines (low) through sovereign equity and offtake to market-tested project finance (high).
- Operational evidence: live workloads, named users, utilization or independent verification; construction is not operations.
| Score | Meaning |
|---|---|
| 5 | Fully realized, operational at scale, independently verified. |
| 4 | Operational with minor gaps. |
| 3 | Substantial: contracted and building, in force and implementing, or production at modest scale. |
| 2 | Partially realized: credibly financed announcement or a small operating base. |
| 1 | Nominal: announced only or general frameworks without AI-specific evidence. |
| 0 | No credible evidence, or credible evidence of absence. |
Evidence hierarchy: tier 1 primary independent (registries, regulators, parliament, grid, benchmarks, audited filings); tier 2 contracts and operator documents (awards, PPP, tender, certification listings); tier 3 independent analysis and reporting; tier 4 official announcements; tier 5 aggregators and vendor marketing, never sufficient above score 2. Rules: announced is not financed; financed is not under construction; construction is not energized; a released model is not a production deployment; a certification listing is not an SL3 claim. Holds: a dimension with insufficient evidence is published as a hold and excluded from totals, with coverage shown.
Limitations
- Evidence asymmetry: countries with more transparent procurement and English-language reporting show fewer holds. A hold means we could not verify, not that the capability does not exist.
- Procurement opacity is structural; five of twelve countries hold on procurement terms in this seed.
- SL3 mapping is approximate; no roster country publishes RAND-equivalent security levels for sovereign AI infrastructure.
- Announced capacity is unstable: the UAE campus shape changed during the research window, and scores reflect the evidence date only.
- Dollar figures are nominal and not comparable across program types; financing is scored by structure, not size.
- Fine-tunes of foreign open weights can be presented as domestic models; base-model dependency is disclosed where known.
Sources
- CNAS, “Sovereign AI’s Second Wave Is Coming Into View” (2026-08-20) and CNAS Sovereign AI Index data through June 2026. cnas.org
- Carnegie Endowment, “Operationalizing AI Sovereignty Through Agency, Interoperability, and Openness” (2026-09-03); “Early Lessons in the Pursuit of Sovereign AI” (2026-06-17). carnegieendowment.org
- RAND, model weight security levels (SL1-SL5) and AI security guide. rand.org
- Reuters, “UAE revises AI data center plan after Iranian attacks” (2026-09-11). reuters.com
- DEPLOY registry, Stargate UAE announced versus energized capacity (accessed 2026-09-18). registry.deploy.report
- The Gulf Tape, “Stargate UAE: Location, Capacity and September 2026 Update” (2026-09-12). thegulftape.com
- AiTech365, HUMAIN production AI compute with AMD and Cisco (2026-09-01); Microsoft and HUMAIN collaboration (2026-08-26); Semafor, “The Switzerland of AI” (2026-09-04); DCD, Oxagon groundbreaking (2026-09-07). aitech365.com
- Lok Sabha Unstarred Question No. 1688, MeitY (2026-07-29); PIB, Sarvam AI profile (2026-02); ET Enterprise AI, Sarvam CEO interview (2026-09-10); AI4Planet, IndiaAI Mission update (2026-06-30). sansad.in
- METI/NEDO GENIAC; Cabinet Office, Second AI Basic Plan (2026-07-14); OECD.AI, ABCI 3.0; ISMAP official site and Digital Agency cloud service list update (2026-09-01). meti.go.jp
- European Commission, sovereign cloud procurement award (2026-04-17) and Sovereign Cloud Framework explained (2026-06-01); EdgeX, France taps Mistral, not OpenAI (2026-08-18); Magica (2026-09-03); ZDNet France (2026-09-04). commission.europa.eu
- Indosat/IOH Sahabat and Sahabat-AI app coverage (Intelligent CIO APAC 2026-03-18; TelcoNews Asia 2026-02-26); Singapore: STT GDC DC-CFA2 award (2026-08-21), DCD on Singtel RE:AI with Mistral (2026-07-22) and HTX GDC Air-Gapped (2026-07-17), AI in Asia SEA-LION analysis (2026-05-17). sttgdc.com
- Brazil: Agencia Brasil (2026-08-21); Estadao (2026-08-20); IBTimes Brazil (2026-09-11). agenciabrasil.ebc.com.br
- Kenya: UNDP Kenya and KoTDA Letter of Agreement (2026-09-11); Konza Technopolis PPP terms of reference (2026-03). konza.go.ke
- Kazakhstan: Prime Minister’s office Data Center Valley progress review (2026-08-13); Times of Central Asia (2026-07-14); Kursiv (2026-09-08); KAZAKH INVEST (2026-08-28). primeminister.kz
- Vietnam: Allen and Gledhill AI Law analysis (2026-03-01); Reg Intel guide (2026-03-24); VinAI PhoGPT repository. allenandgledhill.com
- Mexico: SECIHTI Coatlicue committee (2026-06-06); Mexico News Daily (2026-03-31); AP News (2025-11-26); ICLG Mexico data protection (2026-07-20); Basham AI bill alert (2026-02-26). secihti.mx
- Juncture SAI v1.1 country files (carried evidence, last verified 2026-08-08/11) and Juncture SAI methodology v1.0. juncturepolicy.org/sovereign-ai-index