Skip to main content
How much practical control does a country have over the infrastructure that produces advanced AI — and how exposed is it if that control were tested?
Ranking
Click a column header to sort. Countries marked Provisional or Insufficient Data are included for transparency, not omitted — see Methodology §9 for the display rule.
| Country | Score | Status | Posture | Confidence | Constrained | Last verified |
|---|---|---|---|---|---|---|
| United States | 81.2 | Scored | Sovereign Builder | medium | Constrained | 2026-08-08 |
| China | 75.4 | Scored | Sovereign Builder | medium | Constrained | 2026-08-08 |
| Taiwan | 75.0 | Scored | Strategic Builder | high | Constrained | 2026-08-10 |
| United Arab Emirates | 70.2 | Scored | Strategic Builder | medium | Constrained | 2026-08-08 |
| Saudi Arabia | 70.0 | Scored | Strategic Buyer | high | Constrained | 2026-08-08 |
| Japan | 67.9 | Scored | Strategic Builder | medium | Constrained | 2026-08-08 |
| Germany | 67.1 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| Switzerland | 66.6 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| South Korea | 65.4 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| Italy | 61.7 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| Norway | 61.3 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| Spain | 60.9 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| India | 60.0 | Scored | Strategic Builder | medium | Constrained | 2026-08-08 |
| Vietnam | 59.4 | Scored | Strategic Builder | medium | Constrained | 2026-08-08 |
| France | 59.2 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| Singapore | 59.1 | Scored | Strategic Builder | medium | Constrained | 2026-08-11 |
| Indonesia | 55.4 | Scored | Strategic Builder | medium | Constrained | 2026-08-11 |
| Ireland | 55.0 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| United Kingdom | 53.7 | Scored | Strategic Builder | medium | Constrained | 2026-08-08 |
| Australia | 51.2 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| Kenya | 46.9 | Provisional | Import-Dependent | medium | Constrained | 2026-08-08 |
| Israel | 46.7 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| Canada | 45.4 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| Brazil | 45.3 | Scored | Strategic Builder | medium | Constrained | 2026-08-08 |
| Nigeria | 43.8 | Provisional | Import-Dependent | medium | Constrained | 2026-08-08 |
| Turkey | 38.8 | Scored | Strategic Builder | medium | Constrained | 2026-08-10 |
| Mexico | 35.4 | Scored | Import-Dependent | medium | Constrained | 2026-08-10 |
| Russia | 35.0 | Scored | Import-Dependent | medium | Constrained | 2026-08-10 |
| Moldova | — | Insufficient data | Import-Dependent | low | Constrained | 2026-08-08 |
Country evidence cards
Click a card for the full pillar breakdown, indicator-level sourcing, and notes (or use the full detail list further down the page, which works without JavaScript).
United States
Sovereign Builder
Scored medium Constrained
Unmatched scale of operational and financed compute: existing hyperscaler infrastructure plus ~$725B in guided 2026 capex (Microsoft, Google, Amazon, Meta) and DOE's Genesis Mission compute partnership (Argonne Solstice/Equinox, up to 110,000 Blackwell GPUs combined).
United Kingdom
Strategic Builder
Scored medium Constrained
Genuinely operational, publicly-funded sovereign compute delivered on a fast timeline: Isambard-AI (£225m, 5,448 NVIDIA GH200 superchips, 21 AI-EFLOPS, 11th on TOP500) and Dawn (Cambridge, 1,000+ Intel chips) both switched on together 17 Jul 2025, backed by a £2bn Compute Roadmap to 2030 (published 18 Jul 2025) and a now-operating £500m Sovereign AI Unit (launched 16 Apr 2026, up to £20m equity per startup plus 1m GPU-hours of AIRR access).
Japan
Strategic Builder
Scored medium Constrained
Independently-confirmed domestic control of frontier-scale compute, not merely hosted foreign infrastructure: SoftBank's DGX SuperPOD (NVIDIA DGX B200, >4,000 Blackwell GPUs, operational since 22 Jul 2025, the world's largest operating DGX SuperPOD on DGX B200 at that date; >10,000 GPUs and 13.7 exaflops across SoftBank's combined platform) and AIST's ABCI 3.0 (6,128 NVIDIA H200 GPUs, 6.22 AI-EFLOPS, operational since Jan 2025) are both domestically owned and operated, with foreign hyperscalers (Microsoft, AWS, Oracle, ~$26-28bn combined 2024-26) explicitly playing a secondary, complementary role.
India
Strategic Builder
Scored medium Constrained
Real, multiply-corroborated operational compute base (34,000-38,000 empanelled GPUs, mostly Nvidia H100/H200, across 10 private providers under the IndiaAI Mission) combined with a large private buildout genuinely under construction: Reliance's 1GW Nvidia-Blackwell data center at Jamnagar, part of an announced $110B/7-year sovereign AI infrastructure commitment.
Saudi Arabia
Strategic Buyer
Scored high Constrained
Unusually fast, dated, verifiable execution: HUMAIN moved from its founding (May 2025) through a $10B AMD collaboration, a Nov 2025 US chip-export authorization, and physical receipt of its first Nvidia GPU shipment by December 2025 -- roughly 7 months from framework to hardware, the fastest confirmed announcement-to-delivery cycle of any country in this index.
United Arab Emirates
Strategic Builder
Scored medium Constrained
The July 10, 2026 BIS reclassification of the UAE to Country Group A:5 (the highest US export-control trust tier) combined with Stargate UAE — a 1GW cluster (part of a planned 5GW, 19.2 km² campus) under construction in Abu Dhabi with named partners (OpenAI, Oracle, Nvidia, Cisco, SoftBank), up to $10B committed for the initial phase, backed by Barakah's fully operational 5.6GW nuclear power base.
Brazil
Strategic Builder
Scored medium Constrained
A genuine, functioning, domestically-controlled sovereign cloud: Serpro and Dataprev (both state-owned) operate Brazil's 'Nuvem de Governo' (Government Cloud), with R$5.7 billion of the national AI plan (PBIA) specifically earmarked for sovereign-cloud/supercomputer infrastructure, and the Santos Dumont supercomputer (18.85 petaflops after a 2025 upgrade) is now the most powerful in Latin America.
Moldova
Import-Dependent
Insufficient data low Constrained
An unusually current, active institutional trail on AI policy specifically: signed the Council of Europe AI Framework Convention (Sep 2024), released a first national AI/data-governance policy White Paper (Oct 2024), launched a coordinating Sub-Council for AI (SCIA, Apr 2026) and a public-private Moldova AI Hub (Sep 2025) -- rare for a country this small.
Nigeria
Import-Dependent
Provisional medium Constrained
Real, dated, dollar-figured private-sector investment is actively flowing: MTN's Sifiso Dabengwa Data Centre Phase 1 (4.5MW, Tier III) began operations July 2025, and Airtel's Nxtra $120M hyperscale facility has already received early GPU shipments -- this is genuine build activity, not just announcements.
Vietnam
Strategic Builder
Scored medium Constrained
The most comprehensive, dated, multi-year institutional and legal apparatus among the four originally 'thin-evidence' countries: a national semiconductor strategy through 2050 (Decision 1018/QD-TTg, Sep 2024), a new Cybersecurity Law (effective Jul 2026) and Law on Data (effective Jul 2025) both explicitly creating data-localization and cross-border-transfer restrictions, plus genuine multi-vendor infrastructure diversification (Microsoft/G42, ST Telemedia/VNG, domestic KBC/VietinBank all actively building).
Kenya
Import-Dependent
Provisional medium Constrained
Kenya has the strongest confirmed OPERATIONAL domestic AI-relevant infrastructure among the four thin-evidence countries: the Konza Technopolis data center (Huawei-built, Tier III, state-owned via the Konza Technopolis Development Authority) has been running since ~2020, serving 120+ industry clients and 70+ government applications, financed via a Chinese concessional loan.
China
Sovereign Builder
Scored medium Constrained
Extraordinary state financing scale paired with real (not merely announced) execution: Big Fund Phase III (Y344bn / ~$47.5B registered capital, incorporated May 2024) alongside operational deployment such as Alibaba's in-house-chip AI data center and SMIC's 7nm node in mass production.
South Korea
Strategic Builder
Scored medium Constrained
An unusually large, coordinated, and dated public-private capital commitment: government and private-sector investment orchestration exceeding 1,350 trillion won (~$880B, Samsung/SK Hynix chip and data-center buildout announced June 29, 2026), a 2026 government budget with direct AI spending up over 300% YoY to 10.1 trillion won, and a 550 trillion won plan for 8.4GW of AI data-center capacity -- combined with world-leading domestic HBM memory production (Samsung, SK Hynix) that is itself critical, currently-shipping AI-chip infrastructure.
Singapore
Strategic Builder
Scored medium Constrained
A real, operating national supercomputing base (NSCC Singapore's ASPIRE 2A+ delivering 20 PetaFLOPS on Nvidia H100s, with the next-generation ASPIRE 2B launched June 8, 2026) combined with a sophisticated, government-controlled data-center capacity-allocation regime: Singapore ended its 4-year data-center moratorium in Dec 2025 and now grants new capacity only to operators meeting binding green-power and efficiency conditions, giving the state real negotiating leverage over hyperscalers rather than simply granting open access.
Indonesia
Strategic Builder
Scored medium Constrained
Real, operational, government-run AI/HPC research infrastructure: BRIN (Indonesia's National Research and Innovation Agency), in partnership with South Korea's KISTI, inaugurated the ASEAN-Korea HPC facility in Jakarta on June 18, 2026 (4.28 petaflops peak, 3.10 petaflops tested, a TOP500-candidate submission) -- combined with substantial, named, dated capital specifically earmarked for data centers (Danantara's Rp21 trillion/~$1.36B 'Fukuoka Project,' part of a Rp202.4 trillion/$13.1B 2026 investment plan) and large secured grid capacity (BDx's 1.2GW PLN power commitment, June 2026).
Taiwan
Strategic Builder
Scored high Constrained
Taiwan is the physical fabrication base for the world's most advanced AI accelerators: TSMC's 2nm node entered mass production in late 2025 with five fabs ramping through 2026, and 3nm monthly capacity is projected to reach 180,000-250,000 wafers by 2026-27 -- a domestic, operational, at-scale capability no other country in this assignment set has.
Australia
Strategic Builder
Scored medium Constrained
Genuinely operational, named domestic compute infrastructure (CSIRO's 448-H100 'Virga', Pawsey's Setonix, and commercial GPUaaS such as Sharon AI's 1,024-GPU Blackwell Ultra cluster) combined with one of the world's largest financed data-centre investment pipelines (Firmus Project Southgate: US$10bn Blackstone-led debt plus US$2bn equity, 2026; Microsoft A$25bn and AWS A$20bn multi-year commitments).
Canada
Strategic Builder
Scored medium Constrained
A genuinely funded and executing institutional/financing base: Budget 2024's C$2bn Canadian Sovereign AI Compute Strategy, topped up by Budget 2025's C$925.6m, culminated in the April 2026 launch of the Sovereign Compute Infrastructure Program (SCIP) with binding Canadian-ownership/governance conditions; a real national HPC network is already operating (Fir at SFU, fully operational Sept 2025, ~20 petaflops; Trillium at U of T; tamIA at Laval), and Canada already has a genuine 2025-26 announcement-to-operation execution record (Fir, tamIA, the Cohere/CoreWeave Cambridge facility, TELUS's Rimouski Sovereign AI Factory).
France
Strategic Builder
Scored medium Constrained
A credible, financed, large-scale buildout pipeline anchored in a genuine energy advantage: nuclear provides roughly two-thirds of French generation with an EDF EUR70/MWh pricing ceiling from 2026, and Choose France 2026 (June 1, Versailles) produced a record EUR93bn in investment commitments including SoftBank's EUR75bn/5GW pledge (EUR45bn firm for phase 1) and Brookfield's EUR30bn cumulative commitment via Data4.
Germany
Strategic Builder
Scored medium Constrained
A confirmed, operating national AI-compute base: the JUPITER exascale supercomputer at Juelich (inaugurated Sept 2025, EuroHPC-owned, GH200-based, ranked 4th on the Nov 2025 TOP500 list) is Europe's fastest supercomputer, and Deutsche Telekom's Munich Industrial AI Cloud (built with Nvidia in six months, ~10,000 Blackwell GPUs) is already operating at over one-third capacity.
Spain
Strategic Builder
Scored medium Constrained
A genuinely public, operating sovereign-AI research asset: the Barcelona Supercomputing Center (BSC-CNS, a public consortium of the Spanish state, the Catalan government and UPC) developed and released ALIA, Europe's first public, open, multilingual LLM, trained on the MareNostrum 5 supercomputer across 35 European languages; Spain is also one of only three EU countries (with Germany and Poland) hosting two EuroHPC AI Factories (Barcelona and CESGA/Galicia).
Italy
Strategic Builder
Scored medium Constrained
Italy's flagship national AI-gigafactory bid is anchored by companies with substantial, verified Italian state ownership: Eni (Ministry of Economy and Finance plus CDP hold a combined 33.09% of share capital) and Leonardo (Ministry of Economy and Finance holds ~30.2%) are leading the bid together with TIM -- described as the only coordinated single-country gigafactory proposal submitted to Brussels. Italy was also the first EU member state to pass a national AI law complementing the EU AI Act (Law No. 132/2025, implementing decrees approved June 10, 2026).
Russia
Import-Dependent
Scored medium Constrained
Strong domestic ownership and control of existing AI infrastructure -- Sberbank (operator of the Christofari supercomputers) is majority state-owned (50%+1 share held via the Ministry of Finance since 2020) and Yandex (operator of Russia's largest AI clusters) became 100% Russian-owned on 2024-07-15 when Dutch parent Yandex NV completed a full divestment -- combined with mature, actively-deployed domestic foundation models: GigaChat (Sber) and YandexGPT (Yandex) are both trained from scratch and serve production workloads across banking, public services and enterprise.
Mexico
Import-Dependent
Scored medium Constrained
A genuine, dated hyperscaler execution track record -- AWS, Microsoft, Google and Oracle have all moved from announcement to live cloud regions in Mexico on documented timelines (Oracle operating two regions since 2022-23; AWS live Jan 2025; Google live Dec 2024) -- alongside a credible, if not yet built, national supercomputer program: Coatlicue (MX$6bn, 314-petaflop design, technical committee installed June 2026, sited at IPN Zacatenco).
Ireland
Strategic Builder
Scored medium Constrained
Ireland is an established, continuously-operating European hyperscaler hub: AWS has run its Ireland region since Nov 2007 (now 3 Availability Zones, 33 sites nationally per operator tracking), alongside Microsoft (11 facilities), Google, and Meta, with 72 operational grid-connected data-centre buildings confirmed in a June 2026 government-commissioned KPMG report. This is reinforced by a fresh, large, dated capital commitment: Intel's €5bn ($5.7bn) Fab 34 (Leixlip) expansion, announced 13 Jul 2026.
Switzerland
Strategic Builder
Scored medium Constrained
A fully operational, wholly state-owned, top-10-globally-ranked AI supercomputer -- CSCS's 'Alps' (10,752 NVIDIA GH200 Grace-Hopper sockets, 434.9 petaflops sustained, #10 on the June 2026 TOP500 list) -- that has already been used to train and publicly release Switzerland's own open-source sovereign LLM, Apertus (released September 2, 2025).
Norway
Strategic Builder
Scored medium Constrained
Real, financed, under-construction hyperscale build-out at multiple sites, anchored by cheap hydropower: Nscale/Aker's Narvik campus (230 MW initial phase, 100,000 targeted NVIDIA GPUs, >$10bn estimated total cost per Time's reporting) backed by a Microsoft compute contract reported at $6.2bn (Sept 2025) and expanded to over 30,000 NVIDIA Rubin GPUs (Apr 2026); TikTok/Green Mountain's Hamar campus (90 MW, 3 buildings) is already fully complete and operational.
Turkey
Strategic Builder
Scored medium Constrained
TÜBİTAK ULAKBİM's ARF-ACC national supercomputer is a real, actively-upgraded, independently verified operational asset: it climbed from rank 266 to 145 on the TOP500 global list between June and November 2025 after the addition of 192 Nvidia H200 GPUs (TOP500.org, Nov 18, 2025), sitting on top of a TRUBA e-infrastructure that reports 80,000 processor cores and 312 GPUs serving thousands of researchers.
Israel
Strategic Builder
Scored medium Constrained
A live national AI supercomputer (Nebius-operated, 1,000 Nvidia B200 GPUs of a 4,000-GPU program, allocating since Jan 2026) plus Nvidia's own Israel-1 facility (operational since Nov 2023), combined with the strongest documented hyperscaler bargaining leverage found in this research round: leaked Project Nimbus contract terms (reported by +972 Magazine/Guardian, Oct 2025, based on Finance Ministry documents) reportedly bar Google/AWS from restricting Israel's use of services even where it breaches their own terms of service, and require covert notification if a foreign court orders data turnover, backed by liquidated-damages clauses.
Full country detail
Every country, expanded: pillar scores, strongest capability, principal dependency, catalyst, constraint, and full notes. This section works without JavaScript and is what search engines and archival tools will see.
United States — 81.2
Posture: Sovereign Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-08
Pillar scores
Strongest capability
Unmatched scale of operational and financed compute: existing hyperscaler infrastructure plus ~$725B in guided 2026 capex (Microsoft, Google, Amazon, Meta) and DOE's Genesis Mission compute partnership (Argonne Solstice/Equinox, up to 110,000 Blackwell GPUs combined).
Principal dependency
Leading-edge chip fabrication remains substantially offshore (TSMC Taiwan for the most advanced nodes; TSMC Arizona is operational but not yet at parity), and frontier AI labs carry growing, well-documented Gulf sovereign-wealth capital exposure (UAE's MGX, Saudi Arabia's Humain/PIF, Qatar's QIA — tens of billions of dollars across OpenAI, xAI, and Anthropic rounds in 2025-26).
Key catalyst
DOE's selection of 4 federal sites for AI data-center/energy-infrastructure development (July 24, 2025) and the Genesis Mission NVIDIA/Oracle/Argonne compute partnership (announced Oct 28, 2025).
Key constraint
Grid/interconnection capacity, not capital, is now the binding constraint on further scale-up at multiple flagship sites.
Constraint notes
Grid interconnection delays are a documented, binding constraint on scaling despite abundant capital: OpenAI/Oracle capped the Abilene ('Stargate') site at 1.2 GW specifically because of multi-year interconnection-queue delays; national interconnection queues held ~2,600 GW of proposed capacity in early 2026 with median wait approaching 5 years.
Notes
REVISED after adversarial review (2026-08-08): confidence downgraded from high to medium. Applying methodology section 10's formula directly: only 10 of 16 total indicators are VERIFIED_FACT (62.5%, below the 70% bar for 'high'), and on the highest-weighted pillar (compute_capacity) only 1 of 3 indicators is VERIFIED_FACT with a single source, not the multi-source corroboration the 'high' threshold specifically requires on Pillar 1. This is the correct, reproducible application of the published formula, not a discretionary downgrade. CORRECTS a prior unsourced Juncture page claim of '11 DOE federal AI sites.' DOE selected 4 sites for AI data-center/energy-infrastructure development (energy.gov, 24 Jul 2025), narrowed from an initial 16-site RFI pool (3 Apr 2025). The figure of 11 refers to a separate DOE Nuclear Reactor Pilot Program (established via Executive Order 14301, selections announced 12 Aug 2025) — a nuclear-reactor demonstration/licensing initiative, not an AI-infrastructure siting program, though several of its projects are co-located at Idaho National Laboratory, which is also one of the 4 AI-infrastructure sites. This geographic overlap is the likely source of the conflation. All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
United Kingdom — 53.7
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-08
Pillar scores
Strongest capability
Genuinely operational, publicly-funded sovereign compute delivered on a fast timeline: Isambard-AI (£225m, 5,448 NVIDIA GH200 superchips, 21 AI-EFLOPS, 11th on TOP500) and Dawn (Cambridge, 1,000+ Intel chips) both switched on together 17 Jul 2025, backed by a £2bn Compute Roadmap to 2030 (published 18 Jul 2025) and a now-operating £500m Sovereign AI Unit (launched 16 Apr 2026, up to £20m equity per startup plus 1m GPU-hours of AIRR access).
Principal dependency
The large-scale compute buildout is overwhelmingly foreign-financed and foreign-owned: Microsoft's $30bn, Google's £5bn, and Amazon's £8bn UK commitments (announced/expanding through Sep 2025) dwarf domestic public compute spending by roughly an order of magnitude, and the UK's National Semiconductor Strategy (May 2023) explicitly chooses not to pursue an advanced domestic chip foundry, leaving both Isambard-AI and Dawn running on US-designed silicon with no domestic alternative.
Key catalyst
The MOD-Google £400m sovereign, air-gapped cloud contract (awarded Sep 2025) and the £14bn G-Cloud 15 framework (Sept 2026–2030) show the UK beginning to convert public procurement into genuine sovereignty terms with hyperscalers, rather than relying on capacity alone — the clearest example of UK government leverage over a hyperscaler identified in current evidence.
Key constraint
Grid interconnection capacity, not capital, increasingly caps further scale-up: a 50 GW connection-queue backlog against a 2030 demand target of ~26 TWh is a live risk with only a consultation-stage policy response so far, and the country has no domestic advanced-fab remediation path analogous to Japan's Rapidus.
Constraint notes
Grid interconnection is a documented, binding constraint on future scale-up: GB data-centre electricity demand (~5 TWh, ~2% of UK demand, 2023) is projected to reach ~26 TWh (~9%) by 2030, against a grid-connection queue of up to 50 GW of applications with multi-year waits; the policy response remains under consultation only, not yet enacted. Separately, the UK's 2023 National Semiconductor Strategy explicitly rules out pursuing advanced domestic chip fabrication, leaving compute delivery structurally dependent on foreign (largely US) silicon with no funded remediation path, unlike peers pursuing domestic/allied fab substitution.
Notes
CORRECTS two vague live-page claims per this vintage's verification pass. (1) 'NHS AI infrastructure' should be corrected to 'strategic health-data platform': what actually exists is the NHS Federated Data Platform (built on Palantir Foundry, live since 2024, £330m allocated 2025–2028 — PROVISIONAL, single-source on duration, up to £2.4bn claimed efficiency savings), a data/analytics platform that AI tools run through, not national AI compute/GPU infrastructure. (2) 'Sovereign cloud procurement' does not name an actual program — no single UK program carries that name. The real instruments are the MOD-Google £400m sovereign (air-gapped) cloud contract (awarded Sep 2025) and G-Cloud 15 (£14bn framework, Sept 2026–2030); the Compute Roadmap's 'sovereign, secure, sustainable' framing explicitly does not constitute a formal sovereign cloud policy per legal/industry commentary (Burges Salmon). The live page should name these instruments directly rather than use the generic phrase. control_ownership was scored on 3 of its 4 indicators: domestic_alternatives is left NA — no named UK domestic frontier-model or AI-infrastructure alternative to foreign hyperscalers/labs is identified in available evidence; the £500m Sovereign AI Unit is a startup-equity financing vehicle aimed at eventually seeding such alternatives, not itself a demonstrated one, per methodology section 6 rule 1 (no evidence = NA, not a guessed score). That pillar's weight was redistributed proportionally across its 3 remaining scored indicators per methodology sections 6(9) and 8; the pillar remained above the 60%-weighted-coverage threshold (75%) so it is still reported as 'scored,' not 'insufficient_data.' All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
Japan — 67.9
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-08
Pillar scores
Strongest capability
Independently-confirmed domestic control of frontier-scale compute, not merely hosted foreign infrastructure: SoftBank's DGX SuperPOD (NVIDIA DGX B200, >4,000 Blackwell GPUs, operational since 22 Jul 2025, the world's largest operating DGX SuperPOD on DGX B200 at that date; >10,000 GPUs and 13.7 exaflops across SoftBank's combined platform) and AIST's ABCI 3.0 (6,128 NVIDIA H200 GPUs, 6.22 AI-EFLOPS, operational since Jan 2025) are both domestically owned and operated, with foreign hyperscalers (Microsoft, AWS, Oracle, ~$26-28bn combined 2024-26) explicitly playing a secondary, complementary role.
Principal dependency
Both the compute layer and the emerging fab layer remain foreign-sourced at the technology level even where the physical assets are domestically owned: SoftBank's and AIST's platforms run on NVIDIA silicon, and Rapidus's 2nm pilot line depends on IBM's licensed gate-all-around process and ASML EUV lithography even though the fab itself is sited in Chitose, Hokkaido and is majority Japanese-funded.
Key catalyst
Rapidus's 2nm foundry has moved past announcement into a genuinely operating pilot line with demonstrated working prototypes (successful GAA transistor prototyping with target electrical characteristics, 18 Jul 2025) on ~JPY2.354 trillion (~$14.7-15bn) of cumulative committed public and private funding as of Feb 2026, trending toward ~JPY3 trillion by end-FY2027 — a funded, in-progress path to leading-edge domestic fabrication that most peers, including the UK, have explicitly chosen not to pursue.
Key constraint
Power demand growth (triple by 2034) is outpacing supply-side response despite the Feb 2026 nuclear restart, and Rapidus's foundry — while genuinely operating at pilot scale — has not yet reached mass production (targeted 2H FY2027), so Japan's flagship resilience investment remains an unproven, if credible, bet as of this vintage.
Constraint notes
Power is a documented, quantified constraint: Wood Mackenzie projects Japan's data-centre electricity use to more than triple, from 19 TWh (2024) to 57-66 TWh by 2034, driving ~60% of Japan's total power-demand growth; Japan restarted the Kashiwazaki-Kariwa Unit 6 nuclear reactor (1,356 MW) on 9 Feb 2026 partly to support this, which is an active mitigation but not yet a resolved constraint. Separately, Rapidus's leading-edge fab remains pre-mass-production (pilot line only, target 2H FY2027) and its core process technology is foreign-licensed (IBM 2nm GAA, ASML EUV), so the flagship domestic-fab remediation path is real and funded but not yet delivering independent capacity.
Notes
CORRECTS two vague live-page claims per this vintage's verification pass. (1) 'Economic security AI program' does not name a single actual program. The real, verifiable instrument is METI's Apr 2024 subsidy program under Economic Security Promotion Act Article 9(4) (up to JPY72.5bn across five GPU-cloud projects: Sakura Internet ~JPY50.1bn, KDDI ~JPY10.24bn, Highreso ~JPY7.70bn, RUTILEA/AI Fukushima ~JPY2.56bn, GMO ~JPY1.93bn) — distinct from the separate Economic Security Fund that financed ABCI 3.0, and from the ~JPY500bn 'K Program' (AI/quantum/semiconductors, run by the Cabinet Office with JST/NEDO). The live page should cite the Apr 2024 METI program by its exact name rather than the vague phrase. (2) 'NVIDIA partnership details' should be replaced with the concrete, dated, independently-confirmed fact: SoftBank's DGX SuperPOD (NVIDIA DGX B200, >4,000 Blackwell GPUs), operational since 22 Jul 2025 and independently confirmed as the world's largest operating NVIDIA DGX SuperPOD on DGX B200 at that date — not a generic 'partnership' label. All 16 indicators returned a non-NA value for Japan in this vintage (full weighted coverage); all within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published). Several supporting figures (SoftBank's ~JPY2 trillion 6-year data-centre plan, NTT's tsuzumi model details, Noetra's 1-trillion-parameter model-scale target, AWS/Oracle Japan investment totals, the Taiwan transshipment-diversion case, and IBM/ASML's role in Rapidus's process technology) are flagged PROVISIONAL in the underlying research — sourced but not independently re-verified — and are reflected in the relevant indicators' evidence_status rather than silently treated as VERIFIED_FACT.
India — 60.0
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-08
Pillar scores
Strongest capability
Real, multiply-corroborated operational compute base (34,000-38,000 empanelled GPUs, mostly Nvidia H100/H200, across 10 private providers under the IndiaAI Mission) combined with a large private buildout genuinely under construction: Reliance's 1GW Nvidia-Blackwell data center at Jamnagar, part of an announced $110B/7-year sovereign AI infrastructure commitment.
Principal dependency
Total structural dependence on foreign (overwhelmingly Nvidia, some AMD) chip supply, with no domestic AI-chip manufacturer and fabrication described as "a decade away at best"; compute ownership rests with a handful of private domestic conglomerates (Reliance, Tata, Jio, Yotta, Adani) rather than diversified public capacity.
Key catalyst
The India AI Impact Summit (16-21 Feb 2026, New Delhi), at which the government announced 20,000 additional sovereign GPUs and a 100,000-GPU target under 'AI Mission 2.0,' and Reliance announced its $110B, 7-year sovereign AI infrastructure commitment.
Key constraint
A large, well-documented announcement-to-execution gap: against the ₹10,371.92 crore (~$1.25B) five-year outlay approved by the Union Cabinet in March 2024, only ~₹400 crore (~4%, ~$48M) had actually been disbursed by early-mid 2026, and the FY2026-27 budget was cut rather than increased.
Constraint notes
Financing/execution capacity is the most acute documented constraint: only ~4% of the IndiaAI Mission's five-year outlay had actually been disbursed as of early-mid 2026, and the FY2026-27 budget allocation was cut (to ₹1,000 crore, down from ₹2,000 crore) even as GPU-count ambitions were being publicly scaled up. India also has no domestic AI-chip fabrication capacity and no dedicated bilateral export-control framework comparable to the UAE's, leaving it structurally dependent on ordinary commercial Nvidia/AMD channels.
Notes
The '$1.24B IndiaAI Mission' figure commonly cited elsewhere is essentially correct — more precisely $1.25B / ₹10,371.92 crore, the Union Cabinet-approved 5-year outlay (PIB, Mar 2024) — but that headline figure obscures a striking execution gap: only ~₹400 crore (~4% of the outlay, ~$48M) had actually been disbursed as of early-mid 2026 (₹21.79cr released FY2024-25 against a ₹173cr revised estimate; ₹379.15cr FY2025-26 against an ₹800cr revised estimate; source: MediaNama, citing a Parliamentary panel disclosure), and the Union Budget 2026-27 allocation was CUT to ₹1,000 crore, down from ₹2,000 crore in 2025-26, even as GPU-count ambitions (100,000 by late 2026) were being publicly scaled up. Two indicators are marked NA rather than scored, per methodology section 6 rule 1: power_grid_access (compute_capacity) — evidence of an emerging grid/power constraint is real (Draft National Electricity Policy 2026) but is not sourced to a single primary document and is explicitly flagged low-confidence in the underlying research; and export_control_exposure (compute_access_resilience) — India's Tier-2 classification under the Biden-era BIS AI Diffusion Rule became moot when BIS rescinded the entire rule in May 2025 before its effective date, and no successor India-specific framework has been identified, so India's current formal export-control exposure level is genuinely unknown rather than scoreable. A claim that the Mission's total outlay would double to ₹20,000 crore, found only in one low-confidence aggregator (explainx.ai) and uncorroborated by any primary/official source, is explicitly excluded as UNVERIFIED per methodology section 7. All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
Saudi Arabia — 70.0
Posture: Strategic Buyer · Direction: baseline · Coverage: 100% · Last verified: 2026-08-08
Pillar scores
Strongest capability
Unusually fast, dated, verifiable execution: HUMAIN moved from its founding (May 2025) through a $10B AMD collaboration, a Nov 2025 US chip-export authorization, and physical receipt of its first Nvidia GPU shipment by December 2025 -- roughly 7 months from framework to hardware, the fastest confirmed announcement-to-delivery cycle of any country in this index.
Principal dependency
Total dependence on US-origin compute (Nvidia and AMD, both under direct US export-control authority) with no domestic or allied chip-fabrication capacity identified in this research; the flagship 35,000-chip authorization is explicitly conditioned and monitored, not unconditional.
Key catalyst
The Nov 19, 2025 US Commerce Department authorization of up to 35,000 Blackwell-equivalent chip exports to HUMAIN, and PIF/Aramco's Oct 29, 2025 term sheet for Aramco to acquire a minority stake in HUMAIN while PIF retains majority ownership.
Key constraint
Chip access, while real and substantial, is explicitly conditioned on security/reporting requirements and anti-diversion monitoring, with a reported Huawei/China-divestment pledge whose precise legal form was not confirmed against primary US government text in this research pass.
Constraint notes
The Nov 19, 2025 US chip-export authorization (up to 35,000 Blackwell-equivalent chips) is conditioned on 'rigorous security and reporting requirements' and ongoing BIS compliance monitoring, and is publicly linked to a reported (though not primary-source-confirmed in exact legal language) HUMAIN pledge not to purchase Huawei equipment, plus ALAT's stated willingness to divest from China 'if the US asks.' This is real, negotiated access, not unconditional access.
Notes
REVISED after adversarial review (2026-08-08): (1) committed_buildout was originally scored 75 (VERIFIED_FACT) on the strength of '211 land plots secured' and a Q2 2026 operational target, but per methodology section 6 rule 4 ('financed does not equal under construction; committed capital without physical construction evidence -- site work, procurement, workforce reporting -- scores at most 50'), land acquisition and a target date do not themselves constitute construction evidence in the way, for example, the UAE's Stargate site (>5,000 workers on site, equipment procured) does. Rescored to 50 to apply the same evidentiary bar used for Brazil's directly comparable Rio AI City claim. This lowered compute_capacity from 58 to 50 and the overall score from 72 to 70. (2) Posture reclassified from Strategic Builder to Strategic Buyer: with compute_capacity now Moderate (50) and control_ownership High (81), and HUMAIN's primary pathway to compute being negotiated, government-authorized foreign chip access (the Nov 2025 BIS authorization) and sovereign-fund-backed offtake/JV deals (AMD, Nvidia) rather than large-scale domestic fabrication, Saudi Arabia matches the Strategic Buyer definition's own example language ('negotiated export-control access, sovereign-fund-backed offtake deals') more precisely than Strategic Builder's 'high/rising domestic compute capacity.' Saudi Arabia has the densest primary-source coverage of any thin-to-moderate-evidence country in this dataset: official government statistics (MCIT), a direct US government export authorization (BIS/Commerce), and multiple company press releases (PIF, Aramco, AMD, Cisco, NVIDIA) collectively support VERIFIED_FACT status on 12 of 15 scored indicators (80%), meeting the methodology section 10 'high' confidence threshold. The precise contractual/legal language of the reported Huawei-divestment condition could not be confirmed against primary commerce.gov/bis.gov text (both blocked automated fetch in this research pass); the fact that the pledge was publicly reported and repeated across independent outlets is treated as VERIFIED_FACT, but its exact binding legal form is flagged as unresolved. domestic_allied_fab_substitution (compute_access_resilience) is marked NA -- no evidence of Saudi domestic or allied chip-fabrication capacity was located. All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
United Arab Emirates — 70.2
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-08
Pillar scores
Strongest capability
The July 10, 2026 BIS reclassification of the UAE to Country Group A:5 (the highest US export-control trust tier) combined with Stargate UAE — a 1GW cluster (part of a planned 5GW, 19.2 km² campus) under construction in Abu Dhabi with named partners (OpenAI, Oracle, Nvidia, Cisco, SoftBank), up to $10B committed for the initial phase, backed by Barakah's fully operational 5.6GW nuclear power base.
Principal dependency
Structural dependence on Nvidia/US chip supply with no domestic chip design or fab capacity; G42, the UAE's central AI infrastructure vehicle, carries real, disclosed foreign capital and governance exposure — Microsoft's Intergovernmental Assurance Agreement secured a board seat and a security commitment for G42 to remove Chinese (Huawei) equipment, and Silver Lake, Microsoft, Dalio Family Office, and the Kai Foundation hold minority stakes alongside Royal Group's SEC-filed controlling position.
Key catalyst
The July 10, 2026 BIS reclassification of the UAE to Country Group A:5, and the June 14, 2026 establishment of a unified, Cabinet-reporting Federal Authority for Artificial Intelligence and Data, consolidating what had been fragmented AI/data governance bodies.
Key constraint
The A:5 export-control upgrade is conditional, not unconditional access: it is limited to specific approved entities and automatically sunsets for non-US-owned UAE AI firms 270 days after April 6, 2027 unless they 'become US companies.' Separately, Stargate UAE's flagship 1GW cluster remains under construction, not yet confirmed operational, as of the most recent reporting.
Constraint notes
The July 10, 2026 BIS Country Group A:5 upgrade — the UAE's single most important compute-access fact — is entity-limited (restricted to a named 'approved entities' list in Supplement No. 8) and carries a built-in sunset clause: eligibility for non-US-owned UAE AI firms expires 270 days after April 6, 2027 unless they 'become US companies,' an undefined and unresolved condition. Power is a secondary constraint: Barakah's 5.6GW nuclear base is largely already absorbed by existing demand, requiring incremental gas-turbine buildout to support data-center growth.
Notes
The single most important, most recent Pillar 3 fact for the UAE is the July 10, 2026 BIS reclassification of the UAE to Country Group A:5 — the highest US export-control trust tier, enabling license-free export under STA of advanced computing items including AI chips/servers (BIS official press release, corroborated by Morgan Lewis legal analysis). This is explicitly NOT unconditional access: it is entity-limited to a specific 'approved entities' list in new Supplement No. 8 (the UAE government/Ministry of Defense, two named UAE AI companies, and subsidiaries of certain US-headquartered hyperscalers operating in the UAE), and it carries a built-in sunset clause — eligibility for the two named non-US-owned UAE commercial AI firms automatically expires 270 days after April 6, 2027 unless they 'become US companies' by that date, an undefined term in the rule, conditioned on the UAE making 'matching investments in U.S. AI digital infrastructure buildout.' export_control_exposure is scored 75, not 100, specifically to reflect this conditionality, and should not be read as parity with an unconditional access regime. Four indicators are marked NA per methodology section 6 rule 1: operational_advanced_compute (compute_capacity) — Stargate UAE's first 200MW phase is confirmed under construction (>5,000 workers on site, equipment procured, targeted Q3 2026 completion) but not yet confirmed live/operational as of the most recent reporting located (The National, 5 Dec 2025), and no separate baseline figure for the UAE's pre-Stargate operational compute base was located; domestic_alternatives (control_ownership) and domestic_technical_research_base (institutional_data_sovereignty) — no evidence of a UAE-developed domestic foundation-model program or an operating national research institute analogous to India's Sarvam/Krutrim/BharatGen was located in the research assembled for this vintage; and domestic_allied_fab_substitution (compute_access_resilience) — no evidence of any UAE domestic or allied chip-fabrication capacity or roadmap was located. G42/MGX exact equity percentages and per-round dollar amounts (Mubadala's stake, Silver Lake's stake, MGX's specific position in OpenAI/xAI/Anthropic rounds) are not publicly disclosed; control_ownership indicators are scored on qualitative governance evidence (Sheikh Tahnoun's SEC-filed voting control, Microsoft's IGAA board seat) rather than false-precision percentages, per methodology section 15.5. All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
Brazil — 45.3
Posture: Strategic Builder · Direction: baseline · Coverage: 80% · Last verified: 2026-08-08
Pillar scores
Strongest capability
A genuine, functioning, domestically-controlled sovereign cloud: Serpro and Dataprev (both state-owned) operate Brazil's 'Nuvem de Governo' (Government Cloud), with R$5.7 billion of the national AI plan (PBIA) specifically earmarked for sovereign-cloud/supercomputer infrastructure, and the Santos Dumont supercomputer (18.85 petaflops after a 2025 upgrade) is now the most powerful in Latin America.
Principal dependency
Brazil imports close to $50 billion/year in chips with almost no domestic production, and its largest commercial AI compute buildouts (Rio AI City, the ByteDance Pecem project) are led by foreign private capital (I Squared Capital, ByteDance) rather than Brazilian public or private domestic ownership.
Key catalyst
The Plano Brasileiro de Inteligencia Artificial (PBIA) 2024-2028's final version, published June 2025, and the Senate's unanimous Dec 10, 2024 passage of PL 2338/2023 (Brazil's AI regulatory bill), plus ANPD's Sep 2025 conversion into a full independent regulatory agency.
Key constraint
A large, well-documented announcement-to-execution gap: against the R$23 billion (~$4B) five-year PBIA outlay, BNDES had approved only about R$1 billion in AI-related financing through mid-2025, and BNDES's own dedicated AI/data-center investment fund had only reached a formal manager-selection stage (April 2026), not full capitalization or launch.
Constraint notes
Brazil received no tariff exemption for GPUs -- unlike the negotiated access secured by Saudi Arabia and the UAE, Brazil's government maintained taxation on GPUs even while expanding exemptions for other products, and a 40% IEEPA tariff stacks with underlying import duties on AI chips. Combined with near-total ($50B/year) chip-import dependency, this is a materially less favorable compute-access position than the Gulf comparators in this index.
Notes
REVISED after adversarial review (2026-08-08): posture (Strategic Builder) is based on compute_capacity plus partial, sub-threshold control_ownership evidence (the two scored indicators -- domestic_public_ownership_share 75, foreign_capital_exposure 25 -- average a Moderate 50), not a fully-scored control pillar; this is flagged explicitly per the adversarial review's finding that posture should not be asserted as if derived from a pillar score that does not exist. Overall weighted evidence coverage is 80%, meeting the score_display_threshold exactly (methodology section 9), so a normal numeric score is displayed. control_ownership did not clear the 60% indicator-coverage floor (2 of 4 indicators scored) and is marked insufficient_data at the pillar level -- no evidence of a negotiated hyperscaler agreement or a named Brazilian AI model/infrastructure alternative was located. power_grid_access (compute_capacity) is marked NA -- unlike Nigeria and Kenya, no specific Brazil national-grid capacity/reliability figure was located in the research assembled for this vintage; this should not be read as implying grid access is unconstrained, only that it was not evidenced. Brazil's PL 2338/2023 AI bill status is precisely characterized as Senate-passed (Dec 10, 2024), Chamber-pending, not yet enacted, as of Aug 2026 -- exact future vote-timing claims found in secondary sources were inconsistent across outlets and are treated as PROVISIONAL at most. All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
Moldova — Not scored
Posture: Import-Dependent · Direction: baseline · Coverage: 45% · Last verified: 2026-08-08
Pillar scores
Strongest capability
An unusually current, active institutional trail on AI policy specifically: signed the Council of Europe AI Framework Convention (Sep 2024), released a first national AI/data-governance policy White Paper (Oct 2024), launched a coordinating Sub-Council for AI (SCIA, Apr 2026) and a public-private Moldova AI Hub (Sep 2025) -- rare for a country this small.
Principal dependency
No domestic advanced-compute/AI-HPC infrastructure was found in this research; Moldova appears entirely reliant on cross-border cloud access (no in-country hyperscaler regions), with only legacy state-telecom hosting (Moldtelecom) as domestic infrastructure.
Key catalyst
A EUR 200 million high-tech park (offices, labs, microelectronics/IT innovation infrastructure) funded via the EU Development Plan plus private participation, with construction targeted to start in 2026.
Key constraint
The high-tech park will not be operational until 2030 at the earliest, and is a general tech park rather than confirmed AI/HPC infrastructure; there is no confirmed near-term path to domestic AI compute capacity.
Constraint notes
Geopolitical/energy-security exposure is the dominant, well-evidenced constraint, distinct in kind from the other thin-evidence countries: the Dec 2024 Russian gas cutoff to Transnistria produced a documented 2025 energy crisis (Transnistrian GDP -18%, inflation 14.7%), with grid-redundancy interconnectors not due to complete until 2026-2029. This is a power/energy-security and geopolitical constraint, not a compute-hardware or export-control one -- no evidence of chip export-control exposure was found.
Notes
Overall weighted evidence coverage is 45%, below the 50% provisional threshold (methodology section 9), so no numeric overall score is displayed -- only the two pillars that cleared the 60% indicator-coverage floor (compute_capacity and institutional_data_sovereignty) are scored; control_ownership, compute_access_resilience, and financing_execution are all marked insufficient_data at the pillar level because fewer than 60% of their indicators had locatable evidence. This is the correct methodology outcome for a country where the honest finding across most indicators is 'no evidence found,' per methodology section 6 rule 1 (no evidence = NA, not zero) and section 9's coverage-gated display rule. One financing claim (a EUR 1.9 billion EU support package reportedly conditioned on AI integration) was explicitly excluded as UNVERIFIED -- it traced to a single lower-tier outlet (Pravda Moldova) with no EU or Moldovan government primary source found to corroborate it, per methodology section 7.
Nigeria — 43.8
Posture: Import-Dependent · Direction: baseline · Coverage: 60% · Last verified: 2026-08-08
Pillar scores
Strongest capability
Real, dated, dollar-figured private-sector investment is actively flowing: MTN's Sifiso Dabengwa Data Centre Phase 1 (4.5MW, Tier III) began operations July 2025, and Airtel's Nxtra $120M hyperscale facility has already received early GPU shipments -- this is genuine build activity, not just announcements.
Principal dependency
No confirmed Nigerian public/state equity stake in any AI or cloud infrastructure asset was located; Nigeria depends entirely on cross-border access to South Africa-region hyperscaler infrastructure (AWS Cape Town, Azure South Africa North, GCP Johannesburg), with no full hyperscaler region in-country.
Key catalyst
NITDA's Sovereign Cloud Framework / National Cloud Investment Strategy, signed Aug 5, 2026, and a wave of dated 2026 project announcements (MTN Phase 2 AI-GPU buildout, Tetracore's $400M Ogun State facility).
Key constraint
Chronic, severe grid unreliability -- ~40% transmission/distribution losses and repeated full grid collapses -- is explicitly cited by industry as a binding risk to AI-data-center buildout plans, independent of available private capital.
Constraint notes
Nigeria's power grid is the most robustly evidenced constraint in the entire v1.0 dataset: installed generation capacity of ~12,500MW against actual delivered power that rarely exceeds 5,000MW, transmission/distribution losses as high as 40% (vs. a global standard of 8-12%), and four major grid collapses in 2025 (one crashing generation from 3,660MW to 50MW). Industry executives explicitly cite unreliable power as a risk to Nigeria's AI-data-center ambitions.
Notes
Overall weighted evidence coverage is 60%, in the 50-80% provisional band (methodology section 9), so a numeric score is displayed but flagged Provisional. control_ownership and compute_access_resilience did not clear the 60% indicator-coverage floor and are marked insufficient_data at the pillar level -- in both cases the honest finding was 'no evidence of a Nigerian state equity stake or specific export-control treatment was located,' left NA per methodology section 6 rule 1 rather than assumed. power_grid_access is scored 0 (not NA) because the evidence of inadequate power access is itself strongly verified (World Bank, multiple independent Nigerian financial/trade press) -- this is a positively-established negative finding, distinct from an absence of evidence. All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
Vietnam — 59.4
Posture: Strategic Builder · Direction: baseline · Coverage: 80% · Last verified: 2026-08-08
Pillar scores
Strongest capability
The most comprehensive, dated, multi-year institutional and legal apparatus among the four originally 'thin-evidence' countries: a national semiconductor strategy through 2050 (Decision 1018/QD-TTg, Sep 2024), a new Cybersecurity Law (effective Jul 2026) and Law on Data (effective Jul 2025) both explicitly creating data-localization and cross-border-transfer restrictions, plus genuine multi-vendor infrastructure diversification (Microsoft/G42, ST Telemedia/VNG, domestic KBC/VietinBank all actively building).
Principal dependency
Vietnam's real, dated semiconductor investment (Samsung, Amkor, Intel -- $1.5B+ and $1.6B+ respectively) is confirmed to be concentrated in assembly/packaging/testing (back-end, including HBM integration) rather than front-end AI-chip fabrication or design -- SEMI's own framing describes Vietnam as 'From ATP Workhorse to Front-End Contender,' i.e., aspiring to but not yet possessing fabrication capability.
Key catalyst
Decision 1018/QD-TTg's semiconductor strategy through 2050 (explicitly positioning Vietnam as a supply-chain-diversification '+1' alternative to China) and a Dec 31, 2024 government decree establishing a special-incentive investment fund for AI/semiconductor science and technology.
Key constraint
All confirmed chip-related foreign investment (Samsung, Amkor, Intel) is back-end assembly/packaging/testing, not front-end fabrication -- a real capability gap that should not be conflated with fab capacity when reading the headline dollar figures.
Constraint notes
Vietnam faces a geopolitical constraint distinct in kind from the power/energy constraints affecting Nigeria and Kenya: its position between the US and Chinese chip ecosystems creates transshipment/enforcement exposure even without being a direct export-control target (no source found documenting Vietnam itself as sanctioned or restricted -- this remains an inferential risk, not a confirmed one). A more moderate, actively-mitigated power constraint also exists: 2026 demand growth scenarios of 8.5-14.1% are being met with grid-scale battery storage deployment (530MW Northern, 275MW Hanoi) rather than chronic shortfall.
Notes
REVISED after adversarial review (2026-08-08): committed_buildout was originally scored 75 (VERIFIED_FACT) for the $2B Tan Phu Trung project, but the underlying claim text contains no site-work, procurement, or workforce evidence -- only financing and a scheduled date -- so per methodology section 6 rule 4 it is capped at 50, matching the bar applied to Brazil's directly comparable Rio AI City claim. This lowered compute_capacity from 63 to 50 and the overall score from 64 to 59. Overall weighted evidence coverage is 80%, meeting the score_display_threshold exactly (methodology section 9), so a normal numeric score is still displayed. control_ownership did not clear the 60% indicator-coverage floor (only 2 of 4 indicators had locatable evidence: VietinBank's co-financing role and VNG's joint-venture stake, which together average a Moderate 50) and is marked insufficient_data at the pillar level -- posture (Strategic Builder) is therefore based on compute_capacity plus this partial, sub-threshold control evidence, not a fully-scored control pillar; this is flagged explicitly per the adversarial review's finding that posture should not be asserted as if derived from a pillar score that does not exist. Two indicators were explicitly excluded from scoring per methodology section 7: export_control_exposure, because the research itself flagged Vietnam's geopolitical chip-ecosystem exposure as speculative/inferential rather than a documented restriction, and domestic_technical_research_base (institutional_data_sovereignty), for which no specific citable claim was located. The user's specific verification request -- whether Intel/Samsung/Amkor facilities in Vietnam are AI-chip-relevant fabrication or merely packaging/testing -- was directly confirmed: all three are back-end (assembly/test/packaging) operations, not front-end fabrication, per SEMI's own analysis. All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
Kenya — 46.9
Posture: Import-Dependent · Direction: baseline · Coverage: 60% · Last verified: 2026-08-08
Pillar scores
Strongest capability
Kenya has the strongest confirmed OPERATIONAL domestic AI-relevant infrastructure among the four thin-evidence countries: the Konza Technopolis data center (Huawei-built, Tier III, state-owned via the Konza Technopolis Development Authority) has been running since ~2020, serving 120+ industry clients and 70+ government applications, financed via a Chinese concessional loan.
Principal dependency
The flagship Microsoft/G42 $1 billion Olkaria project -- Kenya's path to hyperscale AI compute -- is foreign-owned (UAE's G42 plus Microsoft) with no confirmed Kenyan government equity stake, and is explicitly reported as stalled over power availability.
Key catalyst
Kenya's National AI Strategy 2025-2030 (launched March 2025) and a further, broader 'Artificial Intelligence and Other Emerging Technologies Policy' released for public consultation through Aug 4, 2026 -- essentially concurrent with this research, indicating active, current policy momentum.
Key constraint
The Microsoft/G42 Olkaria project's reported power-availability 'limbo' directly demonstrates that grid capacity, not capital, is the binding constraint on Kenya's largest named AI infrastructure commitment.
Constraint notes
Kenya's power grid is directly and specifically evidenced as binding on a named, dollar-figured AI project: the Microsoft/G42 $1 billion Olkaria data center campus (100MW initial capacity) is reported to be 'in limbo over power concerns,' with one estimate suggesting the project's full scale could require up to 1,000MW against Kenya's total national generation capacity of only ~2,300MW. This is a rare case where a specific constraint and a specific stalled project are both independently documented together.
Notes
Overall weighted evidence coverage is 60%, in the 50-80% provisional band (methodology section 9), so a numeric score is displayed but flagged Provisional. control_ownership and compute_access_resilience did not clear the 60% indicator-coverage floor and are marked insufficient_data at the pillar level. The '1,000MW requirement' figure for the Olkaria project rests on a single outlet and is treated as illustrative/contested rather than a hard figure; the underlying 'in limbo' status itself is corroborated. power_grid_access is scored 0 (not NA) because the evidence of the constraint directly binding on a named project is itself strongly verified. All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
China — 75.4
Posture: Sovereign Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-08
Pillar scores
Strongest capability
Extraordinary state financing scale paired with real (not merely announced) execution: Big Fund Phase III (Y344bn / ~$47.5B registered capital, incorporated May 2024) alongside operational deployment such as Alibaba's in-house-chip AI data center and SMIC's 7nm node in mass production.
Principal dependency
Near-total exposure to US/allied export controls on the most advanced compute and lithography: Nvidia H20, AMD MI325X/MI308, and ASML EUV systems are all restricted or license-gated, forcing reliance on domestic substitutes (Huawei Ascend, SMIC sub-7nm) whose real output is bottlenecked by HBM memory supply, not just fabrication capability.
Key catalyst
Big Fund Phase III's 2026 strategic pivot toward advanced packaging, semiconductor equipment, and AI chips specifically (reported Aug 7, 2026), and Alibaba's operational deployment of an in-house-chip AI data center (April 2026).
Key constraint
Domestic HBM substitution capacity (CXMT) is projected to fall roughly 5-6x short of pre-restriction stockpile-equivalent supply next year, a quantified ceiling on real AI-chip output regardless of SMIC/Huawei fabrication progress.
Constraint notes
China is the single most export-control-restricted country in the index: Nvidia H20 remains license-gated indefinitely (no licenses granted as of research date), AMD MI325X/MI308 are license-gated or pending, and ASML EUV lithography is fully banned. Domestic substitution is real but resource-constrained: China's pre-restriction HBM stockpile (~13M stacks, sufficient for ~1.6M Ascend 910C packages) is projected to deplete, with domestic substitute CXMT able to supply only ~250,000-300,000 chips'-worth next year — a binding, quantified bottleneck independent of chip-fabrication progress. READER NOTE ON EVIDENCE QUALITY: two compute_capacity indicators (committed_buildout and financing_execution's site_readiness_signal) rely on Chinese state media (China Daily) as their only source for headline scale figures (780,000 PFlops, >CNY200bn hub investment) -- those specific numbers are down-weighted and scored conservatively rather than taken at face value, per methodology section 7's caution against treating state media as sufficient for VERIFIED_FACT status. China's overall score_status is 'scored' (coverage clears 80%), but confidence is deliberately capped at 'medium' rather than 'high' for exactly this reason.
Notes
This profile follows the researching agent's explicit recommendation: score China as Provisional-quality evidence via the confidence field (set to medium) and granular notes, rather than forcing the entire country to score_status='provisional', because coverage genuinely clears the 80% weighted-coverage threshold for a normal scored display (methodology section 9 ties score_status to evidence COVERAGE, not evidence CONFIDENCE, which is a separate field for exactly this situation). Two indicators are marked NA per methodology section 6 rule 1: hyperscaler_bargaining_leverage (control_ownership) — this indicator does not map cleanly onto China, whose largest AI infrastructure operators (Alibaba, Tencent, Baidu) are themselves domestic hyperscalers rather than negotiating access from foreign ones, and no evidence was gathered framing this as a negotiated-leverage question; and domestic_technical_research_base (institutional_data_sovereignty) — the research assembled for this vintage did not include a specific, citable claim about China's national AI research-institute base distinct from the corporate R&D already captured elsewhere, so it is left unscored rather than assumed. SMIC's 5nm yield rate is explicitly EXCLUDED from scoring: estimates in circulation span roughly 20-70% with no verifiable primary source (SMIC does not disclose node-level yields), so domestic_allied_fab_substitution is scored on the independently verifiable 7nm mass-production status only, not on any 5nm yield claim. Huawei's 'state control' is genuinely contested in Western scholarship (formal ownership is an employee trade-union structure; whether that structure is functionally state-controlled is disputed between credentialed scholars on both sides, not a China-vs-West framing) -- domestic_public_ownership_share is scored primarily on SMIC's independently-documented state-linked shareholder register (Big Fund, Datang, state banks), which is on firmer evidentiary ground than the Huawei-control argument. East Data West Computing's headline PFlops and investment totals (780,000 PFlops, >CNY200bn hub investment) are sourced only to Chinese state media (China Daily); the underlying NDRC-established program is independently documented as real, so committed_buildout is scored on the program's existence at a moderate PROVISIONAL confidence, not on the state-media scale figures, consistent with methodology section 7's caution against treating state media as sufficient for VERIFIED_FACT status. All within-pillar indicators are equally weighted in v1.0 (no sub-indicator weighting scheme yet published).
South Korea — 65.4
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
An unusually large, coordinated, and dated public-private capital commitment: government and private-sector investment orchestration exceeding 1,350 trillion won (~$880B, Samsung/SK Hynix chip and data-center buildout announced June 29, 2026), a 2026 government budget with direct AI spending up over 300% YoY to 10.1 trillion won, and a 550 trillion won plan for 8.4GW of AI data-center capacity -- combined with world-leading domestic HBM memory production (Samsung, SK Hynix) that is itself critical, currently-shipping AI-chip infrastructure.
Principal dependency
Nvidia is both the near-universal hardware supplier for Korea's announced AI-factory buildout (260,000+ GPUs committed across Naver, SK, Samsung, Hyundai, and the National AI Computing Center) and, as of a $1B investment announced July 24, 2026, a direct 4.5% equity holder in Naver -- Korea's leading domestic AI/cloud company -- making Korea simultaneously supplier-concentrated and directly capital-exposed to the same foreign counterparty.
Key catalyst
Groundbreaking of the National AI Computing Center at Haenam (Aug 3, 2026), after two failed bidding rounds forced the government to drop a domestic-NPU installation quota and restructure equity terms before Samsung SDS's consortium (with Naver Cloud, Kakao, Samsung C&T, Samsung Electronics, and the National Growth Fund) agreed to build it under a 71%-private/29%-government ownership split.
Key constraint
A severe capital-region grid-interconnection bottleneck, only partially addressed by the still-not-yet-effective (Feb 2027) AI Data Center Special Act, is forcing new capacity into satellite regions and constraining how fast financed/announced GW-scale projects can actually energize.
Constraint notes
Two documented, binding constraints: (1) grid access -- Seoul-region data-center vacancy fell to 1.1% in H1 2026 and grid-impact-assessment approvals in the capital region have been severely bottlenecked, prompting a May 7, 2026 special law (effective Feb 2027) that fast-tracks AI data centers as national strategic facilities and exempts some from grid-impact review; (2) supplier concentration -- virtually all announced GPU-scale AI-factory capacity (Naver/SK Telecom/Samsung/Hyundai, 260,000+ units) is single-sourced to Nvidia, with Samsung's own Mach-1 inference accelerator still pre-scale.
Notes
Coverage is 100% (all 16 indicators across all 5 pillars returned locatable, citable evidence; none required null), so no pillar redistribution was needed and score_status is 'scored' (>=80% weighted coverage). Confidence is 'medium' rather than 'high': 11 of 16 indicators (68.75%) are VERIFIED_FACT, below the 70% bar, even though Pillar 1 (compute_capacity) has two of three indicators VERIFIED_FACT with multi-source corroboration. Posture judgment call: compute_capacity scored 50 (Moderate band) and control_ownership scored 63 (Moderate band, not High) -- under the methodology's bands this is 'Strategic Builder' (building at scale via massive financed buildout, but not yet fully controlling the stack given Nvidia's near-universal hardware role and direct equity stake in Naver). This is distinct from 'Sovereign Builder' (which requires control_ownership to also clear the High >=70 band) and from 'Strategic Buyer' (which requires control to be High via negotiated foreign access rather than domestic construction -- not the case here, since Korea's buildout is overwhelmingly domestic-construction-led). One clarification on operational_advanced_compute: Naver's flagship GAK Sejong 'AI factory' (the Nvidia DSX-platform buildout most heavily covered in press) is NOT yet operational -- it is planned to begin operating at 55MW in H1 2027, expanding to 200MW by 2028 -- so that project was NOT used as the basis for the operational_advanced_compute score; SK Telecom's smaller 'Haein' GPU-as-a-service cluster (live since Aug 1, 2025) was used instead as the genuinely-operational evidence.
Singapore — 59.1
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-11
Pillar scores
Strongest capability
A real, operating national supercomputing base (NSCC Singapore's ASPIRE 2A+ delivering 20 PetaFLOPS on Nvidia H100s, with the next-generation ASPIRE 2B launched June 8, 2026) combined with a sophisticated, government-controlled data-center capacity-allocation regime: Singapore ended its 4-year data-center moratorium in Dec 2025 and now grants new capacity only to operators meeting binding green-power and efficiency conditions, giving the state real negotiating leverage over hyperscalers rather than simply granting open access.
Principal dependency
Every entity awarded new data-center development rights in the Dec 2025 moratorium-lift decision -- Equinix, Microsoft, GDS, and a consortium of AirTrunk and ByteDance -- is foreign-headquartered or foreign-linked, and AWS alone has committed $12B (2024-2028) to Singapore infrastructure; Singapore's AI compute buildout is overwhelmingly foreign-capital-built on domestic soil rather than domestically owned.
Key catalyst
The Dec 2025 end of the data-center moratorium plus the launch of the Data Centre Call-for-Application 2 (DC-CFA2, a 200MW allocation round, applications closed March 31, 2026), alongside the May 2026 National AI Strategy 2.0 update and a new National AI Council chaired by Prime Minister Lawrence Wong.
Key constraint
Grid/power scarcity remains the fundamental limiting factor behind Singapore's entire data-center policy regime, and newly-scrutinized export-control compliance risk (the Aperia Group smuggling case, Nvidia's July 2026 Asian buyer-list cuts) adds a second, more recent constraint specific to chip access.
Constraint notes
Two distinct, well-documented constraints: (1) structural electricity scarcity -- data centers were already ~7% of Singapore's total electricity consumption before the 2019-2022 moratorium and are projected to reach ~24% by 2030, which is why new capacity releases (DC-CFA2) carry binding conditions (>=50% green power, PUE<1.3); (2) export-control/compliance exposure -- Singapore-based entities (the Aperia Group: Aperia International, A-Speed Infotech, Aperia Cloud Services) have been criminally charged in an alleged US$2.5B Nvidia-chip diversion scheme to China, and Nvidia named Singapore (with Malaysia and Japan) as a region of intensified compliance screening when it cut more than half its approved Asian AI-chip buyers in July 2026.
Notes
Coverage is 100% (all 16 indicators across all 5 pillars returned locatable, citable evidence), so score_status is 'scored' with no pillar redistribution needed. Confidence is 'medium': only 10 of 16 indicators (62.5%) are VERIFIED_FACT, below the 70% bar, partly because the widely-repeated '$27B AI infrastructure by 2030' headline figure traces to a lower-tier aggregator (Introl) blending a $740M government figure with an imputed ~$26B private-sector total rather than one primary disclosure -- committed_buildout and committed_capital were scored ANALYTICAL_INFERENCE accordingly rather than VERIFIED_FACT, even though the underlying component facts (AWS $12B 2024-2028, the moratorium-lift awards) are independently well-sourced. Posture is 'Strategic Builder' by the literal numeric bands (compute_capacity=58 Moderate, control_ownership=63 Moderate/not-High), but this is a judgment call worth flagging: Singapore's 'building at scale' is overwhelmingly foreign-capital-driven construction happening on Singaporean soil under a government-negotiated hosting/allocation model (the moratorium and DC-CFA2 process), not domestic-owned construction in the way Korea or Taiwan build -- it reads closer in spirit to a sophisticated hosting/brokering posture than a builder posture, even though it does not cleanly clear Strategic Buyer's numeric control=High threshold (domestic_public_ownership_share of just 25 pulls control_ownership down to Moderate). A human editor may reasonably prefer to override this to Strategic Buyer given the qualitative fit; I kept the literal band result since Strategic Buyer's defined threshold (control High) was not numerically met.
Indonesia — 55.4
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-11
Pillar scores
Strongest capability
Real, operational, government-run AI/HPC research infrastructure: BRIN (Indonesia's National Research and Innovation Agency), in partnership with South Korea's KISTI, inaugurated the ASEAN-Korea HPC facility in Jakarta on June 18, 2026 (4.28 petaflops peak, 3.10 petaflops tested, a TOP500-candidate submission) -- combined with substantial, named, dated capital specifically earmarked for data centers (Danantara's Rp21 trillion/~$1.36B 'Fukuoka Project,' part of a Rp202.4 trillion/$13.1B 2026 investment plan) and large secured grid capacity (BDx's 1.2GW PLN power commitment, June 2026).
Principal dependency
Indosat Ooredoo Hutchison -- the telecom partner central to Indonesia's most visible AI-infrastructure initiatives (the Nvidia/Cisco-backed national AI Center of Excellence, BDx's 'first sovereign AI data center,' and the Sahabat-AI language model) -- is 65.644% owned by Ooredoo Hutchison Asia Pte Ltd, a Qatar-Hong Kong joint venture, with the Indonesian government retaining only 9.6% and domestic PT Tiga Telekomunikasi Indonesia holding 10.8%.
Key catalyst
BRIN and KISTI's inauguration of the ASEAN-Korea HPC facility (June 18, 2026) plus Danantara's Feb 13, 2026 unveiling of a data-center-specific investment tranche within its broader 2026 investment plan.
Key constraint
Indonesia's data-center power buildout, while attracting real large-scale PPAs in the near term, faces documented transmission-bottleneck and 2030 power-shortfall risk according to industry market analysis, and Indonesia's AI-infrastructure supply chain has been directly implicated (via Indosat) in a documented Nvidia-chip routing case to China.
Constraint notes
Three documented constraints: (1) export-control circumvention exposure -- the Wall Street Journal reported (Nov 14, 2025) that Indosat served as the intermediary through which a Chinese AI firm (INF Tech) obtained Nvidia Blackwell-equivalent compute (~2,300-GPU equivalent) from Aibres via a regulatory gap, which raises Indonesia's profile as a routing point even though the specific transaction was not deemed a violation; (2) majority foreign ownership of the telecom partner central to nearly every major named AI-infrastructure project (Indosat Ooredoo Hutchison, 65.644% owned by a Qatar-Hong Kong joint venture); (3) medium-term power-delivery risk -- large individual power purchase agreements are being signed now (BDx 1.2GW, DayOne ~450MW), but industry analysis flags transmission bottlenecks and a possible power shortfall by 2030.
Notes
Coverage is 100% (all 5 pillars cleared the 60% non-null-indicator threshold, so all are 'scored'; no pillar redistribution needed). Two of the 16 indicators are honestly null: hyperscaler_bargaining_leverage (no evidence found of Indonesia negotiating specific procurement terms, board seats, or security conditions with a hyperscaler or Nvidia -- the AI Center of Excellence and PLN power deals are real but are not documented as negotiated-leverage instruments in that sense) and domestic_allied_fab_substitution (no evidence found of any domestic or allied semiconductor fabrication capacity in Indonesia). Confidence is 'medium': 10 of 14 non-null indicators (71%) are VERIFIED_FACT, which technically clears the 70% bar, but Pillar 1 (compute_capacity) indicators rely on single primary/near-primary sources (BRIN's own site, Antara's Danantara coverage) rather than multi-source corroboration -- applying the same strict standard the USA reference file applied to itself, confidence was kept at 'medium' rather than upgraded to 'high'. One claim was deliberately downweighted rather than excluded: an Indonesian-language Kompas report describing a restrictive GPU export quota for Indonesia under the (subsequently rescinded) 2025 AI Diffusion Rule was treated as lower-tier/single-source supporting color, not the primary basis for export_control_exposure's score -- the better-corroborated, WSJ-reported Indosat/Aibres/INF Tech routing case was used as the load-bearing evidence instead. Posture is 'Strategic Builder' via a clean, non-overridden numeric-band match: compute_capacity (58) and control_ownership (58) are both in the Moderate (40-69) band, fitting 'building at scale but not fully controlling it' without requiring a qualitative override.
Taiwan — 75.0
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
Taiwan is the physical fabrication base for the world's most advanced AI accelerators: TSMC's 2nm node entered mass production in late 2025 with five fabs ramping through 2026, and 3nm monthly capacity is projected to reach 180,000-250,000 wafers by 2026-27 -- a domestic, operational, at-scale capability no other country in this assignment set has.
Principal dependency
TSMC, Taiwan's most systemically important AI-infrastructure company, is roughly 73-75% foreign-institutionally owned (Vanguard ~3.2%, BlackRock ~2.1%, plus GIC and Norway's sovereign fund), versus the Taiwan government's National Development Fund holding just 6.38% as the single largest shareholder -- meaning the entity underpinning Taiwan's compute-access leverage is itself majority foreign-held at the shareholder level, even though operational and technological control remains domestic.
Key catalyst
Nvidia's first overseas headquarters, the 'Constellation' campus in Taipei's Beitou Shilin Technology Park, secured via a 50-year land-lease deal finalized with Taipei City Government (Feb 2026) and MOEA investment approval (Jan 2026), alongside Jensen Huang's public commitment that Nvidia's annual Taiwan spending is moving from ~$100B toward $150B/year.
Key constraint
Nano4, the flagship national AI supercomputer at NCHC's new Tainan Cloud Computing Center, is a real but modest-scale operational asset (15MW, ~1,760 H200 + 144 Blackwell GPUs) relative to the GW-scale AI factories being built by larger economies -- Taiwan's own domestically-operated compute remains well behind its fabrication capacity.
Constraint notes
Two documented constraints, plus a widely-analyzed but not indicator-scored geopolitical risk: (1) localized grid strain -- Taipower projects semiconductor+AI electricity demand will add over 5GW between 2026-2030 concentrated in northern metro grids/substations, though data centers remain under 1% of total national electricity use so far; (2) supplier concentration in chip design/architecture and lithography equipment (Nvidia GPU IP, ASML tools) even though wafer fabrication itself is domestic. Separately -- not used in any indicator score, since it is a contextual risk rather than a documented evidentiary claim about a specific indicator -- multiple 2026 industry-risk analyses (e.g. supply-chain-resilience trade press) flag cross-strait tension as a severe, if not judged imminent, tail risk to Taiwan's semiconductor infrastructure, given Taiwan's outsized ~60% share of global advanced-chip production.
Notes
Coverage is 100% (all 16 indicators across all 5 pillars returned locatable, citable evidence; none required null), so no pillar redistribution was needed and score_status is 'scored'. Confidence is 'high': 13 of 16 indicators (81%) are VERIFIED_FACT, above the 70% bar, and Pillar 1 (compute_capacity) has multi-source corroboration on both operational_advanced_compute (NSTC primary source plus DigiTimes/press coverage) and committed_buildout. Posture is a genuine judgment call flagged for review: the numeric bands (compute_capacity=58, Moderate; control_ownership=81, High) technically land in the 'Strategic Buyer' bucket (Moderate-or-Low compute + High control), but Strategic Buyer's defined mechanism is control achieved 'through negotiated foreign access... rather than domestic construction,' which does not describe Taiwan -- Taiwan's control score is high because it wholly, publicly owns its flagship compute asset (NCHC) and because it has genuine bargaining leverage as the world's dominant upstream chip *supplier* (evidenced by Nvidia siting its first overseas HQ there), not because it bought access to someone else's infrastructure. I classified Taiwan as 'Strategic Builder' instead, since it is building real (if not yet globally-massive-scale) domestically-owned compute capacity while retaining unusually strong structural control -- but this is an override of the raw band match and should be double-checked by a human editor. compute_capacity was kept at Moderate rather than High specifically because Taiwan's own operating AI data-center/GPU-cluster footprint (Nano4's ~1,900 GPUs, GMI Cloud's 7,000-GPU factory) is small next to Korea's 260,000+ GPU commitments or US hyperscaler scale; TSMC's fabrication dominance is scored separately under compute_access_resilience (domestic_allied_fab_substitution), consistent with how the USA reference file treats TSMC Arizona.
Australia — 51.2
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
Genuinely operational, named domestic compute infrastructure (CSIRO's 448-H100 'Virga', Pawsey's Setonix, and commercial GPUaaS such as Sharon AI's 1,024-GPU Blackwell Ultra cluster) combined with one of the world's largest financed data-centre investment pipelines (Firmus Project Southgate: US$10bn Blackstone-led debt plus US$2bn equity, 2026; Microsoft A$25bn and AWS A$20bn multi-year commitments).
Principal dependency
The country's largest AI-infrastructure platforms are majority foreign-owned or foreign-financed -- AirTrunk (Australia's largest data-centre platform, 1.8GW) was acquired by a Blackstone-led consortium for A$24bn in Dec 2024, and Firmus's growth is financed by Blackstone, Coatue and NVIDIA -- while chip supply is concentrated in NVIDIA plus three US hyperscalers with no domestic or allied fab substitution.
Key catalyst
The government's 'Expectations of data centres and AI infrastructure developers' framework (23 Mar 2026) and the tied Australia-Microsoft MoU (23 Apr 2026, alongside the A$25bn investment), which for the first time set explicit national-interest, security and grid-cost conditions on hyperscaler investment, even though both instruments remain non-binding.
Key constraint
Grid-connection queues (~2-year timelines per AEMO) and the lapse of Australia's formal Tier-1 export-control status (rescinded May 2025, no replacement framework as of Aug 2026) are the two binding, documented constraints on further scale-up and on chip-access certainty.
Constraint notes
Two documented, binding constraints: (1) grid/interconnection -- AEMO reports ~2-year application-to-energisation timelines for large loads, with data centres growing from <2% to a projected ~32% of NEM transmission-connected demand by 2030, prompting an active AEMC rule-change process (draft determination 12 Mar 2026); (2) export-control uncertainty -- the US AI Diffusion Rule that gave Australia formal Tier-1 (license-free) chip access was rescinded in May 2025, and as of Aug 2026 no replacement tier framework is in force, leaving chip access to discretionary US case-by-case licensing rather than a codified allied status.
Notes
All 15 indicators returned locatable evidence (coverage_pct = 1.0; no pillar fell below the 60% non-null floor). Confidence set to 'medium' rather than 'high': applying the same formula used for USA, only 6 of 16 total indicator fields are VERIFIED_FACT (37.5%, well below the 70% bar), and on Pillar 1 specifically only 1 of 3 indicators (committed_buildout) is VERIFIED_FACT -- the other two (operational_advanced_compute, power_grid_access) are ANALYTICAL_INFERENCE/require synthesis across sources, so the 'high' threshold's Pillar-1 multi-source-VERIFIED_FACT requirement is not met. Posture judgment: compute_capacity scored 58 (Moderate band, 40-69) and control_ownership scored 38 (Low band, <40, driven by CDC's largest shareholder being foreign-listed Infratil and by Blackstone's majority/financing control of AirTrunk and Firmus); per the posture rules this is Moderate-compute paired with control that is not High, which maps to 'Strategic Builder' (building real domestic capacity at increasing scale without yet controlling the dominant share of it) rather than 'Strategic Buyer' (which requires control to be High via negotiated foreign access, which is not the case here -- Australia's negotiated instruments, the Expectations framework and Microsoft MoU, are explicitly non-binding and do not constitute achieved control). export_control_exposure was marked PROVISIONAL rather than ANALYTICAL_INFERENCE because sources actively conflict on Australia's current status: higher-tier legal/policy sources (CRS, Baker McKenzie, US Studies Centre) state the AI Diffusion Rule and its Tier-1 allocation were rescinded in May 2025, while at least one lower-tier industry tracker still describes Australia as 'Tier 1' as of late 2025/early 2026; the higher-tier account was treated as authoritative for scoring but the situation is genuinely unsettled and could be reclassified by a future US rule (Reuters, 5 Mar 2026, reports a possible new framework under consideration).
Canada — 45.4
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
A genuinely funded and executing institutional/financing base: Budget 2024's C$2bn Canadian Sovereign AI Compute Strategy, topped up by Budget 2025's C$925.6m, culminated in the April 2026 launch of the Sovereign Compute Infrastructure Program (SCIP) with binding Canadian-ownership/governance conditions; a real national HPC network is already operating (Fir at SFU, fully operational Sept 2025, ~20 petaflops; Trillium at U of T; tamIA at Laval), and Canada already has a genuine 2025-26 announcement-to-operation execution record (Fir, tamIA, the Cohere/CoreWeave Cambridge facility, TELUS's Rimouski Sovereign AI Factory).
Principal dependency
Virtually all commercial-scale AI compute capacity in Canada is foreign-owned or foreign-financed: Canada's flagship AI firm, Cohere, raised its August 2025 Series D (US$500m, $6.8bn valuation) from predominantly US strategic investors (AMD, Nvidia, Salesforce Ventures) and its own data center is built and operated by US company CoreWeave; hyperscaler capacity (AWS, Microsoft, Google) is entirely foreign-controlled equity; and no domestic or reserved allied leading-edge fab exists to substitute for NVIDIA-dependent chip supply.
Key catalyst
The April 2026 launch of the AI Sovereign Compute Infrastructure Program (SCIP), which for the first time attaches binding conditions -- Canadian-located, Canadian-governed, with data residency and Canadian ownership/contractual control of core compute -- to a national supercomputer buildout, alongside the May 2026 TELUS agreement targeting >60,000 GPUs by 2032.
Key constraint
Provincial grid rationing (Quebec's Bill 69 approval gate, Alberta's AESO connection limits, BC's 400MW cap, Ontario's IESO transmission review) is the binding near-term ceiling on further AI-infrastructure buildout across the country.
Constraint notes
Grid capacity is a documented, binding, near-term constraint in all four major AI-infrastructure provinces simultaneously: Hydro-Quebec has effectively halted large new data-center loads (Bill 69, passed June 2025, requires provincial approval for any load >=5MW); Alberta's system operator says it can reliably integrate only a small fraction of the ~20GW of data-center load that has applied for connection; BC capped new AI/data-center load at 400MW over two years (Jan 2026 allocation process); Ontario's IESO has flagged large data-center loads as requiring transmission upgrades. This rations buildout independent of available capital.
Notes
All 15 indicators returned locatable evidence (coverage_pct = 1.0). Confidence set to 'medium': only 6 of 16 total indicator fields are VERIFIED_FACT (37.5%, below the 70% bar for 'high'), and on Pillar 1 only power_grid_access is VERIFIED_FACT (1 of 3), so the Pillar-1 multi-source requirement for 'high' is not met either. Posture: compute_capacity scored 42 (just inside the Moderate band, 40-69) and control_ownership scored 38 (Low band, <40), mapping to 'Strategic Builder' per the numeric bands -- real domestic buildout underway (SCIP, TELUS, an operating national HPC network) without yet controlling the dominant share of national AI compute, most of which remains foreign-owned. Per the research pass's own transparency notes, several claims were explicitly checked and excluded from scoring as unverified: a distinct 'Canadian Center for AI Compute (CCAI)' government body was searched for and NOT FOUND (procurement capacity is instead attributed to ISED/NRC/the Digital Research Alliance/Canada Infrastructure Bank); a claimed 'federal AI Act' and 'AI Sovereignty Fund' (low-tier source augureai.ca) were excluded; a claimed TSMC Canada fab (mobilityforesights) was excluded; and a circulating 'C$2bn TELUS federal funding' figure (thecanadianwire) was excluded because the official ISED release explicitly states no funding has yet been committed or distributed for that project.
France — 59.2
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
A credible, financed, large-scale buildout pipeline anchored in a genuine energy advantage: nuclear provides roughly two-thirds of French generation with an EDF EUR70/MWh pricing ceiling from 2026, and Choose France 2026 (June 1, Versailles) produced a record EUR93bn in investment commitments including SoftBank's EUR75bn/5GW pledge (EUR45bn firm for phase 1) and Brookfield's EUR30bn cumulative commitment via Data4.
Principal dependency
The country's own frontier AI lab, Mistral AI, is majority foreign-owned by capital (ASML 11% as largest shareholder, heavy US venture capital and Gulf sovereign-wealth participation, no French investor above 10%), and the largest data-center commitments (SoftBank, Brookfield, MGX, Nebius) are foreign-sourced capital building on French soil rather than French capital building abroad or at home.
Key catalyst
The June 1, 2026 Choose France summit at the Château de Versailles, which produced EUR93bn in investment announcements (71 announcements, 15,600 jobs), anchored by SoftBank's EUR75bn/5GW data-center commitment tied explicitly to EDF nuclear sites and RTE grid capacity.
Key constraint
RTE grid interconnection capacity, not capital, is the binding constraint on turning announced gigawatts into operating compute; CRE's new fast-track procedure for state-designated sites only delivers 250MW in 24 months and 1,000MW in 4 years, far slower than the multi-gigawatt scale of the commitments made.
Constraint notes
Grid interconnection is the dominant documented constraint: RTE's connection queue held roughly 11GW of data-center capacity as of mid-2026 against a transmission network nearing exhaustion, with typical delays of 5-10 years outside the new fast-track procedure. France has also announced (~EUR93-109bn at Choose France 2026) more compute capacity than independent analysts believe the grid can energize on the investors' timelines. Ownership/control is a secondary constraint: the bulk of new AI-infrastructure capital (SoftBank/Japan, Brookfield/Canada, MGX/UAE) is foreign-sourced, and Mistral AI -- France's flagship sovereign lab -- has no French investor holding more than 10% of its capital after its Sept 2025 Series C (ASML/Netherlands is the largest single shareholder at 11%).
Notes
Coverage is 100% (16/16 indicators scored, none null) because France is an unusually well-documented case; even so, only 5 of 16 indicators (31%) are VERIFIED_FACT rather than ANALYTICAL_INFERENCE, reflecting that most of the country-level judgments (e.g., grading a still-under-construction data center, or assessing 'control' from a foreign-heavy cap table) require interpretation beyond a single primary-source statement, so confidence is set to medium rather than high. domestic_allied_fab_substitution is scored 0 (not null): STMicroelectronics/GlobalFoundries' Crolles 300mm fab (EUR7.5bn, EUR2.9bn French government support, targeted to ramp by 2026) is real and operating, but it produces FDSOI/bulk-CMOS chips for low-power and wireless applications -- there is no evidence of any domestic or allied-fab capacity for AI-accelerator-relevant advanced logic nodes, which is a materially different claim from 'no fab exists.' Posture is 'Strategic Builder': compute_capacity (67) is Moderate and control_ownership (44) is not High, matching the definition of building at scale without full control, rather than 'Sovereign Builder' (which would require both High) or 'Strategic Buyer' (which requires control achieved High specifically through negotiated foreign access, not the case here since France is host to substantial domestic construction, not just buyer of foreign-controlled capacity).
Germany — 67.1
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
A confirmed, operating national AI-compute base: the JUPITER exascale supercomputer at Juelich (inaugurated Sept 2025, EuroHPC-owned, GH200-based, ranked 4th on the Nov 2025 TOP500 list) is Europe's fastest supercomputer, and Deutsche Telekom's Munich Industrial AI Cloud (built with Nvidia in six months, ~10,000 Blackwell GPUs) is already operating at over one-third capacity.
Principal dependency
Germany's flagship sovereign AI lab, Aleph Alpha, is being absorbed into Canadian-headquartered Cohere in a $20bn deal (announced April 24, 2026) -- a case where a real domestic alternative existed but its ownership is shifting to a foreign parent even as German investors and the German government remain financially and operationally involved (PhariaAI continues to be deployed across German federal ministries and defense agencies).
Key catalyst
The Federal Government's first national Data Centre Strategy, adopted March 18, 2026, which sets a target to double overall data-center capacity and quadruple AI/HPC capacity by 2030 across 28 concrete measures on energy, siting and sovereignty.
Key constraint
Grid capacity is now the primary bottleneck: the Energy Efficiency Act's 100%-renewables-by-2027 requirement and documented near-doubling of data-center electricity demand over the past decade mean new AI capacity increasingly competes for scarce, compliant grid connections rather than capital.
Constraint notes
Power grid strain is the dominant documented constraint: German data-center electricity load nearly doubled from 14.1 to 25.9 TWh/year between 2015 and 2025 (projected to reach 48.3 TWh by 2035), and multiple 2026 reports describe grid capacity, not capital or technology, as the binding limit on AI-infrastructure growth. The Energy Efficiency Act also requires data centers to source 100% of electricity from renewables on a balance-sheet basis from Jan 1, 2027, adding a compliance constraint. A secondary constraint is the pending loss of full domestic control over Germany's flagship sovereign AI lab: Aleph Alpha is being acquired into Canadian-led Cohere at a $20bn valuation (announced April 2026), even though German investors (Schwarz Group, SAP, Bosch) and the German government (as anchor customer) remain involved.
Notes
Coverage is 100% (16/16 indicators scored, none null). 10 of 16 indicators (62.5%) are VERIFIED_FACT, below the 70% threshold for 'high' confidence, so confidence is set to medium. domestic_allied_fab_substitution is scored 25 rather than higher: the ESMC Dresden fab (TSMC/Bosch/Infineon/NXP joint venture, structural build complete late 2025, equipment move-in 2H2026, production targeted late 2027) is a genuine, financed, under-construction allied-fab facility, but its target process nodes (28/22nm and 16/12nm) are far from the leading-edge logic needed for AI accelerators, so it is scored as a credible-but-not-AI-relevant substitution capacity rather than at the higher operational tiers used for confirmed leading-edge fabs elsewhere. Posture is 'Strategic Builder': compute_capacity (75) is High but control_ownership (50) is Moderate, not High, so this does not meet the 'Sovereign Builder' bar (which requires both High); it is not 'Strategic Buyer' because Germany's control position is built on domestic construction and a real (if partially foreign-acquired) domestic lab, not negotiated foreign access alone.
Spain — 60.9
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
A genuinely public, operating sovereign-AI research asset: the Barcelona Supercomputing Center (BSC-CNS, a public consortium of the Spanish state, the Catalan government and UPC) developed and released ALIA, Europe's first public, open, multilingual LLM, trained on the MareNostrum 5 supercomputer across 35 European languages; Spain is also one of only three EU countries (with Germany and Poland) hosting two EuroHPC AI Factories (Barcelona and CESGA/Galicia).
Principal dependency
Spain's largest single AI-infrastructure capital commitment is entirely foreign: Amazon's EUR33.7bn Spain investment (announced March 2, 2026 at MWC Barcelona) dwarfs Spain's own EUR719m AI-gigafactory bid, and Spain's flagship semiconductor ambition under PERTE Chip has not yet produced a confirmed advanced-logic fab.
Key catalyst
Amazon's March 2, 2026 announcement of an additional EUR18bn (total EUR33.7bn) in Spanish data-center and AI infrastructure investment, expanding the AWS Europe (Spain) Region in Aragon that has been operational since November 2022.
Key constraint
Spain's own flagship sovereignty projects -- the AI gigafactory bid (Tarragona/Madrid, up to EUR4bn) and PERTE Chip's advanced-fab ambition -- remain unawarded/unmaterialized, leaving the country's compute buildout disproportionately dependent on foreign hyperscaler capital rather than domestically-directed investment.
Constraint notes
Grid and transmission capacity is a documented constraint despite Spain's renewable-energy abundance: while renewables covered 55.5-56.6% of national electricity in 2025, multiple 2026 reports note that many announced data-center projects depend on transmission or generation upgrades not yet in place. A second, structural constraint is semiconductor-fab execution risk: Spain's PERTE Chip program (EUR12.25bn budget through 2027) has an explicitly unmaterialized flagship ambition -- Broadcom's planned fab was cancelled, and no major advanced-logic fab has been secured as of this research.
Notes
Coverage is 100% (16/16 indicators scored, none null). 9 of 16 indicators (56%) are VERIFIED_FACT, below the 70% threshold for 'high' confidence, so confidence is set to medium. domestic_public_ownership_share is scored on BSC-CNS (a genuinely public consortium operating MareNostrum and ALIA) rather than on commercial telecom/energy holdings, since no sourced evidence of a Spanish state equity stake in a commercial AI-infrastructure operator was located in this pass. domestic_allied_fab_substitution is scored PROVISIONAL/25 because PERTE Chip is financed at the program level (EUR12.25bn) but its central objective -- attracting a major advanced-node fab -- is explicitly documented as unmaterialized (Broadcom's planned fab was cancelled). Posture is 'Strategic Builder': both compute_capacity (67) and control_ownership (44) are Moderate, which fits building at scale without full control rather than 'Sovereign Builder' (needs both High) or 'Import-Dependent' (needs both Low).
Italy — 61.7
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
Italy's flagship national AI-gigafactory bid is anchored by companies with substantial, verified Italian state ownership: Eni (Ministry of Economy and Finance plus CDP hold a combined 33.09% of share capital) and Leonardo (Ministry of Economy and Finance holds ~30.2%) are leading the bid together with TIM -- described as the only coordinated single-country gigafactory proposal submitted to Brussels. Italy was also the first EU member state to pass a national AI law complementing the EU AI Act (Law No. 132/2025, implementing decrees approved June 10, 2026).
Principal dependency
Near-term data-center capital in Italy is dominated by foreign hyperscalers (Microsoft EUR4.3bn, AWS EUR1.2bn over five years, Equinix EUR4bn for seven Milan-area data centers through 2033), and Italy has no domestic AI-accelerator chip-design or leading-edge fabrication capability -- its main semiconductor investment (STMicroelectronics' EUR5bn Catania silicon-carbide campus) targets power electronics, not AI-relevant logic.
Key catalyst
The IT4LIA EuroHPC AI Factory at Bologna's Tecnopolo has been operational since September 2025 (EUR420m co-financed equally by the EU and Italy), and Italy's national gigafactory bid led by Eni, Leonardo and TIM is described by Industry Minister Adolfo Urso as the only coordinated single-country proposal submitted to Brussels' EU-wide gigafactory tender.
Key constraint
Terna's data-center connection queue (50-79 GW of requests against a national generating base of roughly 120 GW) is the binding constraint on converting Italy's substantial announced investment pipeline into operating capacity.
Constraint notes
Grid connection capacity is the dominant documented constraint and is proportionally severe: by January 2026, data-center developers had submitted connection requests to Terna for 78.79 GW -- more than half of Italy's entire generating capacity -- and Terna's own count showed requests above 50 GW across 300+ projects by mid-2025, up from 30 GW six months earlier. Terna's EUR16.6bn grid-investment plan (2024-2028) is a real but partial mitigation, and no Italy-specific fast-track delivery mechanism comparable to France's CRE procedure was located.
Notes
Coverage is 100% (16/16 indicators scored, none null). 10 of 16 indicators (62.5%) are VERIFIED_FACT, below the 70% threshold for 'high' confidence, so confidence is set to medium. domestic_allied_fab_substitution is scored 0 (not null): STMicroelectronics' EUR5bn Catania silicon-carbide campus (~EUR2bn Italian government support via the EU Chips Act, production targeted 2026, full capacity by 2033) is real, financed and under construction, but silicon carbide is a power-electronics technology, not AI-accelerator-relevant logic -- there is no evidence of any domestic or allied-fab capacity in Italy targeting AI-relevant advanced nodes, a positive claim distinct from 'no evidence found.' Posture is 'Strategic Builder': compute_capacity (58) is Moderate and control_ownership (63) is Moderate (not High), fitting the 'building at scale without full control' definition rather than 'Sovereign Builder' (needs both High) or 'Strategic Buyer' (needs control High via negotiated foreign access, which is not the case given Italy's own domestically-anchored, state-owned-company-led gigafactory bid).
Russia — 35.0
Posture: Import-Dependent · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
Strong domestic ownership and control of existing AI infrastructure -- Sberbank (operator of the Christofari supercomputers) is majority state-owned (50%+1 share held via the Ministry of Finance since 2020) and Yandex (operator of Russia's largest AI clusters) became 100% Russian-owned on 2024-07-15 when Dutch parent Yandex NV completed a full divestment -- combined with mature, actively-deployed domestic foundation models: GigaChat (Sber) and YandexGPT (Yandex) are both trained from scratch and serve production workloads across banking, public services and enterprise.
Principal dependency
Total hardware dependency on gray-market NVIDIA chips: every flagship operational cluster (Yandex's Chervonenkis/Galushkin/Lyapunov, Sber's Christofari systems) runs on pre-2022 NVIDIA hardware, and there is no citable evidence of any large GPU cluster commissioned in Russia since the full 2022 embargo -- new hardware arrives almost exclusively via transshipment through Hong Kong, mainland China, and Turkey/Central Asian intermediaries. No domestic or allied fab is within reach of AI-accelerator-class nodes (Mikron is capped at 90nm; the government's own target is 28nm by 2030, a node TSMC achieved in 2011).
Key catalyst
The Center for Artificial Intelligence Development (CAID), created under Government Resolution No. 861 (9 June 2025) as a federal coordination office and operational since August 2025 -- though President Putin himself stated in November 2025 that its current ~50-person apparatus is insufficient and demanded a 'real operational headquarters.'
Key constraint
Russia's classification in the most restrictive US export-control tier (Country Group D:5, presumption of denial) combined with near-exhausted Moscow-region grid capacity are the two binding, simultaneous constraints on any further scale-up.
Constraint notes
Multiple severe, well-documented constraints operate simultaneously: Russia sits in the most restrictive US export-control classification (Country Group D:5 / AI Diffusion Rule Tier 3, presumption of denial for advanced computing ICs and AI model weights); every incremental hardware gain depends on gray-market/transshipment smuggling (primarily via Hong Kong and China) rather than legitimate supply; Moscow-region grid capacity is reported as close to exhausted (~50MW free capacity remaining, new connections described as 'virtually impossible' for new investors); and Moscow Times/Forbes reporting found 38 Russian data-center projects worth ~168.6bn rubles (~$2.26bn) were suspended over the past three years due to financing rates and power constraints.
Notes
All 15 indicators returned locatable evidence (coverage_pct = 1.0). Confidence set to 'medium': 9 of 16 total indicator fields are VERIFIED_FACT (56.25%, below the 70% bar for 'high'), and on Pillar 1 only power_grid_access is VERIFIED_FACT (1 of 3), so the Pillar-1 multi-source requirement is not met either. Posture is flagged as a genuine judgment call worth human double-checking: compute_capacity scored 33 (Low band, <40) but control_ownership scored 69 (Moderate band, 40-69, just under the High threshold of 70) -- an unusual combination the four-way posture taxonomy was not designed for, since Russia's comparatively strong domestic/state ownership reflects nationalization under sanctions pressure (the Kremlin-brokered forced divestment of Yandex NV; long-standing state ownership of Sber) rather than sovereign capacity-building, while the country's ability to grow its compute stock at all depends entirely on illicit imports of foreign (smuggled) hardware. This file classifies Russia as 'Import-Dependent' primarily on that basis -- Low compute plus total dependence on foreign chip imports for any future expansion -- rather than 'Strategic Builder' (which requires compute to be Moderate-or-High) or 'Strategic Buyer' (which requires control to be High via negotiated foreign access, not applicable since Russia's control stems from domestic nationalization, not a foreign-access negotiation). Two items were explicitly downweighted in scoring: Sber's reported ~500 billion-ruble third supercomputer (Vedomosti, 10 Feb 2026) is single-sourced and was publicly disputed by Sber itself as 'extremely inaccurate' (19 Feb 2026), so committed_buildout was scored conservatively; and the Moscow Times/Forbes finding of 38 suspended data-center projects (~168.6bn rubles) was treated as material evidence that committed buildout is partly illusory.
Mexico — 35.4
Posture: Import-Dependent · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
A genuine, dated hyperscaler execution track record -- AWS, Microsoft, Google and Oracle have all moved from announcement to live cloud regions in Mexico on documented timelines (Oracle operating two regions since 2022-23; AWS live Jan 2025; Google live Dec 2024) -- alongside a credible, if not yet built, national supercomputer program: Coatlicue (MX$6bn, 314-petaflop design, technical committee installed June 2026, sited at IPN Zacatenco).
Principal dependency
No AI-training-scale compute cluster currently operates in Mexico at all: MEXDC's own executive director stated in June 2026 that 'there is no data center in Mexico' capable of supporting AI-scale (>250MW) operations, and every major buildout underway (AWS >$5bn, Microsoft $1.3bn, CloudHQ $4.8bn, Odata ~$1.2bn) is 100% foreign-owned, with no Mexican equity, board seats, or governance rights secured in exchange -- confirmed explicitly in this research, including Nvidia's own denial of any direct financial investment in Mexico (13 Nov 2025).
Key catalyst
The June 2026 installation of the Coatlicue supercomputer's technical committee -- Mexico's first genuinely public national AI-compute asset -- with construction targeted for the second half of 2026 and domestic operation targeted around 2028.
Key constraint
Transmission/grid interconnection (CFE/CENACE) is the acknowledged binding bottleneck on further AI-infrastructure buildout, independent of the substantial foreign capital already committed.
Constraint notes
Grid/transmission capacity (CFE/CENACE) is a documented, binding constraint: the Mexican Data Center Association (MEXDC) states it has 'unsuccessfully' sought to coordinate electricity planning with the Energy Ministry (SENER), estimates ~1.7GW of additional capacity is needed to meet announced data-center targets, and connection wait times of up to four years have been reported in Queretaro. Microsoft is running natural-gas generators at one site pending grid connection expected in 2027.
Notes
14 of 15 indicators returned locatable, scorable evidence; hyperscaler_bargaining_leverage returned no evidence at all (explicitly searched for and not found -- Mexico's hyperscaler deals were announced at ceremonial presidential press conferences with no disclosed terms) and was set to null rather than 0, since this is an absence-of-evidence case rather than a confirmed negative claim. Because control_ownership retained evidence on 3 of its 4 indicators (75%, above the 60% pillar floor), all five pillars remain 'scored' and coverage_pct = 1.0. Four indicators were scored 0 rather than null because each rests on an explicit, well-corroborated confirmed-absence finding rather than a mere lack of search results: operational_advanced_compute (MEXDC's own executive quoted stating no AI-scale facility exists); domestic_public_ownership_share (Mexico's one flagship domestically-founded operator, KIO Networks, was acquired by foreign PE fund I Squared Capital in 2021, and no other public/domestic AI-infrastructure ownership was found); foreign_capital_exposure (every major buildout is explicitly confirmed 100% foreign-owned with no Mexican capital participation found); and domestic_allied_fab_substitution ('No operational wafer fabrication exists in Mexico' -- only design/packaging/assembly capacity). Posture: compute_capacity scored 33 (Low band, <40) and control_ownership scored 8 (Low band, <40), which maps directly to 'Import-Dependent' per the posture rules -- both pillars are Low, driven by confirmed absence of both operating domestic compute and any domestic/public ownership stake. Domestic model claims (KAL, Mexico's first national-scale LLM) were scored only 25/nominal: an independent review (Mexico Business News, Dec 2025) found KAL was presented 'with no technical documentation, no code, no benchmarks.'
Ireland — 55.0
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
Ireland is an established, continuously-operating European hyperscaler hub: AWS has run its Ireland region since Nov 2007 (now 3 Availability Zones, 33 sites nationally per operator tracking), alongside Microsoft (11 facilities), Google, and Meta, with 72 operational grid-connected data-centre buildings confirmed in a June 2026 government-commissioned KPMG report. This is reinforced by a fresh, large, dated capital commitment: Intel's €5bn ($5.7bn) Fab 34 (Leixlip) expansion, announced 13 Jul 2026.
Principal dependency
Compute ownership and the AI-accelerator supply chain are almost entirely foreign-controlled. No Irish frontier AI lab or sovereign model was found -- the only Irish-branded model, UCCIX, is a small academic fine-tune of Meta's open-weight Llama-2-13B for the Irish language, not an infrastructure alternative -- and Ireland's only advanced-node semiconductor fab (Intel Fab 34) produces Xeon server CPUs, not GPUs/AI accelerators. For the GPUs that actually run frontier AI workloads, Ireland depends entirely on imports via US hyperscalers.
Key catalyst
Three concrete, dated unlocks followed years of a de facto Dublin grid moratorium: the CRU's 12 Dec 2025 Large Energy User Connection Policy reopened data-centre grid connections nationally (conditional on operators providing their own on-site/proximate generation); the 13 Oct 2025 EuroHPC CASPIr hosting agreement (€25M, University of Galway/ICHEC); and the 2026 launch of Rinn AI, a €460M state-funded (plus ~€500M industry co-funded) national AI research network spanning all 13 Irish universities.
Key constraint
Dublin-specific grid capacity remains the binding constraint on further hyperscaler build-out: EirGrid continues case-by-case assessment in the region even after the national CRU policy reopening, and data centres already account for roughly half of Dublin's electricity demand.
Constraint notes
Grid capacity is the dominant, best-evidenced constraint: data centres draw roughly 48-50% of Dublin's electricity demand, and EirGrid continued case-by-case review of new Dublin connections even after the CRU's 12 Dec 2025 national policy reopening (one 2025 report described an effective halt on new Dublin applications until 2028), pushing new hyperscaler growth toward Cork and Limerick. A secondary constraint is near-total foreign ownership of compute infrastructure and of the only advanced-node fab (Intel), with no domestic AI-accelerator fabrication capacity at all.
Notes
All 5 pillars cleared the 60% per-pillar indicator-coverage floor (16 of 16 indicators received a numeric score; none were null in this research pass), so coverage_pct = 1.0 and overall_score is a straight weighted sum (0.30×50 + 0.20×50 + 0.20×50 + 0.15×67 + 0.15×67 = 55.1, rounded to 55) with no pillar redistribution needed. Two items a human editor should double-check: (1) domestic_alternatives was scored 0 rather than null -- this rests on a 25 Sep 2025 Irish Times feature explicitly framing 'sovereign AI' as an aspiration Ireland has not yet acted on (contrasted with France's ASML-backed Mistral) plus confirmation that UCCIX (the only Irish-branded model found) is a small academic Llama-2 fine-tune, not an infrastructure alternative; this is a defensible but genuinely close call between 0 and null, since absence-of-evidence and evidence-of-absence are easy to conflate here. (2) posture is 'Strategic Builder' on the literal numeric bands (compute_capacity = 50 Moderate, control_ownership = 50 Moderate-not-High), which is the correct rule application, but Ireland's actual model -- an FDI-attraction strategy (IDA Ireland) that imports foreign hyperscaler capital to build physical infrastructure at scale on Irish soil rather than negotiating access to infrastructure built elsewhere -- has some family resemblance to 'Strategic Buyer' too; Strategic Buyer was not used because that classification specifically requires control_ownership to reach the High (>=70) band via negotiated foreign access, which it does not here (foreign_capital_exposure and domestic_alternatives pull control_ownership down to 50). domestic_allied_fab_substitution (score 50) rests entirely on Intel Fab 34, which is real, operational, and on an advanced node (Intel 3, further $5.7bn expansion announced 13 Jul 2026) but fabricates Xeon server CPUs, not GPUs/AI accelerators -- this is flagged explicitly in the indicator claim so it is not mistaken for AI-accelerator fab capacity. Evidence for operational_advanced_compute was thinner than the country's data-centre reputation might suggest: Ireland's first dedicated liquid-cooled AI supercomputer (the Boole upgrade in Cork, Nvidia Blackwell HGX) had been delivered but had not gone live as of the most recent (11 Aug 2026) trade-press report, so the score rests on general-purpose operational hyperscaler cloud regions rather than a confirmed, named, live dedicated AI-training cluster. One aggregate figure (€18.5bn cumulative data-centre investment across 107 facilities, from a single lower-tier business-news aggregator) was deliberately NOT used as scored evidence on its own -- it appears only as unscored context in the committed_capital claim, per the rule that a single low-tier source cannot support a score above 25 or VERIFIED_FACT status.
Switzerland — 66.6
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
A fully operational, wholly state-owned, top-10-globally-ranked AI supercomputer -- CSCS's 'Alps' (10,752 NVIDIA GH200 Grace-Hopper sockets, 434.9 petaflops sustained, #10 on the June 2026 TOP500 list) -- that has already been used to train and publicly release Switzerland's own open-source sovereign LLM, Apertus (released September 2, 2025).
Principal dependency
Near-total reliance on a small number of US suppliers across the stack: NVIDIA GPUs and HPE systems integration for the flagship Alps supercomputer, and AWS/Microsoft Azure/Oracle Cloud Infrastructure for commercial and government cloud, with Microsoft directly investing $400 million (announced June 2, 2025) to expand its own -- not Swiss-owned -- data centers near Zurich and Geneva.
Key catalyst
FlexBase Group's 500 MW 'Technology Center Laufenburg' AI data-center and battery-storage campus, under active construction since its May 5, 2025 groundbreaking (cornerstone laid June 2026) with commissioning targeted for summer 2028, sited at the historic 'Star of Laufenburg' international grid-interconnection hub.
Key constraint
Switzerland has no advanced-node semiconductor fab and no near-term AI-accelerator-relevant fabrication substitution: the SwissChips program targets niche low-power ASICs and sensors, not AI accelerators, and the proposed 'Chip FabLab' remains an unfinanced, drawing-board-stage proposal seeking state support as of August 2025.
Constraint notes
Genuine, well-documented structural constraint: Switzerland was excluded from the Tier-1 'close ally' list under the (now-rescinded) January 2025 US AI Diffusion Rule and would have faced Tier-2 caps on AI chip/cluster imports; the rule was rescinded before taking effect but a replacement framework remains under development, so meaningful forward exposure to US export-control policy persists. Compounding this, Switzerland's compute stack -- both its flagship sovereign supercomputer (Alps: NVIDIA GPUs, HPE systems integration) and its commercial/government cloud base (AWS, Microsoft Azure, Oracle Cloud Infrastructure) -- depends almost entirely on a small number of US suppliers, with no domestic or non-US-controlled hyperscale alternative identified. Power/grid access, by contrast, is not currently a binding constraint per the evidence reviewed.
Notes
All 16 indicators across the 5 pillars returned locatable, citable evidence in this research pass, so coverage_pct = 1.0 (5 of 5 pillars scored; every pillar cleared the 60% non-null indicator-coverage floor with 100% coverage, so no pillar redistribution was needed). JUDGMENT CALL flagged for editor review: posture is classified 'Strategic Builder' because compute_capacity scored High (83 -- driven by the fully operational, state-owned Alps supercomputer and the actively-under-construction Laufenburg campus) while control_ownership scored only Moderate (56) -- pulled down by significant foreign (chiefly US) ownership and operation of the broader commercial data-center/cloud base (Microsoft's $400M investment; AWS/Azure/OCI dominance in government cloud) despite CSCS's fully public ownership of the flagship sovereign compute asset. This is a genuine mixed picture rather than a clean-cut case, and an editor may reasonably weigh the public-ownership-of-the-flagship-asset fact more heavily toward 'Sovereign Builder.' EXCLUDED CLAIM: a secondary-source claim (found only in a single lower-tier aggregator, ecorpit.com) that BIS's prospective 'Redesigned Framework for Artificial Intelligence Diffusion' was submitted to OIRA in February 2026 and withdrawn in March 2026 could not be corroborated against a primary BIS or Federal Register source in this research pass and was excluded; export_control_exposure is instead scored on the well-corroborated facts of the original rule's May 13, 2025 rescission and Switzerland's documented non-Tier-1 ('close ally') status under the original January 2025 rule. VERIFIED-AND-NOT-FOUND: the research brief's suggested lead that CSCS/Alps might be relocating to or expanding at a 'Piora' site could not be corroborated by any source located in this pass (only the existing Lugano site and the EPFL-Lausanne extension of the Alps infrastructure were confirmed); it was therefore not used or scored anywhere in this file.
Norway — 61.3
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
Real, financed, under-construction hyperscale build-out at multiple sites, anchored by cheap hydropower: Nscale/Aker's Narvik campus (230 MW initial phase, 100,000 targeted NVIDIA GPUs, >$10bn estimated total cost per Time's reporting) backed by a Microsoft compute contract reported at $6.2bn (Sept 2025) and expanded to over 30,000 NVIDIA Rubin GPUs (Apr 2026); TikTok/Green Mountain's Hamar campus (90 MW, 3 buildings) is already fully complete and operational.
Principal dependency
The country's largest AI infrastructure project is built and controlled by a foreign (UK-headquartered) operator, Nscale, whose investor base is majority non-Norwegian (Nvidia, Dell, Lenovo, Citadel, Point72, 8090 Industries) even after Aker ASA's stake rose to ~27% of share capital; NVIDIA GPUs are the near-universal compute substrate across every named Norwegian project (Olivia, Narvik, Tydal), and Norway has no domestic semiconductor fabrication capacity of any kind.
Key catalyst
Aker ASA's escalating stake in Nscale (from 9.3% to ~23.9% fully diluted / ~27.3% of share capital, via a ~$350m Series C cash investment plus roll-up of their prior 50% Narvik joint-venture stake, completed 2026) paired with Microsoft's expanded multibillion-dollar compute offtake at Narvik -- together the clearest evidence of committed capital converting into physical, under-construction infrastructure rather than remaining an announcement.
Key constraint
Grid-connection queue pressure (Statnett: 3.4 GW already reserved for data centers as of Jan 2026, out of a national installed base where Statnett warns not all queued projects will be realized) and a dedicated national AI Act that has already slipped its original 2026 target to spring 2027.
Constraint notes
Three documented, material constraints: (1) grid connection is a real bottleneck despite abundant hydropower -- Statnett had already reserved 3.4 GW (>8% of national installed capacity) for 53 registered data-center projects as of Jan 2026, and Statnett itself cautions not all queued projects will be realized; (2) Norway's flagship AI compute build-out is majority foreign-owned and single-vendor dependent -- the Narvik campus operator Nscale is UK-domiciled with a non-Norwegian majority investor base (Nvidia, Dell, Lenovo, Citadel, Point72, 8090 Industries) even after Aker ASA became its largest shareholder at ~27% of share capital, and every named Norwegian AI compute project depends on Nvidia GPUs delivered through the same handful of foreign infrastructure partners; (3) Norway's dedicated AI Act has slipped past its original 'late summer 2026' in-force target to a bill before the Storting only in spring 2027.
Notes
Coverage is 100% (all 5 pillars cleared the 60%-of-indicators-non-null threshold, so coverage_pct = 1.0 and score_status = 'scored'). Within compute_access_resilience, 1 of 3 indicators (domestic_allied_fab_substitution) is null because Norway has no semiconductor fab of any kind (fabless design model only; NorFab is R&D micro/nanofabrication infrastructure, not production capacity) and no evidence of any allied-fab substitution arrangement was located -- the pillar score of 50 is the average of the 2 remaining non-null indicators only, consistent with the methodology used in MDA.json. No UNVERIFIED claims were excluded from this file; all indicators either carry a scored, sourced claim or are null. Posture judgment call: compute_capacity (67, Moderate band) paired with control_ownership (44, Moderate band, not High) maps to 'Strategic Builder' under the stated posture rules (compute Moderate-or-High + control not-High = building at scale without full domestic control) -- this is a reasonably mechanical application of the bands, not a borderline call, but a human editor should double-check the control_ownership sub-scores (foreign_capital_exposure=25 and hyperscaler_bargaining_leverage=50 in particular) since they required the most synthesis/interpretation of any indicators in this file. Also flagging for editor review: the OpenAI angle from the original research brief did not pan out as expected -- OpenAI co-announced 'Stargate Norway' with Nscale/Aker in July 2025, but per DCD/Capacity reporting (Apr 2026) OpenAI did not reach a final agreement with Nscale for the Narvik site and Microsoft took over the capacity instead; this file reflects the corrected, current (Aug 2026) ownership/offtake picture, not the original 2025 announcement.
Turkey — 38.8
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
TÜBİTAK ULAKBİM's ARF-ACC national supercomputer is a real, actively-upgraded, independently verified operational asset: it climbed from rank 266 to 145 on the TOP500 global list between June and November 2025 after the addition of 192 Nvidia H200 GPUs (TOP500.org, Nov 18, 2025), sitting on top of a TRUBA e-infrastructure that reports 80,000 processor cores and 312 GPUs serving thousands of researchers.
Principal dependency
AI-relevant compute and cloud infrastructure is concentrated on a small number of foreign suppliers: Turkey's only confirmed operational GPU cluster runs exclusively on Nvidia hardware, its only committed full hyperscale cloud region is a single Google Cloud partnership with Turkcell (not live until 2028-2029), and AWS's only Turkey presence is an edge Local Zone in Istanbul (GA May 20, 2026), not a full region.
Key catalyst
President Erdoğan's June 13, 2026 unveiling of the 2026-2030 National AI Action Plan, targeting $10 billion in largely private mobilized investment and 1 GW of data-center capacity by 2030, layered on the pre-existing $30B HIT-30 High Technology Investment Program (launched July 26, 2024) and its ~$5B HIT-Chip semiconductor line.
Key constraint
Turkey's prior 2021-2025 National AI Strategy expired without the government presenting an account of whether its flagship targets (top-20 global AI ranking, 5% AI share of GDP) were met; Oxford Insights' 2025 Government AI Readiness Index ranked Turkey just 53rd of 195 countries, explicitly noting that national strategy commitments made years earlier remained unfulfilled -- a documented gap between announced ambition and delivered outcomes at the flagship-strategy level, even where individual infrastructure projects (AWS Local Zone, Equinix IL4, ARF-ACC) did execute.
Constraint notes
Three documented, distinct constraints: (1) export-control exposure -- Turkey was named a Tier 2 country under the US BIS Jan 2025 AI Diffusion Rule (aggregate ~50,000 AI-chip-unit cap 2025-27, no unrestricted access to top-tier parts), and while BIS says it has stopped enforcing that specific rule pending a replacement (a Feb 2026 draft was withdrawn in March 2026), no bespoke bilateral access agreement of the kind the UAE or Saudi Arabia negotiated has been found for Turkey, leaving its current status ambiguous rather than resolved; (2) grid interconnection -- national installed capacity is large (115 GW) but data-center-specific interconnection is strained (2024 utilization at 71% of available power) and a $10B transmission upgrade won't complete until the early 2030s, after the government's 2030 buildout target; (3) supplier concentration -- Turkey's only confirmed operational AI-relevant GPU cluster runs entirely on Nvidia hardware, and its only committed hyperscale cloud region (Google) is not yet live, with no Microsoft Azure region confirmed and AWS present only via a single Istanbul edge Local Zone.
Notes
All 5 pillars cleared the 60% within-pillar indicator-coverage floor (every pillar in this file has 100% non-null indicators), so coverage_pct = 1.0 and all 5 pillar weights (0.30/0.20/0.20/0.15/0.15) are used directly with no redistribution. overall_score = 0.30(42) + 0.20(44) + 0.20(25) + 0.15(50) + 0.15(33) = 38.85, rounded to 39. score_status is 'scored' because weighted coverage is 100% (>=80% threshold). POSTURE JUDGMENT CALL: compute_capacity (42) and control_ownership (44) both fall in the Moderate band (40-69), with control_ownership not reaching the High (>=70) band required for 'Strategic Buyer' or 'Sovereign Builder'. Per the methodology's explicit rule ('Strategic Builder: compute Moderate-or-High paired with control not-High'), this is classified Strategic Builder -- Turkey is building real capacity (TRUBA/ARF-ACC operational, a financed hyperscale Google Cloud region, a $30B HIT-30 program) without yet controlling that capacity outright (a foreign hyperscaler supplies the cloud stack, Nvidia is the sole confirmed GPU vendor, and a disclosed foreign investor (LetterOne, 19.8%) holds a large stake in the state-influenced entity building the domestic side of the cloud region). EXCLUDED UNVERIFIED CLAIM: a claim that Fly BVLOS Technology would begin producing advanced domestic chips in Ankara with a path to 3nm by 2027 (following a March 2025 technology-transfer agreement with Taiwan's Accton Technology) was excluded from the domestic_allied_fab_substitution score. It is corroborated only by low-tier Turkish defense/trade outlets (Defensehere, TURDEF, DonanımHaber), no government or tier 1-4 source was found, and the claimed timeline is technically implausible for a company with no prior fab track record partnering with a networking-hardware maker rather than a fab/IP licensor. The indicator was instead scored on the better-corroborated, government-linked HIT-Chip/YONCA program (Türkiye Today, Jul 28, 2024), which is real but early-stage (16nm initial target, long-run sub-7nm goal, no confirmed operational fab). SINGLE-SOURCE CAPS: power_grid_access and site_readiness_signal are each sourced to a single non-corroborated aggregator/newsletter (a Medium blog and a niche trade newsletter respectively); per methodology rule 2 both are capped at 25 and barred from VERIFIED_FACT status regardless of how plausible the underlying figures are. EXPORT CONTROL STATUS: Turkey's Tier 2 classification stems from a Jan 2025 BIS rule that BIS itself says it has stopped enforcing, with no finalized replacement as of Aug 2026 -- this is flagged as an analytical judgment call a human editor should double-check as the replacement rule firms up.
Israel — 46.7
Posture: Strategic Builder · Direction: baseline · Coverage: 100% · Last verified: 2026-08-10
Pillar scores
Strongest capability
A live national AI supercomputer (Nebius-operated, 1,000 Nvidia B200 GPUs of a 4,000-GPU program, allocating since Jan 2026) plus Nvidia's own Israel-1 facility (operational since Nov 2023), combined with the strongest documented hyperscaler bargaining leverage found in this research round: leaked Project Nimbus contract terms (reported by +972 Magazine/Guardian, Oct 2025, based on Finance Ministry documents) reportedly bar Google/AWS from restricting Israel's use of services even where it breaches their own terms of service, and require covert notification if a foreign court orders data turnover, backed by liquidated-damages clauses.
Principal dependency
Nearly all flagship AI infrastructure operating in Israel is foreign-owned or foreign-operated -- the national supercomputer (Nebius, Netherlands-HQ), Nvidia's Israel-1, and the AWS/Google/Microsoft/Oracle cloud regions -- and Israel's leading domestic foundation-model lab, AI21 Labs, is heavily foreign-invested (Google, Nvidia, Intel Capital among its backers) and was reportedly the subject of Nvidia and then Nebius acquisition interest in 2025-26, calling its independent Israeli-controlled status into question.
Key catalyst
The government's June 2026 National AI Program approval, following the Prime Minister's Office National AI Directorate (established Oct 2025) issuing an RFI on 2026-04-08 to 'advance the establishment of Israeli sovereignty in foundational models.'
Key constraint
The Electricity Authority's July 2026 140-day freeze on new data-center grid connections is the binding near-term constraint on further AI-infrastructure buildout, with officials reportedly unable to determine which of the ~27,000 MW in pending requests are realistic versus speculative.
Constraint notes
A severe, currently binding constraint: on 2026-07-20 Israel's Electricity Authority instructed grid operator Noga to freeze new data-center grid-connection applications for 140 days after connection requests reached roughly 27,000 MW -- about three times Israel's average national consumption (~9,000 MW). Only ~1,500 MW of connections are already committed. Compounding this, the US AI Diffusion Rule's Tier-2 classification for Israel lapsed with the rule's May 2025 rescission, leaving chip-access terms undefined and policy-dependent, and Intel's leading-edge Fab 38 expansion at Kiryat Gat remains frozen with a portion of its government grant withdrawn (2026-07-31).
Notes
All 15 indicators returned locatable evidence (coverage_pct = 1.0). Confidence set to 'medium': 7 of 16 total indicator fields are VERIFIED_FACT (43.75%, below the 70% bar for 'high'), and while Pillar 1 has 2 of 3 VERIFIED_FACT (committed_buildout, power_grid_access), operational_advanced_compute itself is ANALYTICAL_INFERENCE, so the 'high' bar's Pillar-1 requirement is not cleanly met either. Posture judgment call flagged for human review: compute_capacity scored 50 (Moderate band, 40-69) and control_ownership scored 38 (Low band, <40), which numerically maps to 'Strategic Builder' -- but this masks a real tension worth double-checking. Three of the four control_ownership indicators (domestic_public_ownership_share, foreign_capital_exposure, domestic_alternatives) scored low (25 each) because Israel's posture is described in its own National AI Program materials as procuring compute access rather than owning assets, and its critical infrastructure is overwhelmingly foreign-owned. But the fourth indicator, hyperscaler_bargaining_leverage, scored unusually high (75) on the strength of the leaked Project Nimbus contract terms -- real, binding, and specific in a way Australia's non-binding 'Expectations' framework is not. A human reviewer may reasonably judge this closer to 'Strategic Buyer' (control achieved through negotiated foreign-access terms rather than domestic ownership) than 'Strategic Builder'; this file uses the numeric-band default because the other three control indicators are weak enough to hold the pillar average in the Low band, but the underlying evidence supports either read. Per opencode's own search-log transparency section, three claims were explicitly excluded as UNVERIFIED and are not reflected in any score: a reported '$30 billion' national AI sovereignty commitment (low-tier aggregator only), a '100,000 GPU' sovereign-compute target (same low-tier source only; gov.il's own release does not state this figure), and a '$963.99B Israel data center market by 2031' marketing projection (Mordor Intelligence, referenced only secondhand via DCD).
Methodology summary
Pillars and weights (v1.0): Domestic Compute Capacity 30%, Control & Ownership 20%, Compute Access Resilience 20%, Institutional & Data Sovereignty 15%, Financing & Execution Capacity 15%. These weights are Juncture's analytical judgment, not a universal scientific standard — a sensitivity table using alternate weights is included in the downloadable dataset.
Scoring: every indicator is scored on a bounded 0/25/50/75/100 rubric with documented thresholds (announced → financed → under construction → operational, and equivalents for regulation and ownership). No evidence means the indicator is left unscored (NA), not defaulted to zero.
Display rule: a numeric overall score is shown only where at least 80% of weighted pillar coverage is supported by evidence. Between 50–80% coverage, a country is marked Provisional. Below 50%, no numeric score is shown at all — only whatever pillar-level evidence exists.
Evidence hierarchy: government/regulator documents and company filings first, then multilateral institutions and credible research institutions, then high-quality journalism. Low-quality aggregators are never used where a higher-tier source exists, and are never sufficient alone to mark a claim VERIFIED FACT.
Full methodology, including indicator definitions, evidence-status classification, confidence and posture derivation, missing-data rules, and versioning policy: download SOVEREIGN_AI_INDEX_METHODOLOGY.
Weighting sensitivity
A second scenario reduces Domestic Compute Capacity to 25% and raises Institutional & Data Sovereignty and Financing & Execution Capacity to 20% each, to show how sensitive the ranking is to Juncture's specific weighting choice (methodology §12). This is a supplementary check, not a second headline ranking.
| Country | Primary score (v1.0 weights) | Alternative weights | Delta |
|---|---|---|---|
| United States | 81.2 | 81.2 | 0.0 |
| United Kingdom | 53.7 | 54.2 | +0.5 |
| Japan | 67.9 | 69.6 | +1.7 |
| India | 60.0 | 60.0 | 0.0 |
| Saudi Arabia | 70.0 | 71.2 | +1.2 |
| United Arab Emirates | 70.2 | 70.8 | +0.6 |
| Brazil | 45.3 | 46.4 | +1.1 |
| Nigeria | 43.8 | 47.2 | +3.4 |
| Vietnam | 59.4 | 62.5 | +3.1 |
| Kenya | 46.9 | 49.3 | +2.4 |
| China | 75.4 | 76.7 | +1.3 |
| South Korea | 65.4 | 67.9 | +2.5 |
| Singapore | 59.1 | 60.4 | +1.3 |
| Indonesia | 55.4 | 55.8 | +0.4 |
| Taiwan | 75.0 | 77.1 | +2.1 |
| Australia | 51.2 | 51.2 | 0.0 |
| Canada | 45.4 | 46.3 | +0.9 |
| France | 59.2 | 60.0 | +0.8 |
| Germany | 67.1 | 67.9 | +0.8 |
| Spain | 60.9 | 61.7 | +0.8 |
| Italy | 61.7 | 63.3 | +1.6 |
| Russia | 35.0 | 35.8 | +0.8 |
| Mexico | 35.4 | 36.2 | +0.8 |
| Ireland | 55.0 | 55.8 | +0.8 |
| Switzerland | 66.6 | 66.6 | 0.0 |
| Norway | 61.3 | 60.9 | -0.4 |
| Turkey | 38.8 | 39.2 | +0.4 |
| Israel | 46.7 | 46.7 | 0.0 |
Sources & credibility
464 scored/attempted indicators across all countries, of which 426 carry a cited, dated source and the remainder are honestly left null (no evidence located) rather than guessed. Evidence-status breakdown of every cited indicator: 224 VERIFIED_FACT (53%), 163 ANALYTICAL_INFERENCE (38%), 37 PROVISIONAL (9%). No indicator is ever published at UNVERIFIED status -- claims that could not be corroborated are dropped from scoring and, where material, disclosed instead in the country's notes field.
Most-cited sources
By domain, across all indicator-level citations in the dataset. This is a citation-frequency count, not an endorsement ranking -- see the full source list for every citation with its exact claim and date.
| Source domain | Citations |
|---|---|
| datacenterdynamics.com | 22 |
| cnbc.com | 14 |
| bis.gov | 12 |
| cbi.nstc.gov.tw | 7 |
| gov.uk | 6 |
| news.microsoft.com | 6 |
| pib.gov.in | 5 |
| en.vietnamplus.vn | 5 |
| reuters.com | 5 |
| aboutamazon.eu | 5 |
| tomshardware.com | 4 |
| techcrunch.com | 4 |
Full source list: download the complete indicator-level citation table (CSV) -- every claim in this dataset, cited or explicitly marked as no-evidence-found, with its source, URL, and date.
Limitations
This index measures infrastructure and control, not AI capability, safety posture, research output, or talent. A country can score low here while leading on AI research, or vice versa.
Coverage is uneven by design. Countries with more transparent disclosure and more English-language press coverage will systematically have higher evidence coverage — not necessarily higher actual sovereignty. A low score_status means “we could not verify enough to score,” not “this country has low sovereignty.”
Facts move fast. Export-control regimes, financing rounds, and construction status can change within weeks. Every indicator carries a last-verified date for exactly this reason.
Ownership percentages are frequently undisclosed, particularly for Gulf sovereign-wealth participation in US AI labs. Where exact cap-table figures are not public, Control & Ownership is scored on qualitative governance evidence (board seats, security agreements, voting-control filings), not invented percentages.
This is Juncture's original analytical judgment, not a peer-reviewed academic instrument. Pillar selection, weights, and the posture taxonomy are disclosed as such throughout.
Related Juncture research
Data, citation & versioning
Downloadable dataset: country-level CSV export · full indicator-level source list (CSV) · methodology document
Suggested citation:
https://juncturepolicy.org/sovereign-ai-index/
Changelog: v1.1.0 (2026-08-11) — expanded country coverage from 12 to 29. v1.0.0 (2026-08-08) — initial publication, corrected the prior page's unsourced “11 DOE federal AI sites” claim (see banner above and the United States country card). Full changelog: see data/CHANGELOG.md in the methodology download.
License: CC BY 4.0. Working draft — Juncture Policy publishes quarterly re-verification passes; see methodology for cadence and correction policy.