Frontier AI Dependency Explorer

How frontier AI models are governed today—and where countries, institutions, firms, and infrastructure providers hold power over their development and deployment.

0visible nodes

How frontier AI is governed—and who holds power

A simplified system view. The model is at the center. Infrastructure and country capabilities enable it from below; institutions, laws, security bodies, and evaluators constrain it from above. Power comes from controlling something the model developer cannot easily replace.

Start here · click any card to inspect evidence
4Governance mechanisms
6Critical dependency layers
19Middle-power countries
8Frontier / adjacent labs

What governs frontier models today?

A layered patchwork rather than a single regime: company commitments and third-party evaluations; national executive and legislative authority; market regulation such as the EU AI Act; export controls and national-security procurement; and diplomatic or standards processes that shape expectations but rarely bind labs directly.

Where does country power come from?

Control over scarce inputs, market access, energy and compute, capital, military procurement, legal jurisdiction, verification capacity, or diplomatic coalition-building. The strongest leverage is often narrow—and becomes power only when a government or institution is willing and able to use it.

Core insight: frontier labs remain concentrated in the United States and China, but the system they depend on is not. The governance opportunity lies in connecting distributed chokepoints and political authority into coordinated conditions on access, deployment, and safety.
Show Region
Strategic lens
Middle powerUS/China poleResource / chokepointFrontier labInstitution / forumAgreement / policy / law supplyownership / hostingrule / membershipdependency / constraintverification / evaluation

Scores are directional, not predictive. “Leverage” combines breadth across eight AI value-chain axes; “dependency” captures exposure to foreign chips, firms, capital, standards or markets. Click a node to isolate its network; click empty space to clear.

Geographic distribution of leverage

A true geographic projection of the countries in the dataset. Circle area reflects the selected power metric; color reflects coalition role. Curved lines show documented bilateral agreements, strategic dependencies, and supply relationships.

Robinson projection · country positions are geographic
Size nodes byLines
AnchorBridgeEngageCounterpartSystem pole

Actor and dependency network

A topology-first view. Countries sit in the center; resources, labs, institutions, and agreements form surrounding rings. Use the filters to isolate the actor system and click any node to inspect its evidence.

Position shows relationship structure, not geography
DensityNode sizeEmphasis

PageRank: influence through influential connections

A higher score means the actor is connected to other actors that are themselves central. It highlights system-level influence, not simply the number of direct links.

CountryResourceLabInstitutionAgreement / policy

AI value chain and control points

Read this view vertically. Countries sit in the middle, aligned to the stages where they add value. The resources and capabilities they contribute run left-to-right underneath. Governance, security and verification bodies sit above and connect to the specific stages they regulate, procure, secure or test.

Click any card for evidence and dependencies
Layers
Country / poleFrontier labInstitutionAgreement / policycritical dependency

References and source notes

The map synthesizes official policy documents, company disclosures, institutional sources and specialist research. The register below shows the evidence behind every line in the network—not only the general bibliography.

How to read the evidence. A direct connection is explicitly documented (for example, OpenAI’s Stargate UAE project). A membership connection comes from an official participant list. A regulatory connection reflects the legal scope of a rule. An analytical inference combines documented upstream positions into a dependency claim—for example, that leading-edge fabs depend on Dutch–German EUV equipment. Inferences are marked rather than presented as contractual relationships. Lines are visually undirected; the register states the direction and meaning.

Connection evidence register

ConnectionMeaningBasisSources

Source library

Semiconductors and critical inputs

  1. ASML — EUV lithographyEquipment chokepoint and advanced-node manufacturing.
  2. TSMC — investor and capacity reportingLeading-edge fabrication and advanced packaging.
  3. SK hynix — HBM technologyHigh-bandwidth memory supply and roadmap.
  4. IEA — Global Critical Minerals Outlook 2025Mining, refining concentration and supply risks.
  5. USGS — critical minerals informationCountry production and mineral-resource context.

Compute, energy and frontier labs

  1. OpenAI — Stargate announcementUS compute buildout and infrastructure partnerships.
  2. Anthropic — company announcementsModel, cloud and capital relationships.
  3. Google DeepMind — official updatesFrontier-model activity and UK–US institutional exposure.
  4. Mistral AI — company announcementsEuropean frontier-adjacent model capacity.
  5. IEA — Energy and AIPower demand, grid constraints and data-centre geography.

Legislation and regulatory leverage

  1. European Commission — EU AI ActScope, implementation timeline and extraterritorial market effects.
  2. US BIS — advanced semiconductor export controlsChip, equipment and destination constraints.
  3. Republic of Korea MSIT — AI policyKorea’s AI Basic Act and implementation materials.
  4. Brazil MCTI — national AI strategyBrazilian governance and institutional context.
  5. India MeitY — AI and digital policyIndiaAI, digital public infrastructure and governance initiatives.

International forums and standards

  1. International Network of AI Safety InstitutesCross-national evaluation and measurement cooperation.
  2. OECD — GPAI partnershipMembership and institutional integration.
  3. United Nations — Global Digital CompactUniversal political framework for digital and AI cooperation.
  4. G7 Hiroshima AI ProcessInternational guiding principles and code of conduct.
  5. Frontier AI Safety Commitments — Seoul AI Summit 2024Voluntary company commitments to publish frontier-safety frameworks, identify severe-risk thresholds and define mitigations. Sixteen original signatories; four organisations were added by February 2025.
  6. OECD AI PrinciplesBaseline intergovernmental principles used across policy forums.

Power-holding and verification institutions

  1. United States CongressFederal legislation, appropriations, investigations, hearings and statutory authority affecting AI developers and executive agencies.
  2. White House — America’s AI Action PlanExecutive direction across infrastructure, procurement, diplomacy, security and agency implementation.
  3. US Bureau of Industry and SecurityExport-control design, licensing and enforcement for advanced chips, equipment and emerging technologies.
  4. NIST Center for AI Standards and Innovation (CAISI)US-government testing, standards and collaborative research interface with commercial AI developers.
  5. European AI OfficeCentral EU implementation and enforcement capacity for general-purpose AI under the AI Act.
  6. UK AI Security InstituteState-backed model evaluations, measurement science and risk-mitigation research.
  7. METRIndependent third-party evaluations of autonomous and AI-R&D capabilities, including pre-deployment work with frontier developers.

Independent research added in this revision

  1. OECD — Mapping the semiconductor value chain (2025)Cross-country dependencies, bottlenecks and supply-chain segmentation used to test the original report.
  2. Epoch AI — AI-chip supply-chain constraints (2026)Evidence that HBM and advanced packaging, not only leading-edge logic, became binding constraints.
  3. NATO — revised AI strategyAlliance-level military adoption, interoperability, responsible-use principles and protection from adversarial AI.
  4. NATO — emerging and disruptive technologiesData and AI Review Board and institutional machinery for military AI assurance.
  5. UN CCW GGE on lethal autonomous weapons (2025)The principal multilateral negotiating track on autonomous weapons and human control.
  6. US Treasury — Outbound Investment Security ProgramCapital-control leverage over semiconductors, quantum technology and specified AI systems in countries of concern.
  7. US Treasury — CFIUS authorityForeign-investment screening as a national-security lever over ownership, technology transfer and critical infrastructure.
  8. ISO/IEC JTC 1/SC 42 — Artificial intelligenceInternational technical standards that can become procurement requirements, conformity benchmarks and regulatory infrastructure.
  9. US Department of Defense — GenAI.milDefense procurement and deployment can confer market, validation and security leverage on model and cloud providers.
  10. US BIS — rescission of AI Diffusion Rule (2025)Corrects the attached report’s treatment: the January 2025 diffusion framework was rescinded before compliance, while chip-related controls continued.

Cross-country evidence

  1. Stanford HAI — AI Index Report 2025Talent, models, investment and national capabilities.
  2. Oxford Insights — Government AI Readiness IndexState capacity and governance comparisons.
  3. Tortoise — Global AI IndexNational implementation, talent and infrastructure indicators.
  4. Economist Intelligence Unit — Democracy Index 2024Regime classifications used in the underlying assessment.

Underlying analytical resource

  1. The Map of Leverage: AI value-chain leverage assessmentPrepared for Ettore Arpini, June 2026. Country scores use eight axes: minerals; equipment/design/IP; fabrication/packaging/memory; compute/data centres/energy; capital; models/talent; market/data/DPI; and governance/convening.
  2. Interpretation ruleThe index rewards breadth, while the network highlights narrow chokepoints and asymmetric dependencies that a simple score can obscure.

Bilateral and minilateral AI agreements

Data vintage: principally 2024–July 2026. Independent review: the report was treated as a hypothesis set and checked against OECD supply-chain mapping, official regulatory and alliance documents, institutional sources and recent bottleneck research. The map now distinguishes direct evidence, formal membership, regulatory reach and analytical inference. Correction: the January 2025 US AI Diffusion Rule was rescinded before compliance; the map therefore represents current BIS chip controls and licensing power rather than that superseded tier system. Announced investments, planned gigawatts, ownership stakes, export-control rules and legislative status can change quickly. The map is designed as a strategic orientation tool rather than a definitive supply-chain model.
A tool developed by Ettore Arpini · give feedbackFrontier AI Dependency Explorer