THE SCENARIO
A national regulatory authority convenes a cross-ministry working group to design an AI governance framework for domestic deployment. The working group reads the Bletchley Declaration, the EU AI Act, NIST AI RMF, and the UAE AI Governance Principles. Each document is understood to be comprehensive. Each document is selected for its perceived authority. The working group proceeds to map domestic requirements — law enforcement facial recognition, banking credit assessment, civil registry biometrics — against the framework provisions. Twelve months later, the framework is adopted. Two years later, an auditor notes that every technical standard cited in the framework was developed, maintained, and validated by foreign entities. The authority to define, update, or suspend those standards belongs to foreign bodies. The working group never identified this as a sovereignty issue because sovereignty was never in the diagnostic frame.
Why does every AI governance framework transfer sovereignty to foreign entities by design?
The current global architecture of AI governance treats sovereignty as an implementation detail. Sovereignty is assumed to flow from competent technical adoption. If a nation adopts the standards, builds the compliant infrastructure, and trains its regulators in the prescribed methodology, sovereignty is presumed satisfied. This assumption is the central failure of contemporary AI governance design. The mechanism of sovereignty erosion is built into the structure of the frameworks themselves.
“The TEE Method — Test, Evaluate, Exit — is not a checklist. It is an architecture for sovereignty. It does not tell you what to do. It gives you a structure for deciding for yourself.”
Part One: Understanding the Landscape
The foundational error in AI governance is the reframing of sovereignty as a governance challenge rather than a structural governance condition. When sovereignty is treated as a governance challenge, the solution space is populated by governance tools: policy documents, regulatory sandboxes, compliance procedures, audit frameworks, and ministerial committees. These tools operate within the governance domain. They do not reach the structural domain where sovereignty actually resides. Sovereignty is not a matter of what the regulator permits. It is a matter of who controls the decision environment in which the regulator operates.
The international governance architecture functions as a sovereignty extraction mechanism. A multilateral body or standards consortium identifies an AI risk category. It convenes technical working groups dominated by vendors from jurisdictions with established AI industries. The working group drafts a standard or framework. The standard carries the implicit authority of international consensus. A national regulator in a developing jurisdiction adopts the standard through legislation, executive order, or regulatory sandbox admission criteria. At the moment of adoption, the governance authority over that domain of domestic AI activity transfers to the foreign body that owns the standard revision process. The transfer is complete. The national regulator has exercised its authority to adopt a standard. It has not exercised authority over the standard.
This transfer mechanism operates across all current major frameworks. The Bletchley Declaration established a global consensus on AI safety priorities without creating any domestic governance capacity. The EU AI Act translates technical compliance requirements into risk tiers that domestic regulators must implement without owning the technical definitions underneath. NIST AI RMF provides a voluntary framework that becomes mandatory when adopted by reference in regulation. The UAE AI Governance Principles, while domestically originated, embed compliance with international technical standards that are maintained outside national control. The PDPL enforces data protection requirements that align with GDPR logic, ensuring that data sovereignty enforcement pathways converge on frameworks developed by foreign legislators.
The sovereignty erosion is deliberate in structure but accidental in intent. The entities designing these frameworks are not seeking to undermine national sovereignty. They are solving technical problems with technical tools. The sovereignty consequence is a side effect of their conceptual frame, not the object of their design. That distinction explains why the erosion has been so little noticed. The entities involved are performing what they understand to be competent governance work. No one is pointing out that the competent work is producing an outcome that contradicts the purpose of governance itself.
The specific provisions that drive sovereignty erosion follow a recognizable pattern. Efficiency provisions require adoption of internationally recognized standards to avoid duplication of effort and to enable market access. Modernisation provisions mandate integration with global AI ecosystems to maintain technological parity. Compliance provisions create certification pathways that tie domestic approval to foreign qualification bodies. Standardisation provisions establish the international standard as the domestic benchmark by default. Investor protection provisions require regulatory predictability that is best achieved through alignment with established frameworks in major financial centres. Innovation promotion provisions prioritise ecosystem growth over control retention. Market access provisions trade governance autonomy for cross-border commercial opportunity.
Each provision is individually reasonable. Each provision serves a legitimate policy objective. Together, they constitute a comprehensive sovereignty transfer architecture. The nation receives compliant systems, certified vendors, predictable regulation, and investor confidence. What the nation does not receive is authority over the systems it deploys.
The critical case study is illustrative. A nation in East Africa adopted the EU AI Act risk tier structure for its emerging AI regulatory framework in 2025 as a mechanism to attract European technology investment. The legislation was drafted by local legal counsel, translated into domestic statute, and passed through parliamentary procedure. The process appeared to be a sovereign exercise of legislative authority. Three years post-adoption, the auditor discovered that the technical definitions underpinning the risk tiers — the specific model capability assessments, the robustness benchmarks, the data quality criteria — were maintained by ISO/IEC subcommittees in which the nation held no membership and no voting representation. The EU AI Act referred to these standards by reference. The domestic legislation referred to the EU AI Act. The result was that the nation’s domestic AI governance was effectively administered by a foreign technical body with no accountability to the nation’s legislature, its public, or its regulatory authorities. The sovereignty transfer had occurred through a chain of technical references.
The Seven Layer Stack Audit
The sovereignty erosion mechanism unfolds across seven layers. Each layer represents a governance allocation point. Each layer is independently addressable. Together, they determine the entity’s genuine sovereignty position.
| Layer | Governance Authority | Sovereignty Score (1–5) | Critical Dependency |
|---|---|---|---|
| Layer 1: Hardware & Compute | Foreign vendor-controlled (NVIDIA H100/H200 firmware, AMD MI300X firmware); ASML lithography equipment subject to export control | 1 | NVIDIA CUDA firmware ecosystem; PTX ISA updates; hardware-level telemetry and management interfaces; NVSwitch interconnect patent control |
| Layer 2: Foundational Models & Training Data | Foreign model weights (OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet, Meta Llama 3); training data sourced from global corpora no domestic curation | 2 | OpenAI API model endpoints; Anthropic API routing; Llama licensing through Meta; absence of domestic foundational models (Toyota/KAUST, MBZUAI research insufficiently commercialised) |
| Layer 3: API & Middleware | Foreign API providers (OpenAI API, Azure OpenAI Service, Google Vertex AI, AWS Bedrock) | 2 | API rate limits, content policies, request logging retention, model update blackout periods, API endpoint physical location and jurisdiction |
| Layer 4: Platform & Application | Foreign platform vendors (Azure Government Cloud, Google Workspace AI, Microsoft 365 Copilot) | 2 | Application governance policy configurable only within vendor constraints; cross-application data flows governed by vendor infrastructure not domestic law |
| Layer 5: Data Architecture & Sovereignty | Mixed; PDPL enacted (Saudi Arabia); NCA technical standards enforced; but data residency requirements circumvented through hybrid cloud architecture using AWS Bahrain or Azure UAE regions subject to US CLOUD Act extraterritoriality | 2 | AWS Bahrain Region data jurisdiction; Azure UAE Region compliance architecture; Cloudflare CDN and WAF routing; PDPL enforcement gaps in cross-border data flows |
| Layer 6: Governance & Regulatory Framework | Domestic legislation adopted (PDPL, SAMA AI Governance Principles, NCA cybersecurity regulations); but compliance targets reference international standards (ISO/IEC 42001, NIST AI RMF, EU AI Act conformity assessments) | 3 | ISO/IEC 42001 certification bodies; NIST AI RMF adoption by reference in SAMA guidance; EU AI Act class conformity assessment procedures conducted by European notified bodies |
| Aggregate Sovereignty Score | Sovereignty held at Layer 6 partially; Layers 1-5 predominantly foreign-controlled with limited domestic control points | 2.0 / 5 | Layer 1: Hardware & Compute is the critical failure layer |
The aggregate sovereignty score of 2.0 reflects a structural condition, not a policy failure. The layers at which the entity holds genuine governance authority are the most remote from the operational reality of AI deployment. The entity controls the regulatory framework that governs the deployment. The entity does not control the hardware on which the deployment runs, the models that execute the intelligence, the APIs through which the models are accessed, the platforms that package the capability, or the data architecture that determines custody.
The critical failure layer is Layer 1: Hardware and Compute. The reason is not simply that hardware is imported. The reason is that hardware sovereignty failure has cascading structural consequences. GPU firmware defines the instruction set architecture of computation. Updates to firmware are delivered by the vendor without requiring legislative or regulatory approval. The vendor firmware update can change the computational behaviour of every device in the national infrastructure simultaneously. The sovereign entity has no veto over this update. It has no mechanism to audit the update for compliance with domestic policy requirements. It has no authority to refuse the update if the update conflicts with domestic technical standards. The firmware update is delivered as an operational necessity. It is treated as a technical maintenance event. It is, in substance, a governance transfer event that occurs without the sovereign entity’s participation.
NVIDIA’s dominance in the AI compute market makes this dependency acute. The H100 and H200 GPUs, which are the foundation of most sovereign AI compute initiatives in the region, are designed, fabricated, and controlled by NVIDIA. The CUDA software ecosystem is owned by NVIDIA. The NVLink and NVSwitch interconnect technologies are patented by NVIDIA. A government agency or sovereign wealth fund that purchases a cluster of NVIDIA GPUs has acquired compute capacity. It has not acquired compute sovereignty. The governance of that compute resides in Santa Clara, California, not in the nation’s capital.
The same pattern repeats across the MENA region. G42, based in Abu Dhabi, has deployed significant AI compute infrastructure. Its infrastructure depends on NVIDIA H100 clusters. Core42, G42’s parentage and technology arm, manages sovereign cloud infrastructure for UAE federal and emirate-level entities. The compute layer, however, remains beneath the sovereignty line. Khazna, a provider of sovereign cloud and managed services in Saudi Arabia, offers data residency commitments. Its compute layer still depends on imported hardware stacks. Injazat, serving UAE government clients, offers sovereign governance packaging over foreign compute stacks. The pattern is consistent and structurally determined.
At Layer 2, foundational models, the sovereignty picture is equally constrained. The UAE has invested in domestic model development through MBZUAI research outputs and the Jais Arabic-language model developed in partnership with Inception, a G42 company. These are genuine domestic capabilities. The scope of a foundational model developed by a 100-billion parameter research effort against the scale of global foundation models maintained by OpenAI, Anthropic, Google, and Meta is asymmetrical. A domestic model will always have narrower capabilities, shorter training runs, and smaller context windows. The sovereignty question is not whether a domestic model exists. The sovereignty question is whether the domestic model is operationally sufficient to govern the domain. The answer, for the foreseeable future, is no. The nation remains dependent on foreign foundational models while developing its own.
Part Two: The Sovereignty Test Matrix
The sovereignty erosion mechanism is not a single event. It is a structural condition that the specific provisions create across the entity’s operation. The Test Matrix maps this condition across the domain, identifying for each layer where sovereignty is genuinely held, where it is claimed but not exercised, and where it has been transferred without the entity’s explicit recognition.
The Test Matrix is a diagnostic instrument. It requires a binary answer to each question. The test is not whether the entity could, in theory, exercise sovereignty. The test is whether the entity actually exercises sovereignty in practice. The gap between theoretical authority and exercised authority is the sovereignty erosion zone.
Can the entity suspend operation of all AI systems within its jurisdiction by a single regulatory order without vendor cooperation? Can the entity audit the training data of every model deployed operationally within its jurisdiction without vendor agreement? Can the entity require that all data generated by domestic AI operations remain within domestic infrastructure without vendor exceptions? Can the entity define the safety criteria that govern AI deployment without reference to international standards? Can the entity update its regulatory requirements for AI models unilaterally without creating compliance conflicts that vendors cannot meet? Can the entity, in a dispute with a vendor about model behaviour, enforce a domestic determination without international arbitration? Can the entity protect all AI-generated data from extraterritorial access by foreign governments under existing vendor contractual terms? The answers to these questions determine whether sovereignty exists or is claimed.
Part Three: Red Flag Checklist
If three or more of the following apply, AI governance sovereignty is not a future risk — it is an existing condition.
- The national AI governance framework adopts ISO/IEC 42001 as a mandatory or de facto mandatory compliance target without domestic normative content beyond the standard
- Regulatory sandbox admission requires vendor compliance with NIST AI RMF without domestic audit capability
- The foundational models in operational use are entirely sourced from OpenAI, Anthropic, Google, or Meta with no domestic alternative
- AI compute infrastructure is procured through foreign vendors with firmware update cycles beyond domestic regulatory visibility
- The data protection authority enforces PDPL or equivalent through cross-border data transfer mechanisms that route through AWS Bahrain or Azure UAE under US CLOUD Act jurisdiction
- The cybersecurity authority (NCA) certifies AI systems for government use through vendor-supplied security attestations without independent technical verification
- AI governance legislation was drafted with reference to EU AI Act conformity assessment procedures conducted by European notified bodies
- The central bank (SAMA or CBUAE) governs AI deployment in financial services using vendor-authored risk frameworks adopted through regulatory guidance without legislative sovereignty review
Part Four: Phased Implementation Framework
| Phase | Timeframe | Key Actions |
|---|---|---|
| Assessment | Weeks 1-4 | Map every dependency across the seven layer stack for all operational AI systems; complete the Sovereignty Test Matrix for each system category; identify sovereignty traps using the red flag checklist; document every international standard referenced in domestic legislation and regulation |
| Strategic Planning | Months 2-3 | Develop withdrawal protocols for the three highest risk dependencies at Layers 1 and 2; identify and pilot domestic or allied alternative foundational models; begin firmware audit and localisation programs; negotiate contractual protections with vendors including audit rights, update notification, and domestic jurisdiction governing law; establish the Sovereignty Caucus within the legislature |
| Sovereign Transition | Months 4-6 | Execute phase one exit protocols for Layer 1 dependencies with acceptable domestic or allied alternatives identified; deploy domestic foundational models for government operations; establish domestic technical standards where international standards have been adopted by reference; complete legislative review of all international standard adoptions with explicit sovereignty reservation clauses |
| Institutionalisation | Ongoing | Quarterly sovereignty audits against the Seven Layer Stack with published public scorecards; staff training programs on sovereignty architecture in AI procurement; national sovereignty register cataloguing all foreign AI dependencies with exit timelines; annual sovereignty scorecard with legislative and executive accountability mechanisms; strategic adjustment every six months based on vendor and standards body developments |
The framework’s implementation must be adapted to the specific institutional context of the adopting entity. In Saudi Arabia, SAMA’s existing AI governance framework for financial services provides a foundation for Layer 6 activity. The Saudi Arabian Monetary Authority has issued AI governance principles that reference international standards. Phased implementation requires translating these principles into sovereignty-anchored domestic standards that do not depend on foreign qualification bodies for enforcement.
In the UAE, PDPL enforcement by the Data Office and NCA technical standards enforcement provide Layer 5 and Layer 6 anchor points. The UAE’s domestic AI development investment through G42, Core42, and MBZUAI creates the conditions for Layer 2 domestic alternative development. Khazna and Injazat provide the domestic service delivery infrastructure for Layers 3 and 4. The challenge is connecting these domestic capabilities into a sovereignty-anchored stack rather than allowing them to serve as local instances of foreign governance.
Dual calendar governance transitions — coordinating Hijri and Gregorian fiscal and regulatory cycles — create implementation windows that can be aligned with phased deployment. The integration of UAE Pass and Saudi Absher identity infrastructure provides a domestic identity governance layer at Layer 4 that is among the most advanced in the region. These capabilities must be extended into AI governance rather than remaining as siloed identity systems.
The assessments conducted by MISA in Saudi Arabia and ADIO in Abu Dhabi provide the policy instruments for the Assessment phase. Aramco and ADNOC operational AI deployments represent the highest-stakes Layer 1-2 dependencies in the region and should be prioritised in the Strategic Planning phase. DEWA’s AI-driven grid management represents another critical dependency category where Layer 1 compute control has direct citizen impact.
Aramco and ADNOC operational AI deployments represent the highest-stakes Layer 1-2 dependencies in the region and should be prioritised in the Strategic Planning phase. DEWA’s AI-driven grid management represents another critical dependency category where Layer 1 compute control has direct citizen impact. The DHA’s AI-assisted diagnostics programs create health sector dependencies at Layers 4 and 5 that require same-priority treatment. The NESA evaluation framework for autonomous systems in the UAE provides a regulatory instrument that can be redirected from compliance verification to sovereignty audit.
The SFDA’s pharmaceutical and food product regulation creates a vertical where Layer 1-5 AI dependencies have direct public health implications. If a foundational model approved for drug interaction assessment is controlled by a foreign vendor whose firmware update conflicts with domestic calibration requirements, the regulatory outcome is a public health dispute between a domestic regulator and a foreign platform holder. The resolution of that dispute will not be determined by SFDA authority. It will be determined by the foreign vendor’s support cycle and contractual terms. The sovereignty failure in AI governance is not abstract when it enters the pharmacy.
The ZATCA e-invoicing and tax compliance deployment creates a Layer 4-5 dependency where AI-assisted transaction review meets regulatory compliance. The system’s sovereignty is contested because the e-invoicing platform operates across international service delivery boundaries. Hijri/Gregorian dual calendar alignment in fiscal systems introduces additional sovereignty complexity. A tax compliance system whose fiscal calendar logic is embedded in a foreign platform cannot be adjusted unilaterally without vendor cooperation. The domestic fiscal year determination is governed by foreign software updates. This is the sovereignty erosion mechanism operating at the infrastructure layer alongside the governance layer.
The Closing Question
The opening question asked why AI governance frameworks treat sovereignty as a technical parameter. The answer is structural: the frameworks are designed by entities for which sovereignty is not the governing concept. Their governing concept is technical interoperability and global market access. Sovereignty, when it appears in their documents, appears as an implementation consideration — something to be addressed through domestic adaptation of the framework, not something that constrains the framework’s design. The result is a global governance architecture that systematically transfers sovereignty from adopting nations to the entities that design and maintain the standards.
The nation that adopted a foreign governance framework to attract investment discovered, too late, that the framework was architecturally incompatible with sovereign AI aspirations. The incompatibility is not a bug. It is a feature of frameworks designed for global deployment rather than national self-determination. Nation states that seek genuine AI sovereignty must recognise that adoption of off-the-shelf governance frameworks is itself a sovereignty decision — and that the decision, by default, is to transfer.
The binary is acceptance or agency.
This article draws on the TEE Method™ framework from SOVEREIGN: Who Owns the Future? by Toni Shatagoe.