Global Model Governance Frameworks for AI Competition Law

1. Introduction

Model governance refers to the institutional, technical, contractual and legal mechanisms through which an AI model is developed, trained, tested, deployed, monitored, updated and retired. In competition law, model governance has become important because control over a foundation model can translate into control over data, compute, APIs, distribution channels, downstream applications, standards and users.

The central competition-law question is therefore no longer simply:

Who has market power in an AI product?

It increasingly becomes:

Who governs the model, who controls access to the model ecosystem, and whether that governance structure can be used to exclude competitors or entrench market power?

The EU, UK and US competition authorities have expressly recognised this emerging issue. In 2024, the European Commission, UK CMA, US DOJ and FTC issued a joint statement identifying competition risks in generative AI and foundation-model markets. The UK CMA has separately developed principles concerning fair competition, choice, access, transparency and accountability in foundation-model markets.

A global model-governance framework therefore needs to connect AI regulation with merger control, abuse-of-dominance rules, monopolisation law, interoperability, data access, cloud dependence and algorithmic conduct.

2. Meaning of a Global AI Model Governance Framework

A global model-governance framework can be understood as a system containing six interconnected layers:

1. Model-development governance

Rules concerning:

• training data;
• model architecture;
• compute resources;
• safety testing;
• evaluation;
• documentation;
• model weights;
• fine-tuning;
• deployment.

2. Access governance

Rules determining who can access:

• APIs;
• model weights;
• inference services;
• training infrastructure;
• developer tools;
• technical documentation;
• datasets;
• computing capacity.

3. Distribution governance

This concerns whether a powerful AI provider controls:

• app stores;
• operating systems;
• search engines;
• browsers;
• cloud platforms;
• enterprise software;
• device defaults;
• advertising systems.

4. Ecosystem governance

A dominant AI firm may simultaneously operate:

cloud → foundation model → API → operating system → application → marketplace → advertising/data ecosystem.

Competition law must therefore examine vertical and conglomerate relationships rather than looking at the model in isolation.

5. Algorithmic governance

AI systems may independently:

• determine prices;
• rank competitors;
• allocate resources;
• recommend suppliers;
• discriminate between customers;
• exchange information;
• optimise bidding;
• coordinate commercial strategies.

The 2024 EU–UK–US competition statement specifically warned that algorithms can facilitate information exchange, price fixing, collusion and exclusionary conduct.

6. Regulatory governance

Competition authorities increasingly need:

• model audits;
• access obligations;
• interoperability;
• monitoring;
• technical investigations;
• merger review;
• behavioural remedies;
• structural remedies;
• cross-border coordination.

3. Why Model Governance Is a Competition-Law Issue

AI markets have unusual economic characteristics.

A. High fixed costs

Training frontier models requires substantial:

• computing power;
• specialised chips;
• engineering talent;
• data;
• electricity;
• capital.

These costs can create substantial entry barriers.

B. Economies of scale

A successful model can distribute inference costs across millions of users.

C. Data feedback loops

More users can produce more data and feedback, which can improve the system, attracting still more users.

This can create:

users → data → better model → more users → more data.

D. Compute dependence

A model developer may depend upon a small number of cloud or accelerator suppliers.

E. Distribution dependence

A model may be technically competitive but commercially weak if it cannot obtain access to:

• smartphones;
• browsers;
• search interfaces;
• cloud marketplaces;
• enterprise systems.

F. Switching costs

Enterprise customers may become dependent upon:

• proprietary APIs;
• embeddings;
• fine-tuning;
• prompts;
• workflow integrations;
• proprietary agents;
• model-specific infrastructure.

Consequently, model governance can become a mechanism of market foreclosure.

4. Principal Global Competition Risks

A. Model-access foreclosure

A dominant model provider may deny or restrict access to competitors.

Potential theories include:

• refusal to supply;
• essential-facility-type arguments;
• discriminatory access;
• technical degradation;
• discriminatory API terms.

B. Self-preferencing

A vertically integrated AI company might give its own model preferential access to:

• operating-system functions;
• search data;
• device capabilities;
• cloud infrastructure;
• distribution channels.

This resembles traditional self-preferencing problems in digital-platform competition.

C. Cloud–AI vertical integration

A cloud provider may simultaneously control:

compute + model + API + enterprise distribution.

This creates the possibility of:

• tying;
• bundling;
• foreclosure;
• discriminatory cloud pricing;
• preferential access to compute;
• exclusion of rival models.

The EU's current DMA work illustrates the increasing importance of cloud and AI infrastructure: the Commission has identified AI and cloud as critical competition priorities, and in 2026 began considering designation of AWS and Azure as gatekeepers.

D. Data foreclosure

A dominant platform may possess data unavailable to competitors.

Competition concerns may arise when:

• data is necessary for model improvement;
• access is technically restricted;
• portability is ineffective;
• competitors cannot replicate the data feedback loop.

E. Model interoperability restrictions

A platform may prevent rival AI systems from accessing:

• operating-system functionality;
• device sensors;
• user data;
• search data;
• application interfaces.

The EU's 2026 DMA measures concerning Android AI interoperability demonstrate how interoperability is becoming an explicit regulatory instrument.

5. Six Core Principles of Global AI Competition Governance

A useful global framework should incorporate the following principles.

Principle 1 — Contestability

Markets should remain open to new models and applications.

Principle 2 — Non-discrimination

Dominant infrastructure providers should not discriminate arbitrarily between their own AI systems and competitors.

Principle 3 — Interoperability

Competitors should be able to interact with important infrastructure where legally and technically justified.

Principle 4 — Data access and portability

Competition should not be permanently blocked by exclusive control over strategically important datasets.

Principle 5 — Transparency

Authorities must be able to understand:

• model inputs;
• commercial relationships;
• API restrictions;
• technical dependencies;
• pricing structures;
• governance decisions.

Principle 6 — Accountability

AI governance should not become a mechanism for avoiding responsibility by claiming that an algorithm independently made the decision.

6. Global Regulatory Architecture

European Union

The EU combines:

• Article 101 TFEU;
• Article 102 TFEU;
• EU Merger Regulation;
• Digital Markets Act;
• Data Act;
• GDPR;
• AI Act;
• sector-specific regulation.

The distinctive feature is ex ante + ex post regulation.

The DMA is particularly significant because it can impose obligations before conventional Article 102 litigation has established dominance and abuse.

United Kingdom

The UK approach combines:

• Competition Act 1998;
• Enterprise Act 2002;
• Digital Markets, Competition and Consumers Act 2024;
• CMA digital-markets regulation;
• consumer protection;
• sectoral regulation.

The CMA's foundation-model work identified risks associated with powerful firms controlling critical inputs, distribution and ecosystems.

United States

The US relies principally upon:

• Sherman Act §1;
• Sherman Act §2;
• Clayton Act §7;
• FTC Act §5;
• FTC merger enforcement;
• sectoral and consumer-protection powers.

The US model tends to rely more heavily on case-by-case enforcement, although regulatory investigations and agency guidance are increasingly focused on AI.

The FTC, DOJ and international partners have explicitly identified generative-AI competition risks.

7. At Least 6 Important Case Laws

Because dedicated frontier-AI competition jurisprudence is still developing, traditional digital-platform, infrastructure and innovation cases are especially important. They provide the doctrinal foundations for future AI-model disputes.

Case 1 — Google Search (Shopping)

Google Search (Shopping), European Commission, 2017

Principle

The European Commission found Google had abused dominance by systematically favouring its comparison-shopping service in general search results.

Relevance to AI model governance

The case demonstrates the danger of a vertically integrated platform using control over an important infrastructure layer to favour its own downstream service.

The AI equivalent could be:

dominant operating system/search platform → preferential access for own AI assistant → exclusion of rival AI services.

It therefore supports scrutiny of:

• self-preferencing;
• ranking;
• preferential access;
• discriminatory interfaces;
• vertical integration.

Importance

Model governance should not allow a firm controlling the distribution layer to systematically disadvantage competing models.

Case 2 — Microsoft / Commission

Microsoft Corp. v Commission, Case T-201/04

Principle

The General Court upheld important aspects of the Commission's approach to Microsoft's use of its dominant position in operating systems to restrict interoperability.

AI relevance

Interoperability is one of the most important governance issues for AI ecosystems.

An AI platform controlling an operating system could potentially restrict rival models from accessing:

• system functionality;
• device information;
• APIs;
• application interfaces;
• user permissions.

Competition lesson

A model may be technologically superior but competitively ineffective if the dominant infrastructure owner controls interoperability.

Thus:

technical access + interoperability = competitive access.

Case 3 — Microsoft / Activision Blizzard

Microsoft/Activision Blizzard, European Commission, 2023

Principle

The transaction demonstrated how competition authorities can examine:

• vertical integration;
• ecosystems;
• access to infrastructure;
• foreclosure;
• cloud distribution;
• licensing remedies.

AI relevance

AI mergers may similarly combine:

cloud infrastructure + foundation model + applications + distribution.

A merger could therefore create foreclosure risks even where the parties are not direct competitors.

Governance lesson

Merger review should examine the entire AI value chain rather than merely asking whether two foundation models compete directly.

Case 4 — Illumina/GRAIL

Illumina, Inc. v FTC / Illumina–GRAIL

The FTC challenged Illumina's acquisition of GRAIL, arguing that the transaction could reduce innovation in emerging cancer-detection technology. The Fifth Circuit ultimately found substantial evidence supporting the Commission's anticompetitive determination, after which Illumina proceeded with divestiture.

AI relevance

The case is particularly important for innovation competition.

AI competition can involve future innovation rather than only current prices.

A merger between:

• a major model developer and a promising challenger;
• a cloud provider and an emerging AI infrastructure company;
• a model company and an important AI application,

may eliminate future competitive constraints.

Governance lesson

Competition authorities should examine:

Who might become the next important AI competitor?

rather than merely identifying today's competitors.

Case 5 — United States v Microsoft Corp.

United States v Microsoft Corp., 253 F.3d 34 (D.C. Cir. 2001)

Principle

Microsoft's control over the operating-system platform and its efforts to protect that position against browser competition became a foundational US monopolisation case.

AI relevance

The case is highly relevant to AI because AI assistants increasingly resemble platform services embedded within broader ecosystems.

A dominant company could potentially use:

• operating systems;
• browsers;
• search;
• cloud services;
• enterprise software;

to protect its AI position.

Competition lesson

Model governance must therefore distinguish between:

innovation that improves a product

and

integration or restrictions designed principally to prevent competitive entry.

Case 6 — Google LLC v Epic Games

Google LLC v Epic Games, Inc., 2023–2025 litigation

Principle

The dispute concerned Google's control over Android distribution and payment infrastructure and the competitive consequences of platform rules.

AI relevance

AI applications increasingly depend upon:

• app stores;
• operating systems;
• payment systems;
• default settings;
• developer access.

A dominant platform could theoretically disadvantage rival AI applications through:

• discriminatory commissions;
• restrictive distribution;
• default placement;
• API limitations;
• payment restrictions.

Governance lesson

AI competition cannot be separated from the governance of the digital infrastructure through which AI applications reach consumers.

8. Emerging AI-Specific Regulatory Precedent

Although direct AI antitrust case law remains relatively limited, regulatory developments are rapidly filling the gap.

The EU–UK–US competition authorities' 2024 joint statement expressly identified foundation-model risks and emphasised fair, open and competitive AI markets.

The CMA's foundation-model review identified three broad risks to fair, open and effective competition and developed principles for addressing them.

This is significant because it indicates a shift from:

traditional antitrust enforcement

toward

continuous governance of AI ecosystems.

9. Model Governance and Merger Control

AI merger control should examine five questions.

1. Does the transaction combine critical AI inputs?

For example:

• data;
• compute;
• chips;
• cloud;
• models;
• talent.

2. Does it remove a potential competitor?

This is the nascent-competition issue.

3. Does it increase vertical foreclosure?

Example:

Cloud provider acquires foundation-model developer.

4. Does it create ecosystem leverage?

Example:

Search + browser + operating system + AI assistant.

5. Does it increase dependence?

If customers become dependent upon a single integrated AI stack, switching costs can rise dramatically.

10. Model Governance and Algorithmic Collusion

One of the most difficult future problems is autonomous coordination.

Suppose competing AI agents independently learn that higher prices maximise profits.

No explicit human agreement may exist.

The competition-law problem becomes:

Can independent machine learning produce an unlawful coordinated outcome without traditional human communication?

Possible categories include:

1. explicit algorithmic agreement;
2. human-designed coordination;
3. information exchange through algorithms;
4. conscious parallelism;
5. autonomous adaptive coordination.

Competition authorities will need to determine whether liability attaches to:

• developers;
• deployers;
• users;
• platform operators;
• firms supplying the algorithm.

The international competition statement already recognises that algorithms can facilitate price fixing and other coordination.

11. Model Governance and Essential Facilities

A particularly difficult question concerns whether a frontier AI model could ever constitute an essential facility.

Possible scenarios include a model becoming indispensable because it controls:

• unique technical capabilities;
• critical data;
• specialised infrastructure;
• enterprise integrations;
• a large developer ecosystem.

However, essential-facility doctrine traditionally imposes demanding requirements.

Therefore, regulators should not automatically classify every powerful model as essential.

Instead, they should consider:

• indispensability;
• replicability;
• availability of alternatives;
• investment incentives;
• interoperability;
• technical feasibility;
• proportionality.

12. Model Governance and Cloud Dependency

The AI ecosystem increasingly has the structure:

chips → cloud → training → foundation model → API → application → distribution.

Control at one layer can affect competition at another.

For example:

If a cloud provider gives its affiliated AI model cheaper or technically superior access to computing resources, rival model developers may face artificial cost disadvantages.

This is why cloud governance has become an important part of AI competition policy.

The EU's current DMA work explicitly links cloud market power and AI ecosystems, including the role of AI tools and partnerships in cloud procurement.

13. Global Regulatory Coordination

A genuine global framework should establish cooperation between:

• European Commission;
• national European competition authorities;
• UK CMA;
• US DOJ;
• US FTC;
• other national competition authorities.

The 2024 EU–UK–US joint statement is an important institutional example because the three jurisdictions expressly agreed to cooperate on competition issues in generative AI.

Coordination is particularly important because AI companies operate globally while:

• models are trained internationally;
• data is sourced internationally;
• cloud infrastructure is multinational;
• users are global;
• mergers may require multiple filings.

14. Proposed Global AI Competition Governance Model

A sophisticated framework could operate as follows:

Layer 1 — Registration and transparency

↓

Layer 2 — Model-risk and competition assessment

↓

Layer 3 — Access and interoperability obligations

↓

Layer 4 — Continuous monitoring

↓

Layer 5 — Merger and investment screening

↓

Layer 6 — Algorithmic-conduct enforcement

↓

Layer 7 — Remedies

Possible remedies include:

• interoperability;
• data access;
• API access;
• non-discrimination;
• licensing;
• firewall obligations;
• behavioural commitments;
• divestiture;
• structural separation.

15. Key Legal Tension: Safety vs Competition

An important difficulty is that AI safety governance itself can create competition problems.

For example, a consortium of major AI firms may create common safety standards.

This can be beneficial.

But if the same firms collectively establish technical requirements that smaller competitors cannot satisfy, safety regulation could become a barrier to entry.

Therefore:

Safety standards must be proportionate, transparent, technically justified and competitively neutral.

Otherwise, incumbent firms could potentially use governance standards to transform regulatory compliance into an entry barrier.

16. Regulatory Capture Risk

Model governance also creates a new form of regulatory capture.

Large AI firms may possess:

• superior technical expertise;
• enormous computing resources;
• specialised lawyers;
• proprietary safety information;
• extensive lobbying capacity.

Regulators may therefore become dependent upon the firms they regulate for technical knowledge.

A strong framework should consequently provide:

• independent technical expertise;
• regulator access to model documentation;
• audit powers;
• secure testing environments;
• independent model evaluations;
• information-sharing between competition authorities.

17. Conclusion

Global model governance for AI competition law represents a transition from conventional antitrust toward continuous governance of technological ecosystems.

The central competition concern is not merely whether one AI model is dominant. It is whether a firm can combine control over:

data + compute + model + API + cloud + operating system + distribution + applications

to create an enduring competitive bottleneck.

The major legal principles emerging from cases such as Microsoft, Google Shopping, Microsoft/Activision, Illumina/GRAIL and Google/Epic suggest that future AI competition enforcement will focus heavily on interoperability, self-preferencing, vertical foreclosure, innovation competition, access restrictions, ecosystem leverage and potential competition.

The most appropriate global framework is therefore a hybrid model combining:

1. ex ante AI governance;
2. traditional Articles 101/102-type competition enforcement;
3. US monopolisation and merger law;
4. UK digital-market regulation;
5. DMA-style interoperability and access obligations;
6. continuous monitoring of foundation-model ecosystems;
7. international cooperation between competition authorities.

The emerging regulatory direction is already clear: AI competition policy is moving from asking “Is the model dominant?” toward the broader question “Does the governance architecture of the model ecosystem allow competition to remain contestable?”