Distributed AI Governance and Competition Fragmentation

1. Introduction

Distributed AI governance refers to governance arrangements in which control over the development, deployment, monitoring, updating, auditing, data access, model standards, or enforcement of artificial-intelligence systems is divided among multiple undertakings, platforms, infrastructure providers, developers, regulators, standard-setting bodies, consortiums, and autonomous software agents.

This structure can have significant competition-law consequences. Distribution of decision-making does not necessarily produce decentralised competition. Several formally independent participants may collectively control essential inputs, technical standards, data, computing capacity, APIs, model access, identity systems, or compliance infrastructure. Conversely, excessive fragmentation of governance can make it difficult for competition authorities to determine who possesses market power, who coordinated the conduct, and against whom a remedy should be imposed.

The central competition-law problem can therefore be expressed as:

Distributed governance may reduce visible concentration while increasing functional concentration, or may fragment markets so extensively that effective competition itself becomes difficult to maintain.

This issue can be analysed particularly through Articles 101 and 102 TFEU, UK competition law under the Competition Act 1998, merger control, abuse-of-dominance principles, information-exchange doctrine, essential-facilities reasoning, and digital-platform regulation.

2. Meaning of Competition Fragmentation

Competition fragmentation occurs when a market is divided among numerous technical, institutional, geographical, or governance layers in a manner that affects competitive conditions.

In AI markets, fragmentation may arise between:

1. Foundation-model developers
2. Cloud-computing providers
3. GPU/accelerator suppliers
4. Data providers
5. Model distributors
6. API intermediaries
7. Application developers
8. AI safety/audit providers
9. Identity and authentication providers
10. AI marketplaces
11. Standard-setting organisations
12. Regulators and supervisory bodies
13. Autonomous AI agents

Fragmentation is not automatically harmful. Multiple independent suppliers can increase competition. The problem arises where interoperability, switching, access, standards, data, or governance rules cause the fragmented layers to operate as a single strategic bottleneck.

3. Distributed Governance Does Not Necessarily Mean Decentralisation

A fundamental distinction should be made between distributed decision-making and distributed economic power.

Formal distribution

Several independent entities make decisions.

Functional centralisation

A small number of entities nevertheless control the resources necessary for everyone else to participate.

For example:

AI developer → cloud provider → GPU supplier → model registry → API gateway → application

Six entities may appear to participate, but if one or two firms control the critical interfaces, effective competition may remain highly concentrated.

Thus:

Organisational fragmentation ≠ competitive decentralisation.

4. Principal Forms of Distributed AI Governance

A. Multi-layer governance

Different firms control different stages:

• data acquisition;
• training;
• compute;
• model development;
• deployment;
• monitoring;
• auditing.

Competition authorities must therefore identify whether power exists at a particular layer or across the ecosystem.

B. Consortium governance

AI developers may cooperate through industry associations or technical consortia.

This can produce legitimate efficiencies, but common standards may also become mechanisms for:

• exclusion;
• discriminatory certification;
• coordinated pricing;
• information exchange;
• restrictions on interoperability.

The legal question is whether cooperation is genuinely necessary for standardisation or instead reduces independent competitive decision-making.

C. Open-source governance

Open-source AI can fragment development by allowing many participants to modify and distribute models.

However, governance can become concentrated around:

• repositories;
• maintainers;
• licensing decisions;
• compute infrastructure;
• model weights;
• safety certification.

A nominally open ecosystem may therefore develop centralised gatekeeping points.

D. Decentralised or DAO-style AI governance

Voting may be distributed among token holders or participants.

Competition concerns can nevertheless arise where:

• voting rights are concentrated;
• large token holders influence governance;
• validators coordinate economic conduct;
• governance participants exchange competitively sensitive information.

The decentralised label does not itself determine the application of competition law.

5. Competition Risks Created by Distributed AI Governance

5.1 Hidden concentration

Several independent firms may collectively depend upon one infrastructure provider.

For example:

Multiple AI companies → one dominant cloud → common compute layer

The downstream market may look competitive while the upstream infrastructure market remains concentrated.

5.2 Governance bottlenecks

A technical body may determine:

• who obtains certification;
• which models are compatible;
• which safety standards apply;
• what API protocols are recognised.

If participation is indispensable, governance itself may become a competitive bottleneck.

5.3 Standard-setting exclusion

A dominant undertaking may influence technical standards to disadvantage competing models.

Potential strategies include:

• proprietary compatibility requirements;
• discriminatory certification;
• exclusionary safety requirements;
• refusal to recognise rival protocols.

This can implicate abuse-of-dominance principles.

6. Information Exchange Through Distributed Governance

AI governance frequently requires participants to share information.

Examples include:

• model-performance data;
• capacity forecasts;
• pricing information;
• safety information;
• customer demand;
• deployment schedules;
• future product plans.

The same governance structure that produces safety cooperation can therefore create an infrastructure for coordinated conduct.

The competition-law question is whether information sharing reduces strategic uncertainty between competitors.

7. Algorithmic Coordination

Distributed AI governance may create particularly difficult Article 101 problems.

Suppose competing AI firms use a common governance platform that:

1. collects market information;
2. processes it through AI;
3. recommends prices;
4. updates recommendations automatically;
5. communicates them to participating firms.

No human employee may explicitly agree to fix prices.

Nevertheless, competition law can examine whether the institutional arrangement itself facilitates coordinated conduct.

The absence of an email saying "let us fix prices" does not necessarily resolve the issue.

8. Common Algorithms and Competition

A further problem arises where competitors purchase AI optimisation services from the same provider.

For example:

Competitor A → common pricing algorithm ← Competitor B

If the provider receives competitively sensitive information from both firms and uses it to influence their pricing, the provider can potentially become a coordination intermediary.

This makes attribution particularly complex.

9. Interoperability Fragmentation

AI ecosystems may become fragmented because models, agents and applications use incompatible:

• APIs;
• model formats;
• identity systems;
• data standards;
• agent protocols;
• safety certifications.

Interoperability restrictions can raise competition concerns when they increase:

• switching costs;
• entry barriers;
• ecosystem dependence;
• customer lock-in.

A dominant undertaking may therefore use technical fragmentation as an exclusionary strategy.

10. Data Fragmentation

AI development depends heavily on data.

Governance may distribute data among:

• public bodies;
• platforms;
• data trusts;
• publishers;
• model developers;
• cloud providers.

Fragmented access can make it difficult for new entrants to obtain sufficient training or evaluation data.

Where a dominant undertaking controls indispensable data, competition law may potentially address:

• discriminatory access;
• refusal to supply;
• self-preferencing;
• tying;
• exclusionary licensing.

11. Compute Fragmentation

AI markets are also dependent upon specialised computing resources.

Compute may be divided among:

• GPU manufacturers;
• cloud providers;
• data-centre operators;
• model developers;
• AI-compute brokers.

This can create vertical concentration without conventional horizontal market shares.

A firm with modest downstream AI-market share may nevertheless possess substantial competitive power because it controls an indispensable upstream input.

12. Case Laws

1. United States v. Apple Inc. (2024)

The U.S. government's antitrust case against Apple illustrates how control over an ecosystem can generate competition concerns even where numerous independent developers and businesses operate within it.

The relevant conceptual lesson for distributed AI governance is that control over interfaces, interoperability, access conditions and ecosystem rules can confer market power beyond the firm's immediate product.

For AI ecosystems, similar questions may arise where a platform controls:

• AI-agent access;
• APIs;
• app distribution;
• identity;
• device-level AI functionality.

Principle

Ecosystem governance can itself be an important source of competitive power.

2. United States v. Google LLC — Search (2024)

The Google search litigation demonstrates the importance of analysing distribution arrangements and access points rather than looking only at the underlying technology.

Google's agreements concerning default search distribution were examined as mechanisms that could reinforce its position.

The AI analogy is significant: control over distribution channels can allow a firm to strengthen an apparently contestable technology market.

Principle

Control over distribution and default access can reinforce durable market power.

3. United States v. Microsoft Corp. (2001)

The Microsoft litigation remains one of the most important authorities for analysing technological ecosystems.

Microsoft's conduct concerning Internet Explorer demonstrated how a firm with substantial power in one layer of technology could use that position to influence competition in adjacent layers.

The case is especially relevant to AI because AI ecosystems involve numerous adjacent layers:

operating system → cloud → model → API → application → agent.

Principle

Market power in one technological layer may be leveraged into neighbouring markets.

4. European Commission v. Microsoft Corp. (T-201/04)

The EU Microsoft case concerned Microsoft's refusal to provide interoperability information necessary for competing work-group server products.

The case is highly relevant to distributed AI governance because interoperability can become a competitive necessity.

Where an AI platform controls an interface required by competing AI systems, the legal analysis may ask whether denial or discriminatory provision of interoperability information excludes rivals.

Principle

Interoperability can become competitively significant where access is necessary for effective rivalry.

5. Google Android — European Commission (Case AT.40099)

The EU Android decision examined Google's contractual arrangements concerning Android, including tying and restrictions affecting competing search and distribution arrangements.

Its broader significance is the examination of ecosystem architecture as a mechanism through which dominance in one layer can influence competitive conditions elsewhere.

For AI ecosystems, comparable issues may arise where:

• a dominant cloud provider bundles AI models;
• a dominant operating system favours its own AI assistant;
• an AI marketplace restricts competing models.

Principle

Vertical restrictions can reinforce ecosystem-level dominance.

6. Google Shopping — Google and Alphabet v Commission (C-48/22 P)

The Google Shopping litigation concerned Google's preferential treatment of its comparison-shopping service within general search results.

The case is particularly relevant to distributed AI governance because AI interfaces increasingly determine which information, services, models or applications users encounter.

An AI platform could potentially favour:

• its own model;
• affiliated applications;
• preferred agents;
• preferred data providers.

Principle

A dominant platform's control over ranking and access can become an exclusionary competition issue.

7. Bronner v Mediaprint (C-7/97)

Bronner is a leading EU authority concerning refusal to supply and the essential-facilities doctrine.

The Court established a demanding framework for imposing an obligation upon a dominant undertaking to provide access to infrastructure.

Its importance to AI governance is substantial because firms may argue that:

• specialised compute;
• proprietary datasets;
• model interfaces;
• AI infrastructure

are indispensable facilities.

Principle

Not every important input is an essential facility; indispensability must satisfy a demanding legal test.

8. Slovak Telekom v Commission (C-165/19 P)

The Slovak Telekom litigation concerns access to infrastructure and exclusionary conduct by a dominant undertaking.

It demonstrates how competition law can address strategies through which a dominant firm controls an upstream infrastructure layer while competing downstream.

The analogy to AI is particularly strong where a cloud or infrastructure provider simultaneously supplies compute and competes in AI services.

Principle

Vertical integration can create exclusionary incentives where infrastructure control and downstream competition coexist.

9. Intel v Commission (C-413/14 P)

Intel is important for analysing exclusionary rebates and the assessment of competitive effects.

For AI markets, the underlying lesson is that contractual incentives should not be analysed solely by their formal structure.

A dominant AI infrastructure provider could theoretically provide:

• compute rebates;
• preferential capacity;
• model-hosting incentives;
• exclusivity discounts.

Principle

The practical competitive effects of conduct may matter more than its contractual label.

10. Eturas v Lietuvos Respublikos konkurencijos taryba (C-74/14)

This case is especially valuable for digitally mediated coordination.

A common electronic booking system communicated pricing restrictions to participating travel agencies. The Court addressed the circumstances in which participation in a digital system can contribute to an infringement.

The case offers a useful analogy for AI governance systems in which a central digital infrastructure communicates commercial parameters to competing undertakings.

Principle

Digital systems can facilitate coordination even when the mechanism of communication is technological rather than traditional human-to-human communication.

13. Synthesis of the Case Law

The cases collectively establish several propositions relevant to distributed AI governance:

14. Distributed Governance and Article 101

Article 101 may become relevant where independent AI undertakings coordinate through governance mechanisms.

Potential arrangements include:

• common pricing algorithms;
• common procurement systems;
• coordinated capacity allocation;
• collective exclusion standards;
• information-sharing platforms;
• joint model-access rules.

The key issue is whether the arrangement has the object or effect of restricting competition.

A governance arrangement should therefore distinguish between:

Legitimate cooperation

• cybersecurity;
• genuine technical interoperability;
• safety standards;
• responsible AI testing;
• non-commercial research.

Potentially problematic cooperation

• common pricing;
• output restrictions;
• customer allocation;
• exclusionary certification;
• exchange of future strategic plans;
• coordinated refusal to deal.

15. Distributed Governance and Article 102

Article 102 becomes relevant where one undertaking occupies a dominant position.

AI dominance may arise from control over:

• data;
• compute;
• cloud infrastructure;
• foundation models;
• APIs;
• distribution;
• AI-agent ecosystems.

Potential abuses include:

Refusal to provide access

A dominant infrastructure provider refuses access to essential AI infrastructure.

Discrimination

The provider gives affiliated AI products better infrastructure or API access.

Self-preferencing

A platform systematically promotes its own AI services.

Tying

Access to one essential service is conditioned on purchasing another.

Exclusive dealing

Customers are discouraged from using competing AI systems.

Interoperability restrictions

Technical barriers make switching to competing models difficult.

16. The "Fragmentation Paradox"

The most important theoretical concept is the fragmentation paradox.

Stage 1 — Decentralisation

AI governance is distributed among many participants.

Stage 2 — Dependence

Those participants rely upon common infrastructure.

Stage 3 — Standardisation

They adopt common technical and governance standards.

Stage 4 — Bottleneck formation

Control migrates toward the entities administering those standards.

Stage 5 — Functional centralisation

The market becomes effectively controlled by a limited number of infrastructure or governance providers.

Thus:

A market can become more institutionally distributed while becoming economically more centralised.

17. Fragmentation Can Also Weaken Competition

The opposite problem is also possible.

Excessive fragmentation may create:

• incompatible standards;
• duplicated compliance requirements;
• high transaction costs;
• fragmented liquidity;
• limited interoperability;
• regional barriers;
• duplicated AI-safety certification.

These effects can raise entry costs for smaller competitors.

Consequently, competition law should not mechanically treat decentralisation as pro-competitive.

18. Collective Dominance and Distributed AI

A particularly difficult future question is whether several firms can collectively exercise market power.

Suppose:

• three cloud providers control most AI compute;
• all three adopt a common access standard;
• the standard restricts interoperability;
• customers cannot practically switch.

The firms may remain legally separate but economically behave as an interconnected governance structure.

Competition authorities would need to determine whether the evidence establishes:

• an agreement;
• concerted practice;
• coordinated effects;
• collective dominance;
• or merely parallel conduct.

19. AI Agents as Governance Participants

Autonomous AI agents introduce a new problem.

An agent may:

• negotiate prices;
• allocate resources;
• select suppliers;
• execute contracts;
• adjust output;
• respond to competitors.

The legal question becomes:

Who is responsible when an autonomous agent produces competitively restrictive conduct?

Potentially relevant actors include:

1. model developer;
2. deploying company;
3. agent operator;
4. infrastructure provider;
5. data provider;
6. governance platform.

Autonomy does not itself eliminate legal responsibility.

20. Competition Fragmentation and Merger Control

Distributed AI ecosystems also create novel merger issues.

A transaction need not combine two dominant foundation-model companies.

Competitive significance may arise from acquisitions involving:

• AI startups;
• cloud infrastructure;
• specialised chips;
• datasets;
• model distributors;
• AI safety companies;
• agent platforms.

A vertical acquisition can eliminate a future competitor or give the acquiring company privileged access to a critical input.

Therefore, conventional market-share analysis may understate competitive harm.

21. Remedy Problems

Distributed governance makes remedies particularly complicated.

Suppose an AI ecosystem contains:

cloud provider + model developer + API intermediary + marketplace + application.

An authority may need to choose between:

Structural remedies

• divestiture;
• separation;
• prohibition of acquisition.

Behavioural remedies

• interoperability;
• non-discrimination;
• data access;
• API access;
• transparency.

Governance remedies

• independent oversight;
• audit mechanisms;
• neutral standards;
• access committees.

Technical remedies

• portability;
• API interoperability;
• model switching;
• data portability.

The remedy must address the actual bottleneck, not merely the most visible company.

22. Regulatory Fragmentation

Competition enforcement itself may become fragmented.

AI companies may simultaneously face:

• competition authorities;
• data-protection regulators;
• AI regulators;
• telecommunications regulators;
• cybersecurity authorities;
• sector regulators.

Different authorities may impose inconsistent obligations.

For example:

Competition law → interoperability
Data protection → restricted data sharing
AI regulation → safety restrictions
Cybersecurity regulation → restricted system access

The resulting regulatory conflict may itself affect market entry.

23. Key Legal Questions for Competition Authorities

When analysing distributed AI governance, authorities should ask:

1. Who controls the critical input?
2. Who controls access?
3. Who controls interoperability?
4. Who controls technical standards?
5. Who receives commercially sensitive information?
6. Who determines ranking or allocation?
7. Can users switch models easily?
8. Can developers migrate their data?
9. Can competing models access equivalent compute?
10. Does governance favour affiliated undertakings?
11. Are autonomous agents coordinating economic conduct?
12. Does formal decentralisation conceal functional centralisation?

24. Competition-Law Analytical Framework

A useful framework is:

AI ecosystem mapping
↓
Identify governance layers
↓
Identify bottlenecks
↓
Define relevant markets
↓
Assess market power
↓
Analyse vertical relationships
↓
Examine information flows
↓
Test coordination risks
↓
Assess exclusion/interoperability restrictions
↓
Evaluate efficiencies
↓
Design targeted remedies

25. Distinguishing Legitimate Distributed Governance From Anticompetitive Fragmentation

26. Overall Legal Position

Distributed AI governance presents competition law with a structural challenge: traditional antitrust analysis focuses heavily on identifiable firms and identifiable markets, whereas AI ecosystems increasingly distribute economic decision-making across networks.

The decisive question should therefore not simply be:

"How many firms participate?"

It should be:

"Where is effective economic control located?"

A system containing twenty nominally independent participants may still be competitively centralised if a small number of firms control compute, data, interfaces, standards, distribution, or governance.

Conversely, a genuinely decentralised AI ecosystem can promote competition by lowering entry barriers and enabling interoperability.

27. Conclusion

Distributed AI governance can produce both more competition and less competition. Its pro-competitive potential lies in decentralised innovation, interoperability, open standards and multiple independent suppliers. Its danger lies in hidden centralisation, governance bottlenecks, common algorithms, data concentration, compute dependence, interoperability restrictions and coordinated decision-making.

The existing case law—from Microsoft, Bronner, Slovak Telekom, Intel, Eturas, Google Android and Google Shopping—provides important doctrinal tools even though these cases predate many contemporary AI governance structures.

The future competition-law task will therefore be to move from a purely firm-centred analysis toward an ecosystem-and-governance analysis, while preserving the requirements of causation, market definition, dominance, agreement or concerted practice, competitive effects, objective justification and proportionality.