Distributed AI Firms And Hidden Centralization Mechanisms

1. Introduction

Distributed AI firms are enterprises or networks in which AI development, computing, data collection, model training, inference, software deployment, decision-making, or governance is spread across multiple legally distinct entities. These may include foundation-model developers, cloud providers, GPU suppliers, data brokers, AI application firms, independent developers, model marketplaces, API providers, and autonomous agents.

The apparent decentralisation of such an ecosystem can, however, conceal substantial economic centralisation. A market may contain hundreds of nominally independent participants while a small number of firms control the indispensable inputs or coordination points—such as:

• computing capacity;
• advanced GPUs and accelerators;
• foundation-model weights;
• proprietary training data;
• APIs and inference infrastructure;
• model repositories;
• app stores and distribution channels;
• interoperability standards;
• identity and authentication systems;
• cloud switching mechanisms;
• safety and compliance infrastructure; and
• access to customers and procurement markets.

Competition law therefore cannot examine only legal ownership or the number of firms. It must also examine functional control.

The central question is:

Does a formally distributed AI ecosystem actually operate as a centrally controlled competitive structure because one or a few firms control the critical coordination points?

2. Meaning of Hidden Centralisation

Hidden centralisation exists where decision-making or economic dependency is concentrated without necessarily being reflected in corporate ownership.

A useful distinction is:

Formal decentralisation

Many independent firms appear to make their own decisions.

Functional centralisation

A smaller number of firms determine the conditions under which those firms can operate.

For example:

100 AI application companies → dependent on 2 cloud providers → dependent on 3 accelerator suppliers → dependent on 1–2 foundation-model ecosystems

The market may therefore look competitive at the application layer while being highly concentrated at the infrastructure layer.

3. Major Centralisation Mechanisms

A. Cloud-Compute Centralisation

AI firms often require enormous quantities of specialised computing resources.

If a few cloud providers control access to:

• GPUs;
• TPUs;
• high-performance networking;
• storage;
• distributed training systems;
• inference infrastructure; and
• AI-specific software stacks,

they may become gatekeepers to AI markets.

A cloud provider can potentially influence downstream competition through:

1. preferential pricing;
2. capacity allocation;
3. technical compatibility;
4. bundled services;
5. preferential access to new hardware;
6. data-transfer charges;
7. restrictive contractual terms; and
8. preferential treatment of affiliated AI products.

Competition concern

The issue is not merely whether cloud prices are excessive. The deeper concern is whether cloud infrastructure becomes an essential competitive bottleneck.

4. Foundation Models As Central Coordination Points

A supposedly distributed AI ecosystem may have many applications but depend upon a small number of foundation models.

For example:

Independent developers → common foundation model → common API → common cloud infrastructure

The foundation-model provider can therefore influence:

• model access;
• pricing;
• technical functionality;
• API availability;
• usage restrictions;
• safety rules;
• data policies;
• fine-tuning rights; and
• downstream compatibility.

This creates a possible vertical leverage problem.

A dominant foundation-model provider that also operates applications competing with its customers may have incentives to discriminate against downstream rivals.

5. API Dependence

APIs can become centralisation mechanisms even where AI models are technically replaceable.

A dominant API provider may create:

• proprietary interfaces;
• high switching costs;
• incompatible data formats;
• application-specific integrations;
• contractual restrictions;
• rate limits;
• preferential latency;
• differential access to new functionality.

Once developers build applications around a particular API, replacement becomes costly.

This creates technological lock-in rather than conventional contractual lock-in.

6. Data Centralisation

AI markets are also susceptible to concentration through data.

A firm may control:

• consumer interaction data;
• search data;
• behavioural data;
• transaction data;
• location data;
• enterprise datasets;
• labelled datasets; or
• feedback generated through millions of AI interactions.

Even when models and applications are distributed, control over a unique dataset can give a firm a structural advantage.

The competition-law question becomes:

Can rivals realistically reproduce the data advantage, or is the dataset functionally indispensable?

7. Hidden Centralisation Through Standards

Centralisation may also arise through technical standards.

A dominant firm may control:

• model formats;
• authentication systems;
• safety protocols;
• agent protocols;
• cloud interfaces;
• developer tools;
• benchmark systems; or
• interoperability standards.

Standards can reduce transaction costs and encourage innovation. But if controlled by a dominant undertaking, they can also become a mechanism for foreclosing competing technologies.

The danger is particularly significant where the standard setter can determine who receives interoperability.

8. AI Agent Ecosystems

Autonomous AI agents create another form of hidden centralisation.

A large number of agents may appear independent:

Agent A + Agent B + Agent C + Agent D

but all may rely upon:

Common model → common API → common cloud → common identity system → common payment infrastructure

The agents are therefore decentralised in execution but centralised in infrastructure.

This distinction is important because competition authorities should examine the dependency graph, rather than simply counting market participants.

9. Common Algorithmic Infrastructure

Another risk occurs when apparently independent firms use the same:

• pricing engine;
• recommendation system;
• demand forecasting model;
• optimisation software;
• procurement algorithm; or
• AI market-making platform.

The firms may not communicate directly with one another, yet a common algorithmic intermediary can coordinate market outcomes.

This creates an important distinction between:

Direct coordination

Competitors communicate with each other.

Algorithmically mediated coordination

Competitors independently provide information to a common system that produces strategically similar outcomes.

The latter can be particularly difficult for traditional antitrust doctrine because the central coordinating mechanism may not itself be a competitor in the downstream market.

10. Corporate Partnerships As Centralisation Mechanisms

Formal independence does not necessarily mean economic independence.

AI companies may enter:

• joint ventures;
• licensing agreements;
• cloud partnerships;
• exclusive distribution agreements;
• strategic investments;
• model-hosting arrangements;
• compute-sharing agreements; and
• long-term supply contracts.

A network of minority investments and exclusive agreements can produce de facto control without outright acquisition.

Competition authorities therefore need to investigate:

• board rights;
• veto rights;
• information rights;
• exclusivity;
• economic dependence;
• technical dependence;
• financing;
• common personnel; and
• contractual termination rights.

11. Relevant Case Laws

The following cases do not all concern generative AI specifically. They provide important competition-law principles that can be applied to distributed AI ecosystems.

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

The Microsoft litigation is foundational for understanding control over technological bottlenecks.

Microsoft's dominance in PC operating systems gave it the ability and incentive to disadvantage competing technologies, particularly through control over interfaces and distribution.

Relevance to distributed AI

An AI infrastructure provider could similarly possess a bottleneck through:

• cloud infrastructure;
• operating environments;
• APIs;
• developer tools; or
• distribution channels.

The lesson is that competition law can address exclusionary conduct occurring at adjacent technological layers, not merely within the dominant firm's primary product.

2. Google Search (Shopping) — European Commission / General Court

The Google Shopping litigation concerned Google's use of dominance in general search to favour its own comparison-shopping service.

The importance of the case lies in leveraging dominance from one layer of a digital ecosystem into another.

Application to AI

A dominant AI infrastructure provider might theoretically use control over:

compute → model hosting → API → application distribution

to favour its own downstream applications.

If rival applications receive inferior access, visibility, interoperability, latency, or functionality, competition concerns can arise.

3. Google Android — European Commission / General Court

The Android proceedings involved Google's use of contractual arrangements and ecosystem control to reinforce its position across interconnected digital markets.

The case illustrates how apparently separate products can operate as part of a single ecosystem of competitive leverage.

AI relevance

A comparable AI ecosystem might combine:

• operating systems;
• cloud services;
• foundation models;
• app stores;
• browsers;
• developer tools; and
• AI assistants.

A competition authority should therefore consider whether contractual and technological integration creates ecosystem foreclosure.

4. United Brands v Commission (1978)

The case is a classic authority on dominance and the ability of a powerful undertaking to behave to an appreciable extent independently of competitors, customers, and consumers.

AI relevance

The United Brands principle is useful because AI dominance may not necessarily be demonstrated through traditional market share alone.

Indicators may include:

• dependency of customers;
• technological barriers to entry;
• control over unique inputs;
• switching costs;
• network effects; and
• inability of downstream firms to operate without the infrastructure provider.

Thus, economic dependency can be highly relevant to determining market power.

5. Bronner v Mediaprint (1998)

Bronner is particularly relevant to access to indispensable infrastructure.

The Court established a demanding framework for imposing access obligations under the essential-facilities doctrine.

AI relevance

Suppose a dominant firm controls an AI infrastructure resource that:

1. is indispensable;
2. cannot realistically be replicated;
3. cannot reasonably be substituted; and
4. prevents effective competition if access is denied.

A refusal to provide access may then attract competition-law scrutiny.

Potential examples include:

• unique compute infrastructure;
• indispensable interoperability interfaces;
• irreplaceable datasets; or
• critical AI distribution infrastructure.

However, Bronner also demonstrates that not every commercially valuable AI resource constitutes an essential facility.

6. IMS Health v NDC Health (2004)

IMS Health further developed the European essential-facilities framework concerning access to intellectual-property-protected resources.

The case is significant because it demonstrates the tension between:

• protecting innovation and intellectual property; and
• preventing exclusionary control over indispensable inputs.

AI relevance

Foundation-model weights, proprietary datasets, model interfaces, and specialised AI technologies may be protected by intellectual-property rights.

Competition law cannot simply transform every proprietary AI technology into a compulsory-access resource.

The difficult question is whether withholding access eliminates effective competition in a distinct downstream market and whether access is objectively indispensable.

7. Aspen Skiing Co. v Aspen Highlands Skiing Corp. (1985)

The US Supreme Court found liability in a refusal-to-deal situation involving a previously established cooperative arrangement.

The case is important for identifying circumstances in which a dominant firm abandons a profitable relationship in a manner that appears designed to exclude a rival.

AI relevance

Suppose a dominant AI infrastructure provider previously permitted interoperability or access to a competing AI service and then abruptly withdraws it.

Evidence such as:

• historical cooperation;
• commercial sacrifice;
• exclusionary intent;
• lack of legitimate business justification; and
• harm to competition

could become important.

8. Lorain Journal Co. v United States (1951)

Lorain Journal demonstrates the principle that a dominant firm cannot necessarily use its control over an important distribution channel to exclude competitors.

AI relevance

AI platforms increasingly operate as distribution channels.

An AI assistant, app marketplace, cloud marketplace, or model marketplace could potentially become a gateway through which customers reach competing AI services.

If the gateway operator discriminates against competing products to preserve its own position, traditional exclusionary principles may become relevant.

12. Comparative Case-Law Matrix

13. The “Distributed Surface, Centralised Core” Problem

One of the most important concepts is the distinction between surface-level decentralisation and core-level centralisation.

Surface

• thousands of developers;
• numerous AI applications;
• independent agents;
• multiple startups;
• open-source projects.

Core

• few cloud providers;
• few accelerator suppliers;
• few foundation models;
• few data repositories;
• few distribution platforms.

This can produce:

Many firms at the edge + few gatekeepers at the centre = hidden concentration.

Consequently, conventional measures of concentration may understate actual market power.

14. Network Effects

AI ecosystems exhibit strong network effects.

More users generate:

more data → better models → more users → more developers → more applications → more data

This creates a feedback loop.

A dominant platform may therefore become progressively harder to challenge even if competitors can initially enter.

Network effects can reinforce centralisation through:

• data accumulation;
• developer ecosystems;
• model improvement;
• reputation;
• interoperability;
• complementary applications; and
• user familiarity.

15. Switching Costs

Switching costs are particularly significant in AI.

An enterprise may have invested heavily in:

• fine-tuning;
• prompt libraries;
• model-specific workflows;
• API integrations;
• employee training;
• proprietary data pipelines;
• monitoring systems; and
• compliance architecture.

Moving from Provider A to Provider B may therefore require significant technical restructuring.

This creates artificial persistence of market power even where nominal alternatives exist.

16. Vertical Foreclosure

A vertically integrated AI undertaking might operate simultaneously at several levels:

GPU access → cloud → foundation model → API → application → distribution

The firm could potentially disadvantage competitors through:

• tying;
• bundling;
• self-preferencing;
• discriminatory access;
• exclusive contracts;
• interoperability restrictions;
• predatory pricing;
• technical degradation; or
• preferential allocation of scarce compute.

The more layers controlled by one undertaking, the greater the potential for vertical leverage.

17. Killer Acquisitions and Hidden Consolidation

Distributed AI markets can also become centralised through acquisition.

A major platform may acquire:

• promising AI startups;
• model developers;
• specialised data firms;
• agent companies;
• developer-tool providers; or
• AI safety firms.

Even where the acquired firm has limited current revenue, it may possess significant future competitive potential.

Competition authorities must therefore examine:

• innovation competition;
• nascent competitors;
• data assets;
• technical capabilities;
• developer ecosystems; and
• potential displacement of incumbent technology.

18. Minority Investments

Centralisation can occur without formal mergers.

A dominant infrastructure company may hold minority stakes in several AI firms.

Although each investment may appear individually harmless, the combined network may create:

• information flows;
• strategic alignment;
• common incentives;
• reduced rivalry;
• board influence; and
• dependence on the investor.

This creates a form of network concentration rather than conventional corporate concentration.

19. Algorithmic Governance As Centralisation

AI ecosystems may increasingly use automated governance.

A central platform can determine:

• who obtains compute;
• which models are approved;
• which applications receive distribution;
• what content is permitted;
• which developers receive API access;
• which transactions are flagged; and
• which users receive particular services.

The critical competition question becomes:

Who controls the algorithm that determines access to the market?

This is potentially more important than who owns the individual applications.

20. Competition-Law Theories Potentially Engaged

Hidden centralisation can implicate several doctrines.

Article 102 TFEU / equivalent abuse-of-dominance rules

Potential theories include:

• refusal to supply;
• discriminatory access;
• tying and bundling;
• self-preferencing;
• exclusionary rebates;
• margin squeeze;
• predatory pricing;
• exploitative conduct; and
• leveraging.

Article 101 TFEU / equivalent cartel provisions

Relevant where independent AI firms coordinate through:

• information exchange;
• common algorithms;
• standard-setting;
• joint purchasing;
• restrictive agreements; or
• common intermediaries.

Merger control

Relevant where centralisation occurs through:

• acquisitions;
• joint ventures;
• serial acquisitions;
• minority investments; or
• acquisitions of nascent competitors.

Essential-facilities principles

Potentially relevant where an infrastructure input is genuinely indispensable.

21. Measuring Hidden Centralisation

Traditional HHI analysis may be insufficient.

Authorities should consider a multi-layer concentration analysis:

Layer 1 — Hardware

Who controls advanced accelerators?

Layer 2 — Compute

Who controls AI-capable cloud capacity?

Layer 3 — Models

Who controls foundation models?

Layer 4 — Data

Who controls irreplaceable datasets?

Layer 5 — APIs

Who controls model access?

Layer 6 — Distribution

Who controls access to customers?

Layer 7 — Governance

Who determines technical and safety standards?

The resulting structure can be represented as:

Hardware concentration
↓
Compute concentration
↓
Model concentration
↓
API concentration
↓
Application dependence
↓
Distribution concentration

This provides a much better picture of actual competitive power.

22. Key Enforcement Indicators

Competition authorities should investigate whether:

1. independent firms depend on the same infrastructure;
2. customers face substantial switching costs;
3. interoperability is restricted;
4. proprietary standards create lock-in;
5. dominant firms control scarce compute;
6. data advantages cannot be replicated;
7. infrastructure providers compete downstream;
8. access is discriminatory;
9. minority investments create common incentives;
10. acquisitions eliminate potential competitors;
11. algorithms facilitate coordination; and
12. technical architecture makes competitors dependent on a single ecosystem.

23. Remedies

Potential remedies include:

Structural remedies

• divestiture;
• separation of infrastructure and downstream operations;
• restrictions on acquisitions.

Behavioural remedies

• non-discriminatory access;
• interoperability;
• API portability;
• data portability;
• transparent access criteria;
• non-exclusive contracts.

Technical remedies

• open interfaces;
• model portability;
• interoperable model formats;
• switching tools;
• data-export mechanisms.

Monitoring remedies

Authorities may require:

• algorithmic audits;
• access logs;
• discriminatory-treatment monitoring;
• reporting of compute allocation;
• notification of major infrastructure changes.

24. Central Legal Challenge

The greatest difficulty is distinguishing legitimate technological integration from anticompetitive centralisation.

Integration can produce substantial efficiencies:

• lower costs;
• better security;
• faster innovation;
• improved reliability;
• better model performance.

Competition law should therefore not condemn concentration merely because an AI ecosystem is vertically integrated.

The crucial inquiry is:

Does centralisation result from superior efficiency and innovation, or is it being maintained through exclusionary mechanisms that prevent rivals from competing effectively?

25. Conclusion

Distributed AI markets require competition law to move beyond the simplistic question of “How many firms exist?”

The more important question is:

Who controls the infrastructure through which all those firms must operate?

A market containing hundreds of AI firms can still be highly concentrated if a handful of undertakings control compute, data, foundation models, APIs, distribution, standards and governance.

The principles emerging from Microsoft, Google Shopping, Google Android, United Brands, Bronner, IMS Health, Aspen Skiing and Lorain Journal demonstrate that competition law already possesses conceptual tools for analysing technological bottlenecks, ecosystem leverage, indispensable infrastructure, distribution control and exclusionary conduct.

The future challenge is to adapt those principles to an AI economy in which centralisation may be invisible at the corporate level but obvious at the infrastructure level.

Core proposition

Distributed ownership ≠ distributed market power.

The appropriate competition-law framework should therefore examine functional control, infrastructure dependency, interoperability, algorithmic governance, network effects, switching costs, vertical integration and ecosystem-wide leverage alongside conventional market shares.