Competition Law and Artificial Intelligence Market Concentration

1. Introduction

Artificial Intelligence (AI) market concentration refers to a situation in which a relatively small number of undertakings control a substantial portion of one or more economically important layers of the AI ecosystem.

AI markets can become concentrated at several levels:

semiconductor and accelerator chips;

cloud computing;

data centres and computing capacity;

foundation models;

training datasets;

AI development tools;

application programming interfaces (APIs);

AI assistants;

enterprise AI;

AI-agent marketplaces; and

distribution platforms.

The competition-law problem is not concentration by itself. Large market share or concentration is not automatically unlawful. Competition law generally becomes concerned when market power is acquired, maintained, or exercised through conduct that restricts competition, or when mergers substantially reduce competitive constraints.

AI markets are particularly important because technological scale, enormous computing requirements, data advantages, network effects, switching costs and vertical integration can reinforce each other.

2. Structure of the AI Market

The AI industry can be conceptualised as a vertically connected chain:

Semiconductors → Cloud/Compute → Data → Foundation Models → APIs → Applications → Distribution

A single undertaking may operate at several levels.

For example, an integrated AI company could simultaneously:

develop a foundation model;

operate cloud infrastructure;

provide developer APIs;

operate an AI application;

distribute AI applications through a platform; and

acquire AI startups.

This vertical integration can produce efficiencies, but it can also create opportunities for foreclosure and leveraging.

3. What Is AI Market Concentration?

Concentration may arise where:

Horizontal concentration

A small number of firms provide most AI models or services.

Vertical concentration

One undertaking controls several successive levels of the AI supply chain.

Data concentration

One firm possesses an unusually large or commercially valuable dataset.

Compute concentration

Access to advanced GPUs, specialised accelerators or cloud infrastructure is concentrated among a small number of suppliers.

Distribution concentration

A small number of platforms control access to AI applications and users.

Ecosystem concentration

A company becomes central to several interconnected AI markets simultaneously.

4. Why AI Markets Can Become Concentrated

A. High Fixed Costs

Training sophisticated models may require enormous expenditure on:

computing infrastructure;

specialised processors;

data;

research personnel;

energy;

model training;

safety testing; and

deployment infrastructure.

High fixed costs can create substantial economies of scale.

B. Economies of Scale

Once a model has been trained, serving additional users may involve comparatively lower marginal costs.

This can produce:

more users → greater revenues → more computing resources → better models → more users.

Such economies can reinforce concentration.

5. Network Effects

AI platforms can experience direct and indirect network effects.

A developer platform becomes more valuable when more developers use it.

Similarly:

more users → more applications → more developers → more users.

AI-agent ecosystems could make these effects even stronger because users may prefer platforms offering a large variety of compatible agents.

6. Data Advantages

Large AI providers can obtain substantial amounts of:

user interaction data;

behavioural data;

feedback;

technical information;

application-performance data;

synthetic training data.

A large dataset does not automatically create dominance because data may be replicable or obtainable elsewhere.

However, where data are:

difficult to replicate;

continuously generated;

highly specialised; and

necessary for improving performance,

they can contribute to durable competitive advantages.

7. Computing Power as a Bottleneck

Advanced AI development depends heavily on computational resources.

Competition concerns may arise if access to critical computing resources is concentrated among a small number of providers.

Potential issues include:

discriminatory cloud access;

preferential allocation to affiliated AI companies;

long-term exclusive arrangements;

discriminatory pricing;

refusal to supply;

tying cloud services to AI products.

8. Foundation-Model Concentration

Foundation models may become an important layer of AI competition.

If only a small number of companies operate economically significant foundation models, downstream application developers may become dependent upon those models.

This creates a possible structure:

Foundation-model provider → API → Application developer → End user

The foundation-model provider may therefore possess bargaining power over downstream firms.

9. Vertical Integration

Vertical integration can generate legitimate efficiencies.

For example, a company controlling:

cloud infrastructure + AI model + application

may achieve:

lower costs;

faster innovation;

improved security;

better technical integration.

However, vertical integration can also create foreclosure possibilities.

A dominant cloud provider might theoretically:

favour its own AI model;

provide its affiliated model with cheaper computing;

restrict competitors' access;

use cloud data to improve its own model;

impose contractual restrictions on competing models.

The competition-law question is therefore whether integration produces efficiency-enhancing competition or exclusionary leverage.

10. Relevant Market Definition

Competition authorities must determine the relevant product and geographic markets.

Possible product markets include:

Market 1

AI foundation models.

Market 2

Generative-AI assistants.

Market 3

Enterprise AI services.

Market 4

AI cloud infrastructure.

Market 5

AI accelerator chips.

Market 6

AI developer APIs.

Market 7

AI-agent marketplaces.

The correct market cannot simply be assumed. It depends on substitutability, customer behaviour, technology and competitive constraints.

11. Zero-Price AI Services

Traditional competition analysis often focuses on price.

Many AI products, however, may be offered:

free of monetary charge.

That does not mean there is no competition.

Relevant competitive parameters may include:

privacy;

quality;

accuracy;

latency;

functionality;

interoperability;

advertising;

data collection;

innovation.

Consequently, AI competition analysis may need to examine non-price competition.

12. Market Share

Market share remains relevant but should not be treated as conclusive.

Authorities may consider:

revenue share;

number of users;

computing capacity;

API usage;

developer numbers;

model deployment;

cloud capacity;

access to data;

switching costs;

barriers to entry.

A company with substantial market share may nevertheless face effective competition.

Conversely, a company with a smaller current share could possess significant competitive importance if it represents a rapidly growing or technologically important alternative.

13. Barriers to Entry

AI markets may have substantial barriers to entry, including:

capital requirements;

access to advanced chips;

energy costs;

cloud infrastructure;

specialised talent;

training data;

intellectual property;

regulatory compliance;

distribution;

reputation;

computing capacity.

These barriers can contribute to persistent concentration.

14. Acquisition of AI Startups

One of the most significant competition concerns involves acquisitions by established technology companies.

A dominant firm might acquire a startup with:

a novel model architecture;

specialised training technology;

AI safety technology;

unique data;

a promising AI application;

a competing AI assistant.

The target may have little current revenue but significant future competitive potential.

This creates the familiar potential-competition and innovation-competition problem.

15. Killer Acquisition Concerns

A transaction can theoretically reduce competition even where the target is not currently a major competitor.

Competition authorities may therefore examine:

the target's technological capabilities;

future expansion plans;

pipeline products;

research capabilities;

potential competition;

alternative technologies.

The objective is to determine whether the acquisition removes an important competitive constraint.

16. Exclusive Arrangements

Concentration may be strengthened through exclusive agreements.

Examples could include:

exclusive cloud contracts;

exclusive AI distribution;

exclusive model licensing;

exclusive access to computing capacity;

exclusive arrangements with developers.

The competition analysis would consider:

duration;

market coverage;

market power;

foreclosure;

efficiencies;

availability of alternatives.

17. Self-Preferencing

An integrated AI platform might operate its own AI services while hosting competing applications.

It could potentially:

rank its own AI tools more prominently;

give them privileged access;

reduce competitors' visibility;

provide its own products with superior technical access.

This creates potential self-preferencing concerns.

The issue is especially significant when the platform controls a major distribution channel.

18. Tying and Bundling

A concentrated AI provider could condition access to one service on the purchase of another.

For example:

dominant cloud service + mandatory use of affiliated AI model.

Or:

AI operating system + mandatory AI assistant.

Competition authorities would examine whether the products are distinct and whether the arrangement forecloses competing providers.

19. Refusal to Supply

A dominant undertaking may control infrastructure that competitors require.

Examples include:

cloud computing;

AI APIs;

model access;

technical interfaces;

essential datasets.

A refusal to provide access can raise competition concerns in appropriate circumstances, although competition law does not ordinarily impose a general obligation on firms to deal with competitors.

The Bronner and IMS Health doctrines are particularly relevant here.

20. AI and Abuse of Dominance

Where a company has a dominant position, potentially abusive practices could include:

Exclusionary practices

predatory pricing;

exclusive dealing;

discriminatory access;

refusal of interoperability;

tying;

self-preferencing.

Exploitative practices

excessive prices;

unfair contractual conditions;

unreasonable data requirements.

Leveraging

Using dominance in one market to obtain or protect dominance in another.

21. AI Market Concentration and Article 102 TFEU

Under EU competition law, Article 102 TFEU prohibits the abuse of a dominant position where the relevant legal conditions are satisfied.

Potential AI scenarios include:

Cloud dominance → foreclosure of rival AI models

AI-platform dominance → discrimination against competing applications

Foundation-model dominance → exclusion of downstream developers

AI marketplace dominance → self-preferencing

Data dominance → discriminatory access

The underlying competition-law principles remain applicable even though the technology is new.

22. Article 101 TFEU and AI Concentration

Article 101 TFEU concerns agreements and concerted practices between undertakings.

AI concentration can increase coordination risks where competing companies:

share commercially sensitive information;

coordinate pricing;

use common algorithms;

coordinate capacity;

divide customers or markets.

Algorithmic coordination does not automatically constitute an infringement. The legal analysis depends on the nature of the communication, agreement or concerted practice and its competitive effects.

23. Algorithmic Collusion

A highly concentrated AI market could facilitate algorithmic coordination.

Suppose several competing AI providers deploy pricing algorithms capable of:

observing competitors' prices;

predicting their responses;

adjusting prices rapidly.

This could potentially make coordinated outcomes easier to sustain.

Competition authorities therefore need to distinguish between:

independent algorithmic optimisation

and

coordination attributable to independent undertakings.

24. Competition in Innovation

AI competition is frequently innovation competition rather than price competition.

Firms compete over:

model quality;

reasoning;

multimodality;

safety;

speed;

accuracy;

cost efficiency;

autonomy;

privacy;

interoperability.

A merger could therefore reduce competition even if current prices remain unchanged.

25. Case Law

1. United Brands v Commission

Case 27/76, United Brands Company and United Brands Continentaal BV v Commission

The Court of Justice explained the concept of dominance and emphasised the ability of an undertaking to behave independently of competitors and customers.

Importance for AI

The principle can be applied when assessing whether a major AI provider possesses sufficient economic power to act independently of effective competitive constraints.

Relevant factors could include:

market share;

barriers to entry;

infrastructure;

data;

customer dependence.

26. Hoffmann-La Roche v Commission

Case 85/76, Hoffmann-La Roche & Co AG v Commission

The Court developed the foundational EU concept of a dominant position.

AI relevance

AI market concentration should not be assessed exclusively through market share.

The analysis can include:

economic strength;

entry barriers;

customer dependence;

technological advantages;

competitive constraints.

27. Microsoft v Commission

Case T-201/04, Microsoft Corp v Commission

The General Court considered Microsoft's conduct concerning interoperability and tying.

AI relevance

This is particularly significant for AI ecosystems.

An AI platform could potentially use control over one technological layer to restrict competing products in another.

Examples include:

operating system → AI assistant;

cloud → AI model;

AI marketplace → AI applications.

The Microsoft judgment demonstrates the importance of interoperability and technological integration in competition analysis.

28. Google Shopping

Case T-612/17, Google and Alphabet v Commission

The case concerned Google's treatment of its own comparison-shopping service within its general search results.

AI relevance

The case provides an important precedent for analysing potential self-preferencing.

A dominant AI platform operating a marketplace could potentially favour its own AI applications or agents over competing services.

29. Google Android

Case T-604/18, Google LLC and Alphabet Inc. v Commission

The General Court examined several contractual restrictions surrounding the Android ecosystem.

AI relevance

The case demonstrates how competition law can scrutinise contractual arrangements that affect the ability of rival services to obtain access to users and distribution.

This could be relevant where AI services become embedded within operating systems, cloud platforms or application ecosystems.

30. Intel v Commission

Case C-413/14 P, Intel Corporation Inc. v Commission

The Court of Justice examined the assessment of exclusionary effects associated with rebates granted by a dominant undertaking.

AI relevance

A dominant AI infrastructure provider might offer:

preferential cloud pricing;

computing rebates;

volume discounts;

capacity incentives

to customers that use its own ecosystem.

The Intel judgment is relevant to the need for an appropriate effects-based assessment in the circumstances covered by the case.

31. Bronner

Case C-7/97, Oscar Bronner GmbH & Co. KG v Mediaprint

The case established important principles concerning refusal of access to infrastructure.

AI relevance

Suppose a dominant AI infrastructure provider controls a facility that competitors argue is indispensable.

Bronner provides an important framework for analysing when refusal of access can constitute abuse.

32. IMS Health

Case C-418/01, IMS Health GmbH & Co. OHG v NDC Health GmbH & Co. KG

The Court considered exceptional circumstances in which refusal to license intellectual-property-related material could amount to abuse.

AI relevance

AI companies may control:

proprietary datasets;

model architectures;

APIs;

technical interfaces;

specialised intellectual property.

IMS Health is therefore relevant to the difficult relationship between intellectual-property rights and competition law.

33. Slovak Telekom

Joined Cases C-152/19 P and C-165/19 P, Deutsche Telekom AG and Slovak Telekom v Commission

The case concerned exclusionary conduct associated with access to telecommunications infrastructure.

AI relevance

The case provides useful guidance for analysing situations in which competitors depend upon infrastructure controlled by a dominant undertaking.

The analogy may become important where AI competitors depend upon dominant cloud or computational infrastructure.

34. Servizio Elettrico Nazionale

Case C-377/20, Servizio Elettrico Nazionale SpA v AGCM

The Court examined exclusionary conduct by a dominant undertaking and the importance of distinguishing competition on the merits from conduct capable of foreclosing competitors.

AI relevance

The case is useful for analysing whether a dominant AI company is succeeding because of:

better technology and efficiency

or because it is using existing market power to exclude competitors.

35. AI Concentration and Merger Control

Merger control may become as important as abuse-of-dominance law.

Authorities can examine:

Horizontal mergers

Two AI developers combine.

Vertical mergers

A cloud provider acquires an AI-model company.

Conglomerate mergers

A large technology ecosystem acquires an AI application.

Potential-competition acquisitions

A dominant AI company acquires a small but technologically significant startup.

36. Competitive Effects of AI Mergers

Authorities may consider whether a transaction could produce:

higher prices;

reduced quality;

reduced innovation;

reduced choice;

increased data concentration;

reduced interoperability;

increased dependence on one ecosystem;

foreclosure of rivals.

The analysis should also account for legitimate efficiencies.

37. Remedies

Where competition concerns are established, possible remedies could include:

Behavioural remedies

non-discrimination;

interoperability;

access obligations;

restrictions on exclusivity;

transparency requirements.

Structural remedies

divestiture;

separation of business units;

restrictions on vertical integration.

Merger remedies

divestment;

licensing;

access commitments;

interoperability commitments.

The appropriate remedy depends upon the specific competitive harm established.

38. Competition Between Open and Closed AI Models

Another important issue is the relationship between:

open-source/open-weight AI

and

proprietary AI models.

Open models can potentially reduce entry barriers because developers can build on existing technology.

However, proprietary models may provide:

greater security controls;

specialised performance;

technical support;

integrated infrastructure.

Competition law should therefore focus on actual competitive effects rather than assuming that either model is inherently preferable.

39. Small AI Firms and Entry

Concentration analysis should also consider whether smaller firms can realistically enter.

Important questions include:

Can startups obtain sufficient computing capacity?

Are chips available on reasonable terms?

Can startups access cloud infrastructure?

Can models obtain sufficient training data?

Can startups reach users?

Can users switch between models?

Are APIs interoperable?

If these conditions deteriorate, concentration may become more durable.

40. Economic Effects of AI Concentration

Potential positive effects include:

economies of scale;

lower computing costs;

faster innovation;

greater research investment;

improved reliability;

enhanced safety;

integrated products.

Potential competitive risks may include:

exclusion of competitors;

reduced innovation;

higher prices;

reduced choice;

increased switching costs;

excessive data concentration;

dependency on dominant infrastructure.

These effects must be assessed on the evidence of the particular market.

41. Key Legal Tests

An AI concentration investigation may therefore proceed through several questions:

Step 1

What is the relevant market?

Step 2

What is the firm's position in that market?

Step 3

What barriers prevent entry or expansion?

Step 4

Are there effective competitors?

Step 5

Is the undertaking dominant?

Step 6

What conduct is being investigated?

Step 7

Does the conduct have exclusionary or exploitative effects?

Step 8

Are there objective justifications or efficiencies?

Step 9

Would intervention improve competitive conditions?

Step 10

What remedy is proportionate to the established harm?

42. Conclusion

Artificial intelligence market concentration presents competition law with a technologically complex version of familiar antitrust problems. The central concern is not simply that a few companies become large. Concentration becomes legally significant where market power produces or reinforces barriers to entry, enables exclusionary conduct, facilitates leveraging across vertically connected markets, or substantially reduces competitive constraints.

The most important competition issues include:

concentration of computing infrastructure;

foundation-model concentration;

data advantages;

network effects;

vertical integration;

exclusive cloud and AI arrangements;

self-preferencing;

interoperability restrictions;

refusal of access;

AI startup acquisitions;

potential-competition concerns;

algorithmic coordination; and

reduction of innovation competition.

The cases of United Brands, Hoffmann-La Roche, Microsoft, Google Shopping, Google Android, Intel, Bronner, IMS Health, Slovak Telekom and Servizio Elettrico Nazionale provide substantial doctrinal foundations for analysing these emerging problems. Although those cases did not concern today's AI ecosystem in its current form, their principles concerning dominance, interoperability, tying, exclusionary effects, infrastructure access, rebates, self-preferencing and leveraging can be applied to AI markets where the factual and legal conditions are comparable.