Competition Policy in AI Foundation Models

1. Introduction

AI foundation models (FMs) are general-purpose AI systems trained on very large datasets and capable of supporting multiple downstream applications, including chatbots, coding assistants, search, image generation, enterprise software, autonomous systems and decision-support tools. Examples include large language models and multimodal models.

Competition policy in foundation models is unusual because market power can arise simultaneously at several vertically connected layers:

1. Compute and semiconductor infrastructure
2. Cloud infrastructure
3. Training data and datasets
4. Foundation-model development
5. Model access/API markets
6. Fine-tuning and deployment infrastructure
7. Downstream applications
8. Distribution through operating systems, search engines, app stores and enterprise software

The UK CMA has identified competition risks in foundation-model markets and developed principles aimed at maintaining fair, open and effective competition. The US FTC, DOJ, European Commission and UK CMA have also jointly identified competition concerns in generative AI foundation models and AI products.

A central policy question is therefore:

How can competition law prevent incumbent firms from using control over compute, cloud, data, models or distribution to foreclose competing foundation-model developers without discouraging investment and innovation?

2. Economic Structure of Foundation-Model Markets

A simplified AI foundation-model value chain is:

Chips → Compute → Cloud → Data → Training → Foundation Model → API/Model Access → Applications → Distribution → Users

The same company may operate at several levels.

For example, a vertically integrated technology firm may simultaneously possess:

• cloud infrastructure;
• data centres;
• AI accelerators;
• a foundation model;
• an operating system;
• a search engine;
• an app store;
• enterprise software;
• advertising infrastructure.

The European Commission has specifically identified vertical integration and partnerships involving cloud providers and AI developers as important competition issues.

This creates the possibility of leveraging:

power at one AI layer → into another AI layer → to restrict competitors.

3. Main Competition-Policy Problems

A. Concentration of Compute

Training frontier foundation models requires enormous computational resources.

If access to advanced GPUs, AI accelerators or high-performance cloud infrastructure becomes concentrated, smaller AI developers may face:

• higher costs;
• capacity shortages;
• discriminatory access;
• contractual restrictions;
• long-term reservation requirements;
• technical switching costs.

This creates a possible essential-input problem.

However, competition authorities must distinguish between:

• legitimate economies of scale; and
• exclusionary control over an indispensable input.

The FTC's investigation of major AI partnerships specifically examined access to computing resources and other AI inputs.

4. Cloud–Foundation Model Vertical Integration

This is one of the most important competition-policy issues.

Suppose:

Cloud Provider A → invests in Foundation Model A → distributes Model A → competes with independent Model B.

The cloud provider may have incentives to:

• provide better compute availability to its affiliated model;
• impose higher switching costs;
• offer preferential pricing;
• bundle cloud services with AI models;
• restrict access to competitors;
• obtain commercially sensitive information from AI customers.

The FTC's study of Microsoft–OpenAI, Amazon–Anthropic and Alphabet–Anthropic partnerships identified potential concerns involving compute access, switching costs and access to sensitive technical/business information.

Competition-policy response

Possible tools include:

• merger control;
• abuse-of-dominance rules;
• vertical foreclosure analysis;
• interoperability obligations;
• non-discrimination requirements;
• monitoring of exclusivity clauses;
• switching and portability measures.

5. AI Partnerships and Minority Investments

Traditional merger control can become difficult where AI firms use:

• minority investments;
• strategic partnerships;
• convertible arrangements;
• exclusive cloud agreements;
• long-term capacity commitments;
• revenue-sharing arrangements.

These transactions may fall short of conventional acquisition structures while nevertheless producing significant competitive effects.

The FTC therefore used its Section 6(b) authority to investigate major AI partnerships and investments rather than treating the issue exclusively as conventional merger control.

Policy significance

Competition authorities increasingly need to examine economic control rather than merely legal ownership.

Relevant questions include:

• Who controls compute?
• Who controls model distribution?
• Who receives sensitive information?
• Who controls technical development?
• Is the AI developer free to switch cloud providers?
• Does an investor obtain veto or consultation rights?
• Does an agreement restrict partnerships with rival clouds?

6. The Microsoft–OpenAI Example

The Microsoft–OpenAI relationship illustrates the difficulty of applying traditional merger principles to AI partnerships.

The UK CMA investigated whether the relationship constituted a relevant merger situation and whether it could substantially lessen competition. In March 2025, the CMA concluded that the partnership did not qualify for investigation under the UK merger provisions.

This is important because it demonstrates that:

A strategically significant AI partnership does not automatically constitute a legally reviewable merger.

The broader competition question nevertheless remains relevant under other legal theories, including:

• vertical foreclosure;
• exclusive dealing;
• information exchange;
• leveraging;
• cloud lock-in;
• access discrimination.

7. Data as a Competitive Input

Foundation models depend heavily upon training data.

Competition problems may arise where a company controls:

• proprietary datasets;
• large-scale user-generated data;
• search-query data;
• transaction data;
• enterprise data;
• social-network data;
• specialised scientific datasets.

A dominant platform may possess a data advantage that competitors cannot reproduce economically.

However, possession of valuable data does not automatically constitute an antitrust violation.

The competition inquiry should examine:

1. Is the data genuinely difficult to replicate?
2. Is it competitively significant?
3. Does the firm possess market power?
4. Is access being strategically denied?
5. Does the data advantage reinforce another bottleneck?
6. Would access obligations reduce incentives to invest?

8. Data Combination and Cross-Platform Advantages

AI creates a particularly important issue because data can be combined across ecosystems.

For example:

Search data + consumer behaviour + cloud data + productivity data + advertising data + AI interactions

may generate a substantial informational advantage.

Competition authorities therefore need to consider whether a dominant firm is using data accumulated in one market to strengthen its position in an emerging AI market.

This is closely related to the European competition-law cases concerning data leverage and digital ecosystems.

9. Self-Preferencing by Foundation Models

A platform controlling a foundation model may favour its own downstream applications.

For example:

Foundation Model → own chatbot → own search → own productivity suite

could be structured so that competing applications receive:

• slower API access;
• inferior functionality;
• less favourable pricing;
• reduced interoperability;
• inferior distribution;
• restricted system permissions.

The relevant competition-law concept is self-preferencing or discriminatory access.

The Google Shopping jurisprudence is particularly relevant by analogy because it demonstrates how dominance at one digital layer can interact with preferential treatment of an affiliated service.

10. API Gatekeeping

Foundation-model APIs can function as gateways.

A model provider may control:

• API access;
• pricing;
• latency;
• rate limits;
• technical documentation;
• safety restrictions;
• model availability;
• access to advanced versions.

If an API becomes commercially indispensable, discriminatory API access could become an important competition concern.

The analysis should consider whether restrictions are:

• technically necessary;
• security-related;
• genuinely safety-related;
• commercially justified; or
• designed to disadvantage rivals.

11. Interoperability

Interoperability may become increasingly important.

A dominant AI ecosystem could restrict competitors from interacting with:

• operating systems;
• productivity applications;
• smartphones;
• cloud environments;
• enterprise databases;
• search infrastructure;
• identity systems.

The European Commission has recently taken measures concerning AI interoperability on Android and access to Google Search data, specifically to facilitate competition involving rival AI services.

This illustrates a broader policy movement from purely ex-post antitrust enforcement toward ex-ante contestability obligations.

12. Switching Costs and Lock-In

AI developers can become dependent upon a particular cloud/model ecosystem because of:

• proprietary APIs;
• model-specific fine-tuning;
• specialised infrastructure;
• data formats;
• inference optimisation;
• enterprise contracts;
• technical integration;
• accumulated prompts and workflows.

Switching may therefore be technically possible but economically difficult.

This creates functional lock-in.

The FTC specifically identified increased switching costs as a potential competitive consequence of major AI-cloud partnerships.

Competition policy can respond through:

• data portability;
• API portability;
• interoperability;
• contractual restrictions on lock-in;
• cloud switching facilitation;
• transparency concerning exit costs.

13. Bundling and Tying

Foundation models may be bundled with:

• cloud computing;
• office software;
• operating systems;
• search;
• advertising;
• cybersecurity;
• enterprise subscriptions.

For example:

Cloud subscription + proprietary AI model + enterprise software

could make independent AI suppliers less competitive.

The traditional tying framework therefore remains relevant.

Authorities would need to establish:

1. separate products/services;
2. market power in the tying product;
3. coercion or effective conditioning;
4. competitive foreclosure;
5. absence of sufficient objective justification.

14. Predatory Pricing and AI Subsidisation

Foundation-model providers may offer models:

• free of charge;
• below incremental cost;
• through subsidised APIs;
• bundled with unrelated products.

Low prices are not inherently anticompetitive.

But competition authorities may investigate whether a dominant firm uses profits from another market to subsidise AI services with the purpose or effect of eliminating competitors.

The challenge is that AI markets naturally exhibit:

• high fixed costs;
• low marginal inference costs;
• rapid technological improvement;
• network effects.

Therefore, traditional price-cost tests may need careful adaptation.

15. Network Effects and Ecosystem Tipping

Foundation models can benefit from feedback loops:

More users → more usage data → better product → more developers → more applications → more users

This can produce tipping.

Developers may also cluster around the model with:

• the largest user base;
• most tools;
• best documentation;
• most integrations.

Once a model ecosystem reaches a critical scale, competitors may struggle to enter even if their technology is viable.

Competition policy should therefore examine dynamic competition, not merely present market shares.

16. Killer Acquisitions

A major foundation-model company could acquire:

• promising model developers;
• AI safety companies;
• specialised data providers;
• model-serving platforms;
• AI infrastructure startups;
• fine-tuning companies;
• AI application companies.

Traditional merger thresholds may fail to capture acquisitions of startups whose present revenues are low but whose future competitive significance is substantial.

Relevant considerations include:

• innovation pipelines;
• technological capabilities;
• access to specialised talent;
• alternative foundation-model architectures;
• potential future competition.

17. Talent Acquisition as a Competition Issue

AI markets are unusually dependent upon specialised talent.

A large company may hire most of a smaller AI firm's:

• researchers;
• engineers;
• safety specialists;
• technical leadership.

Even without formally acquiring the company, this may substantially diminish an emerging competitor.

Competition authorities should distinguish legitimate labour mobility from arrangements designed to eliminate a competitor.

18. Algorithmic Collusion

Foundation models could facilitate coordination among competing firms.

Potential mechanisms include:

• AI pricing agents;
• common optimisation systems;
• shared APIs;
• automated competitor monitoring;
• common data suppliers;
• algorithmically generated prices.

The fundamental competition-law question is whether coordination can occur without traditional human communication.

Traditional cartel law generally focuses on agreement or concerted practice. AI creates difficult questions concerning:

• algorithmic communication;
• autonomous agents;
• shared optimisation objectives;
• hub-and-spoke coordination;
• common software providers.

The 2024 international competition statement expressly recognised risks arising from AI and emphasised preserving competitive markets and innovation.

19. Essential-Facility Theory and AI

Foundation-model infrastructure raises an important question:

Can a particular AI model, compute resource, dataset or API become an essential facility?

Traditional essential-facility doctrine generally requires stringent conditions.

A useful analytical framework is:

Step 1 — Indispensability

Can competitors reasonably obtain the input elsewhere?

Step 2 — Replicability

Can the infrastructure be economically reproduced?

Step 3 — Elimination of competition

Would denial substantially eliminate competition?

Step 4 — Objective justification

Is there a legitimate technical, security or safety reason for refusal?

Step 5 — Proportional remedy

Would access obligations be feasible without destroying investment incentives?

The Bronner and IMS Health decisions are especially important for understanding the limits of mandatory access.

20. Six Key Case Laws

There is currently no mature body of reported judicial precedent specifically deciding the legality of foundation-model conduct. Consequently, competition policy for AI foundation models relies substantially on established digital-platform, essential-facilities, vertical-foreclosure and data cases.

Case 1 — Microsoft Corp. v Commission, Case T-201/04

The EU General Court upheld major elements of the Commission's finding that Microsoft had abused its dominant position through practices involving interoperability information and tying.

Relevance to foundation models

The case demonstrates how control over a technologically important platform can create competitive concerns where interoperability is restricted.

It provides an analytical precedent for:

• AI-platform interoperability;
• API access;
• model integration;
• tying AI services to operating systems;
• technological foreclosure.

21. Case 2 — Google Shopping, Case AT.39740 / Google LLC v Commission

The European Commission found that Google had abused dominance by giving preferential positioning to its own comparison-shopping service.

Relevance

The case is important for self-preferencing.

An analogous AI scenario would involve a dominant platform giving its own foundation-model application:

• privileged placement;
• preferential access;
• superior system integration;
• better data;
• better visibility.

The legal question would remain fact-specific: mere vertical integration would not itself establish abuse.

22. Case 3 — Google Android, Case AT.40099

The Commission found several practices concerning Google's Android ecosystem to be abusive, including tying and restrictions affecting competing services.

Foundation-model relevance

The case demonstrates the importance of ecosystem control.

AI competition increasingly depends on:

OS → default assistant → search → model → application

If a dominant operating-system provider gives its own foundation model privileged access while restricting rival AI systems, Android provides a useful competition-law analytical precedent.

23. Case 4 — Bronner v Mediaprint, C-7/97

The Court of Justice established a demanding test for compulsory access to infrastructure under Article 102 TFEU.

Relevance to AI

This case is highly relevant to claims that:

• compute infrastructure;
• cloud infrastructure;
• AI APIs;
• datasets;
• model infrastructure

should be made available to competitors.

The principle is important because competition law should not automatically transform every commercially valuable input into a mandatory-access facility.

24. Case 5 — IMS Health v Commission, Joined Cases C-418/01 P and C-7/97

The Court of Justice considered when refusal to license an intellectual-property-related resource could constitute abuse of dominance.

Relevance to foundation models

The case is relevant to:

• model weights;
• proprietary AI interfaces;
• specialised datasets;
• technical interfaces;
• AI-related intellectual property.

The case illustrates the exceptional nature of compulsory licensing and the need to consider whether refusal would eliminate competition in a distinct downstream market.

25. Case 6 — Slovak Telekom v Commission, Joined Cases C-152/19 P and C-165/19 P

The Court of Justice considered exclusionary conduct concerning access to telecommunications infrastructure.

Relevance

The case is useful for analysing:

• infrastructure access;
• technical foreclosure;
• discriminatory access;
• network bottlenecks.

AI infrastructure may similarly develop bottlenecks where a small number of firms control critical infrastructure used by competing AI developers.

26. Case 7 — Qualcomm, Case AT.39711

The European Commission's Qualcomm proceedings concerned exclusionary practices and pricing arrangements in the chipset sector.

AI relevance

The case provides useful precedent for examining upstream technology inputs.

AI foundation models depend on upstream infrastructure such as:

• GPUs;
• AI accelerators;
• networking equipment;
• memory;
• cloud compute.

Where an upstream supplier possesses substantial market power, competition authorities may need to assess whether contractual arrangements foreclose downstream AI competitors.

27. Case 8 — Meta Platforms Inc. v Bundeskartellamt, Case C-252/21

The Court of Justice addressed the relationship between competition law and the processing of personal data in the context of Meta's business model.

Relevance to foundation models

This is particularly significant because AI models rely upon data.

The case demonstrates that:

competition analysis can intersect with data-protection considerations.

For AI, this becomes relevant where dominant firms combine data across:

• social platforms;
• search;
• cloud;
• productivity software;
• advertising;
• AI services.

Competition authorities may therefore have to examine whether data practices reinforce market power while respecting the separate legal framework governing personal data.

28. Lessons From the Case Law

The cases collectively establish several important principles.

These cases do not mean that foundation models automatically satisfy the legal tests developed in them. They provide analytical frameworks that can be adapted to AI markets.

29. Competition Policy Model for Foundation Models

A comprehensive competition-policy framework should examine six dimensions.

A. Access

Are competitors able to obtain:

• compute;
• cloud services;
• datasets;
• APIs;
• specialised chips;
• model infrastructure?

B. Choice

Can developers switch between:

• clouds;
• foundation models;
• APIs;
• deployment environments?

C. Neutrality

Does an integrated platform treat:

• its own AI model; and
• competing AI models

on equivalent terms?

D. Transparency

Are businesses able to understand:

• pricing;
• API restrictions;
• access criteria;
• switching costs;
• contractual commitments?

E. Contestability

Can a new foundation-model provider realistically enter and expand?

F. Innovation

Will intervention preserve incentives to invest in:

• compute;
• research;
• model development;
• safety;
• data acquisition;
• infrastructure?

30. Ex-Ante Regulation Versus Ex-Post Antitrust

Traditional competition law is generally ex post:

conduct occurs → investigation → infringement finding → remedy.

AI markets may require some ex-ante measures because tipping can occur rapidly.

Possible ex-ante mechanisms include:

• interoperability;
• switching rights;
• data portability;
• non-discrimination;
• transparency;
• restrictions on certain exclusivity arrangements;
• merger notification;
• access obligations for designated bottlenecks.

The EU's current digital-policy approach increasingly treats AI and cloud infrastructure as important areas for maintaining contestability.

31. Competition Policy and AI Safety

A major complication is that competition and safety can point in different directions.

For example, a model provider may restrict API access because of:

• cybersecurity;
• misuse;
• privacy;
• model safety;
• fraud prevention.

A competition authority should therefore avoid assuming that every access restriction is exclusionary.

The proper inquiry is:

Is the safety justification genuine, evidence-based and proportionate, or is safety being used as a pretext for exclusion?

This distinction will become increasingly important.

32. National-Security and Industrial-Policy Considerations

Foundation models also have strategic significance.

Governments may support domestic AI companies through:

• subsidies;
• government procurement;
• research grants;
• public compute;
• semiconductor policies;
• sovereign cloud projects.

Competition policy must distinguish between legitimate public investment and state support that substantially distorts competition.

A competition-neutral public AI infrastructure may actually lower entry barriers if it is accessible on transparent and non-discriminatory terms.

33. Public Compute as Competition Infrastructure

Governments could create shared AI infrastructure providing:

• compute capacity;
• research datasets;
• testing facilities;
• model evaluation;
• cloud credits.

The competition-policy design should ensure:

• open eligibility;
• transparent allocation;
• non-discriminatory access;
• objective pricing;
• independent governance.

Otherwise, public infrastructure could itself become a mechanism for favouring selected AI firms.

34. Foundation Models as a New Type of Bottleneck

Traditional antitrust often analyses bottlenecks such as:

• telecommunications networks;
• operating systems;
• payment systems;
• search engines;
• app stores.

Foundation models may become a new category:

algorithmic infrastructure bottleneck.

A dominant foundation model could potentially become indispensable because thousands of downstream applications rely on it.

But the crucial legal distinction remains:

large or successful model ≠ automatically dominant model.

Market definition, substitutability, countervailing power, entry conditions and actual competitive effects remain essential.

35. Market Definition

Potential relevant markets might include:

• general-purpose foundation models;
• specialised foundation models;
• text-generation models;
• multimodal models;
• AI inference services;
• AI APIs;
• model hosting;
• AI cloud infrastructure;
• training compute;
• AI application distribution.

The SSNIP test may become difficult where services are free or subsidised.

Authorities may therefore need to consider:

• quality;
• latency;
• accuracy;
• privacy;
• safety;
• interoperability;
• switching costs;
• data advantages;
• innovation;
• developer ecosystem.

36. Dynamic Competition

Static market share is particularly unreliable in AI.

A company with a smaller market share today may possess:

• superior architecture;
• better efficiency;
• stronger research;
• better open-source technology;
• rapidly growing developer adoption.

Therefore, competition authorities should examine innovation competition.

Relevant questions include:

• Who is developing credible substitutes?
• Are incumbents acquiring emerging competitors?
• Are partnerships eliminating independent routes to market?
• Is access to compute preventing expansion?
• Are distribution bottlenecks preventing adoption?

37. Open-Source Foundation Models

Open-weight or open-source models can potentially reduce entry barriers by allowing developers to build without purchasing access from a dominant proprietary model provider.

They may promote:

• model diversity;
• experimentation;
• portability;
• local deployment;
• lower switching costs.

But open models also create different competition questions concerning:

• distribution;
• cloud hosting;
• downstream ecosystem control;
• safety;
• licensing restrictions;
• access to compute.

Competition policy should therefore remain technologically neutral rather than assuming that either open or closed models are inherently more competitive.

38. Overall Competition-Policy Framework

The emerging framework can be expressed as:

Foundation Model Competition

↓

Inputs

• Compute
• Chips
• Data
• Talent
• Capital

↓

Model Development

• Training
• Fine-tuning
• Model weights
• Safety systems

↓

Access

• API
• Cloud
• Licensing
• Deployment

↓

Distribution

• Search
• OS
• App stores
• Enterprise software

↓

Downstream Competition

• Applications
• Agents
• Productivity
• Coding
• Search
• Creative tools

At every stage, competition authorities should ask:

Is a firm winning because of superior innovation, or because control over one bottleneck is being used to exclude competitors at another level?

39. Conclusion

Competition policy for AI foundation models is developing at the intersection of antitrust, digital-platform regulation, data governance, intellectual property, cloud infrastructure and innovation policy.

The principal risks are:

1. concentration of compute;
2. cloud–model vertical integration;
3. exclusionary AI partnerships;
4. data advantages;
5. API gatekeeping;
6. interoperability restrictions;
7. self-preferencing;
8. ecosystem tipping;
9. switching costs;
10. killer acquisitions;
11. algorithmic coordination; and
12. control over AI infrastructure.

The most relevant existing jurisprudence comes from Microsoft, Google Shopping, Google Android, Bronner, IMS Health, Slovak Telekom, Qualcomm and Meta, rather than from a mature body of foundation-model judgments.

The emerging policy direction is therefore not to regulate AI simply because it is technologically powerful, but to preserve contestability, interoperability, access, innovation and freedom to switch while avoiding remedies that unnecessarily undermine investment or legitimate safety measures. The CMA has expressly identified risks to fair, open and effective competition in foundation-model markets, while the FTC's AI partnership study has highlighted compute access, switching costs and sensitive-information advantages as potential competition concerns.