Competition Law and Cognitive Infrastructure Market Power

1. Introduction

Cognitive infrastructure refers to the underlying technological infrastructure required to develop, train, deploy and operate advanced computational and artificial-intelligence systems. It includes:

• AI chips, GPUs, CPUs and accelerators;
• cloud-computing infrastructure;
• data centres and high-performance computing;
• networking and interconnection technologies;
• AI training and inference capacity;
• model-serving infrastructure;
• specialised datasets and computational resources;
• AI development platforms, APIs and software layers; and
• technical standards and interoperability interfaces.

The competition-law issue arises when control over these infrastructure layers gives a firm the ability to exclude competitors, raise their costs, restrict access, tie complementary products, discriminate between downstream users, or reinforce its position across adjacent AI markets.

This is increasingly important because cloud infrastructure and AI computing are becoming interconnected. The European Commission has described cloud computing as an important gateway and, in June 2026, preliminarily considered AWS and Microsoft Azure to possess sufficiently entrenched positions to warrant designation under the DMA despite not satisfying the ordinary quantitative thresholds.

2. Meaning of Cognitive Infrastructure Market Power

Market power traditionally means the ability of an undertaking to behave to a significant extent independently of competitive constraints.

In cognitive infrastructure, market power may arise from:

1. Control over scarce computational capacity
2. Economies of scale
3. Large capital requirements
4. Access to specialised AI chips
5. Data-centre network effects
6. Cloud switching costs
7. Interoperability barriers
8. Long-term contractual commitments
9. Vertical integration
10. Control over complementary AI technologies

Thus, a company need not possess a conventional monopoly over "AI" as a whole. It may have substantial power in a narrower relevant market such as:

AI accelerator chips → cloud GPU infrastructure → AI model training → model-serving infrastructure → AI applications.

A competition authority therefore has to determine where the relevant market boundaries lie.

3. Relevant Product Markets

Several distinct markets may potentially be identified.

A. AI Accelerator Market

This includes GPUs and specialised accelerators used for AI training and inference.

Relevant competitive concerns include:

• exclusive supply arrangements;
• discriminatory allocation;
• refusal to supply;
• technological interoperability;
• software ecosystem restrictions;
• bundling chips with networking equipment.

B. Cloud AI Infrastructure

Cloud providers can supply:

• virtual machines;
• GPU clusters;
• storage;
• networking;
• AI development tools;
• model-serving facilities.

Cloud markets may exhibit substantial switching costs and customer lock-in.

The European Commission's current cloud investigation specifically examines issues such as interoperability, access to business-user data, tying and bundling, and contractual conditions.

C. AI Data-Centre Infrastructure

Data centres provide:

• computational capacity;
• electricity;
• cooling;
• networking;
• storage;
• specialised AI hardware.

Control over strategically located or technologically specialised facilities can potentially constitute infrastructure market power.

D. AI Development Platforms

A cloud provider may simultaneously supply:

computing + storage + AI models + APIs + developer tools.

This creates opportunities for leveraging market power between vertically related markets.

4. Sources of Market Power

4.1 Economies of Scale

AI infrastructure requires enormous capital investment.

Large providers can spread:

• data-centre costs;
• electricity costs;
• networking costs;
• chip procurement costs;
• engineering expenses

over enormous customer bases.

This may make entry difficult for smaller firms.

4.2 High Switching Costs

A customer migrating from one cloud provider to another may need to change:

• APIs;
• storage systems;
• databases;
• machine-learning pipelines;
• security architecture;
• identity systems;
• model deployment tools.

Consequently, even where several providers formally exist, contestability may be substantially weaker than the number of providers suggests.

The European Commission has specifically identified lock-in effects and high switching costs as relevant features of the AWS and Azure competitive position.

4.3 Network Effects

The larger the infrastructure ecosystem becomes, the more attractive it may become to:

• developers;
• AI startups;
• enterprise customers;
• chip designers;
• software providers.

This can generate a self-reinforcing cycle:

More customers → more developers → more applications → greater ecosystem value → more customers.

5. Vertical Integration

A particularly important issue is vertical integration.

Consider:

AI chip → server → cloud → AI model → AI application

A firm controlling several levels may possess the ability and incentive to favour its own downstream products.

For example:

dominant cloud infrastructure + proprietary AI accelerator + proprietary AI model

could create incentives to make rival AI developers less attractive on the same infrastructure.

The concern is not vertical integration itself. Competition law generally permits vertical integration. The issue is whether the integration is used to foreclose rivals or exploit market power.

6. Important Competition-Law Theories

A. Refusal to Deal

A dominant infrastructure provider may potentially violate competition law if it unjustifiably refuses access to infrastructure that rivals cannot reasonably reproduce.

The difficult question is whether the infrastructure is genuinely indispensable.

B. Discriminatory Access

A provider may supply infrastructure to competitors on:

• worse prices;
• inferior technical conditions;
• slower access;
• reduced capacity;
• discriminatory service levels.

Where the provider also competes downstream, discrimination becomes particularly significant.

C. Tying and Bundling

A dominant infrastructure provider might condition access to computing capacity upon purchasing:

• its storage;
• its AI platform;
• its networking;
• its proprietary model;
• its cybersecurity tools.

Competition authorities would examine whether such arrangements foreclose competing suppliers.

D. Exclusive Dealing

Long-term agreements may reserve scarce AI capacity for one customer.

This can be commercially legitimate, but excessive exclusivity can potentially prevent rivals from obtaining sufficient computational resources.

E. Margin Squeeze

A vertically integrated provider might:

• charge competitors high wholesale infrastructure prices; while
• offering its own downstream services at prices that competitors cannot profitably match.

This can raise a margin-squeeze theory.

7. Six Major Case Laws

1. United States v. Terminal Railroad Association, 224 U.S. 383 (1912)

Principle

The U.S. Supreme Court dealt with control over essential railway terminal facilities.

The defendants controlled infrastructure necessary for competitors to reach the relevant market.

Competition-law significance

The case is an early foundation for the essential-facilities/access theory.

Applied to cognitive infrastructure, the analogy could arise where:

• a particular infrastructure facility is indispensable;
• competitors cannot reasonably duplicate it;
• access is necessary to compete; and
• the dominant owner unjustifiably excludes competitors.

Relevance to AI infrastructure

Potential examples could include:

• uniquely accessible computational facilities;
• critical interconnection infrastructure;
• indispensable technical interfaces.

However, modern competition law applies the essential-facilities doctrine cautiously.

2. Aspen Skiing Co. v. Aspen Highlands Skiing Corp., 472 U.S. 585 (1985)

Principle

The U.S. Supreme Court found antitrust significance in the termination of a previously profitable cooperative arrangement by a dominant firm.

Relevance

The case is important for refusal-to-deal analysis.

For cognitive infrastructure, questions may include:

• Did the infrastructure provider previously supply access?
• Was access commercially beneficial?
• Was access suddenly withdrawn?
• Is there an anticompetitive explanation?
• Does the refusal disadvantage an effective competitor?

The case does not establish a general obligation for dominant companies to cooperate with competitors.

3. Verizon Communications Inc. v. Law Offices of Curtis V. Trinko, 540 U.S. 398 (2004)

Principle

The U.S. Supreme Court substantially limited compulsory-dealing theories under U.S. antitrust law.

It emphasised that competition law generally does not require firms to assist competitors merely because they possess infrastructure.

Relevance to cognitive infrastructure

This creates an important counterweight to Terminal Railroad and Aspen Skiing.

A competitor seeking access to:

• GPU capacity;
• cloud infrastructure;
• networking;
• proprietary APIs

cannot automatically demand access merely because another company controls it.

The legal inquiry must satisfy the applicable doctrinal requirements.

4. Commercial Solvents v. Commission, Joined Cases 6/73 and 7/73 (1974)

Principle

The European Court of Justice addressed refusal to supply an essential input by a dominant undertaking that was also active in a downstream market.

Competition significance

The case illustrates the danger of a vertically integrated undertaking using control over an upstream input to disadvantage downstream competitors.

Application to cognitive infrastructure

Suppose a company controls a scarce AI computing input while competing with customers who use that input to develop competing AI services.

Competition authorities could examine:

Upstream infrastructure power + downstream competition + discriminatory/refused supply = potential foreclosure problem.

5. Bronner v. Mediaprint, Case C-7/97 (1998)

Principle

The ECJ established a demanding framework for treating infrastructure as an indispensable facility.

The facility must, broadly speaking, be indispensable and there must be no realistic alternative.

Importance for cognitive infrastructure

This is particularly relevant to AI infrastructure because companies may argue:

"Competitors can simply build their own infrastructure."

The legal question therefore becomes whether realistic alternatives exist.

Relevant factors can include:

• technical feasibility;
• economic feasibility;
• time required for construction;
• availability of substitute suppliers;
• scale of investment;
• geographic constraints.

6. IMS Health v. Commission, Case C-418/01 P (2004)

Principle

The ECJ considered refusal to license intellectual property under the exceptional circumstances associated with compulsory access.

The Court identified stringent conditions concerning indispensability, elimination of competition, and absence of objective justification.

Application to cognitive infrastructure

Modern AI infrastructure often combines:

• hardware;
• software;
• APIs;
• proprietary interfaces;
• datasets;
• intellectual property.

IMS Health demonstrates why competition authorities must distinguish between:

legitimate protection of intellectual property

and

use of intellectual property to eliminate competition in a related market.

8. Additional Important AI-Infrastructure Cases

7. FTC v. NVIDIA Corporation / Arm

The proposed NVIDIA–Arm transaction became a major example of competition scrutiny in AI-related infrastructure.

The FTC alleged that acquisition of Arm could allow NVIDIA to undermine competitors dependent upon Arm's processor technology. The transaction was ultimately abandoned.

The FTC had specifically identified concerns involving processors used by cloud-service providers and other markets.

Legal importance

It demonstrates how competition analysis can focus on:

• control of critical technological inputs;
• vertical foreclosure;
• innovation competition;
• future competition;
• access to processor technologies.

8. NVIDIA / Mellanox

The NVIDIA–Mellanox transaction involved AI-relevant computing and networking infrastructure.

The European Commission examined possible conglomerate effects but cleared the transaction unconditionally after concluding that the identified concerns would not materially harm competition. China's SAMR ultimately approved the transaction subject to conditions.

Significance

The transaction demonstrates that competition analysis of AI infrastructure may extend beyond horizontal market shares to:

chips + networking + data-centre architecture.

9. Microsoft–OpenAI Competition Inquiry

The Microsoft–OpenAI relationship illustrates a newer form of infrastructure competition issue.

The UK CMA investigated whether Microsoft's partnership with OpenAI qualified for merger investigation and ultimately concluded in March 2025 that it did not qualify under the relevant UK merger provisions.

Separately, the U.S. FTC's 6(b) study examined Microsoft–OpenAI, Amazon–Anthropic and Google–Anthropic partnerships.

The FTC identified possible concerns involving:

• access to computing resources;
• engineering talent;
• switching costs;
• contractual rights;
• sensitive technical information;
• control and exclusivity arrangements. 

This is important because cognitive infrastructure market power may arise through investment and contractual relationships, rather than traditional ownership alone.

10. Cloud Infrastructure as a Competition Bottleneck

Cloud computing is particularly important because it provides the computational foundation for AI.

A simplified structure is:

Data centres
↓
Cloud infrastructure
↓
GPU/AI accelerator access
↓
AI model training
↓
AI models/APIs
↓
AI applications

Control at an upstream level can therefore affect competition downstream.

The UK's CMA completed its cloud-services market investigation in 2025 and identified competition concerns sufficient to recommend consideration of strategic market-status investigations concerning Microsoft and AWS.

11. Interoperability as a Competition Remedy

Interoperability can reduce infrastructure market power.

Possible remedies include:

Data portability

Customers should be able to transfer:

• datasets;
• workloads;
• configurations;
• applications.

API interoperability

Rival services should be capable of communicating through standardised interfaces.

Cloud switching

Customers should not face unreasonable technical barriers when moving workloads.

Open technical standards

Standards can reduce dependency on a single infrastructure provider.

12. Cognitive Infrastructure and Ecosystem Lock-In

The most significant competition concern may not be a traditional monopoly.

Instead, market power may result from ecosystem control.

For example:

Cloud provider

• proprietary AI chips
• AI development platform
• proprietary model
• storage
• identity system
• enterprise software

can create a highly integrated ecosystem.

A competitor may technically be able to enter every individual market but still face substantial cumulative barriers.

This is sometimes described as ecosystem foreclosure.

13. Market Power Through Scarcity

AI infrastructure can generate temporary or structural scarcity.

For example:

Limited GPUs → limited training capacity → higher infrastructure costs → fewer potential entrants.

A dominant provider controlling scarce capacity could potentially:

• allocate capacity preferentially;
• enter exclusive agreements;
• discriminate between customers;
• reserve capacity for its own AI operations.

Competition authorities therefore need to distinguish legitimate capacity management from exclusionary conduct.

14. Data and Compute as Complementary Inputs

AI competition increasingly involves two major inputs:

Data

and

Compute

A firm possessing both can potentially achieve significant advantages.

The competitive structure may therefore resemble:

Data + Compute + Talent + Distribution + Capital

rather than simply:

Market share of AI models.

This makes traditional market-definition techniques more difficult.

15. Merger Control

Competition authorities may scrutinise mergers involving:

• AI-chip manufacturers;
• cloud providers;
• data-centre operators;
• networking companies;
• AI model developers;
• AI infrastructure software;
• cybersecurity providers.

The analysis should consider:

Horizontal effects

Will the merger eliminate an actual competitor?

Vertical effects

Could the merged firm restrict access to an upstream input?

Conglomerate effects

Could the firm bundle products across adjacent markets?

Innovation effects

Could the transaction eliminate future technological competition?

Ecosystem effects

Could it reinforce a closed technological ecosystem?

The NVIDIA–Arm and NVIDIA–Mellanox matters illustrate how AI infrastructure transactions can raise these issues.

16. Exclusive AI Compute Agreements

Long-term agreements for AI compute may create competition concerns when they involve:

• exclusive GPU allocation;
• minimum-purchase obligations;
• restrictions on multi-cloud deployment;
• preferential pricing;
• capacity reservations.

The economic question is not simply whether an agreement is exclusive.

Authorities may examine:

What proportion of available capacity is foreclosed?

and

Can competing AI firms obtain an economically viable alternative?

17. Self-Preferencing

A vertically integrated cloud provider may operate:

• cloud infrastructure;
• an AI model platform; and
• competing AI applications.

Possible self-preferencing could involve:

• better computing prices for internal models;
• preferential access to GPUs;
• privileged API access;
• superior technical integration;
• preferential search or marketplace placement.

Such conduct must be analysed under the applicable jurisdiction's abuse-of-dominance or digital-platform rules.

18. Competition Law Framework

United States

Relevant tools include:

• Sherman Act §1;
• Sherman Act §2;
• Clayton Act §7;
• FTC Act §5;
• merger review;
• monopolisation analysis;
• refusal-to-deal doctrine;
• tying and exclusive-dealing doctrines.

European Union

Relevant provisions include:

• Article 101 TFEU — anticompetitive agreements;
• Article 102 TFEU — abuse of dominance;
• EU Merger Regulation;
• Digital Markets Act;
• essential-facilities jurisprudence;
• interoperability and access remedies.

The EU's current cloud investigations show that these principles are being adapted to infrastructure markets involving AI.

United Kingdom

Relevant legislation includes:

• Competition Act 1998;
• Enterprise Act 2002;
• Digital Markets, Competition and Consumers Act 2024.

The CMA's cloud investigation provides a particularly important contemporary example of infrastructure-focused competition analysis.

19. Key Competition Risks

20. Defences and Legitimate Business Justifications

Not every infrastructure restriction violates competition law.

A provider may have legitimate reasons relating to:

• cybersecurity;
• capacity constraints;
• reliability;
• intellectual-property protection;
• network security;
• technical compatibility;
• fraud prevention;
• investment incentives;
• data protection;
• service quality.

Competition authorities therefore need to establish both:

1. market power, and
2. anticompetitive conduct or effects where required by the applicable legal rule.

21. Emerging Concept: Cognitive Infrastructure as an Essential Input

The traditional essential-facilities doctrine may become increasingly important as AI infrastructure becomes more specialised.

The legal question could evolve from:

"Is this railway or telecommunications facility indispensable?"

to:

"Is this computational infrastructure indispensable for effective competition in a downstream AI market?"

However, Bronner, Trinko and IMS Health demonstrate that compulsory-access theories remain exceptional and require demanding conditions.

22. Six Core Case-Law Principles — Exam Table

23. Conclusion

Cognitive infrastructure market power represents a new competition-law problem in which economic power can arise from control over the physical and technological foundations on which AI systems depend.

The principal competition concerns are:

1. control over scarce AI computing resources;
2. cloud concentration;
3. GPU and accelerator dependence;
4. high switching costs;
5. interoperability restrictions;
6. exclusive compute arrangements;
7. vertical foreclosure;
8. tying and bundling;
9. self-preferencing;
10. strategic acquisitions; and
11. ecosystem lock-in.

The traditional cases—Terminal Railroad, Aspen Skiing, Trinko, Commercial Solvents, Bronner and IMS Health—provide the doctrinal foundation, while the NVIDIA–Arm, NVIDIA–Mellanox and Microsoft–OpenAI matters demonstrate how those principles are increasingly being applied to AI-related infrastructure.

The central competition-law challenge is therefore to distinguish legitimate technological integration and infrastructure investment from the strategic use of infrastructure control to exclude or weaken competing AI ecosystems.