Competition Law and Cognitive Infrastructure Gatekeepers

1. Introduction

Cognitive infrastructure refers to the underlying technological infrastructure that enables advanced computation, artificial intelligence, machine learning, data processing, knowledge services, and automated decision-making. It may include:

• cloud-computing infrastructure;
• AI compute and GPU infrastructure;
• data centres;
• AI model-serving infrastructure;
• semiconductor and accelerator ecosystems;
• operating systems and developer platforms;
• data repositories and large-scale datasets;
• APIs and interoperability layers;
• AI application marketplaces;
• identity, authentication and payment infrastructure; and
• networks connecting AI developers, enterprises and end-users.

A cognitive infrastructure gatekeeper is an undertaking that controls an infrastructure layer through which competitors, developers, businesses or consumers must pass to reach another market.

The competition-law concern is not simply that a company is large. The concern arises where control over an indispensable or strategically important infrastructure layer enables the undertaking to exclude rivals, raise their costs, leverage dominance into adjacent markets, restrict interoperability, foreclose access to data or compute, or disadvantage downstream competitors.

The European Commission's recent work illustrates the growing importance of this issue: it has specifically investigated whether cloud services such as AWS and Azure function as important gateways and has examined interoperability barriers, conditioned access to data, tying/bundling and imbalanced contractual terms in cloud markets.

2. Meaning of a Cognitive Infrastructure Gatekeeper

A gatekeeper may occupy several layers simultaneously:

A. Compute gatekeeper

Controls:

• GPUs/AI accelerators;
• high-performance computing;
• AI clusters;
• specialised inference infrastructure;
• access to computational capacity.

B. Cloud gatekeeper

Provides:

• virtual machines;
• storage;
• AI-as-a-service;
• model hosting;
• databases;
• networking;
• developer tools.

C. Data gatekeeper

Controls commercially or technically important:

• datasets;
• training data;
• behavioural data;
• search data;
• telemetry;
• proprietary databases.

D. Platform gatekeeper

Controls:

• operating systems;
• application stores;
• AI marketplaces;
• APIs;
• developer ecosystems;
• authentication systems.

E. Model infrastructure gatekeeper

An undertaking may control access to:

• foundation models;
• model APIs;
• inference services;
• model fine-tuning;
• model deployment;
• agent infrastructure.

The key competition question is therefore:

Does control over the infrastructure give the undertaking the ability and incentive to restrict competition in a neighbouring or downstream market?

3. Relevant Competition-Law Framework

A. Dominance

The first issue is usually whether the undertaking possesses substantial market power.

Relevant factors include:

1. market share;
2. barriers to entry;
3. switching costs;
4. network effects;
5. economies of scale;
6. control over proprietary data;
7. access to scarce computing resources;
8. technical interoperability;
9. ecosystem dependence;
10. vertical integration;
11. customer lock-in; and
12. access to capital and infrastructure.

In cognitive infrastructure markets, conventional market-share analysis may be insufficient because an undertaking can exercise significant competitive power through control of an ecosystem bottleneck even where several nominal competitors exist.

4. Essential-Facility and Access Problems

A central issue is whether infrastructure constitutes an essential facility.

The basic concern is:

A dominant undertaking controls infrastructure that competitors cannot reasonably reproduce, and refusal or discriminatory access prevents effective competition in a related market.

However, competition law generally does not impose a universal duty upon dominant firms to deal with competitors.

Courts have traditionally considered factors such as:

• indispensability;
• absence of realistic alternatives;
• elimination of effective competition;
• technical feasibility of access;
• legitimate business justification; and
• whether access obligations would undermine investment incentives.

5. Six Major Case Laws

Case 1 — United Brands v Commission (1978)

Principle

The European Court of Justice established important principles concerning abuse of dominance and the special responsibility of dominant undertakings.

Relevance to cognitive infrastructure

A cognitive infrastructure provider with substantial market power cannot use its infrastructure position in a manner that exploits or excludes competitors contrary to competition law.

For example, a dominant cloud provider could potentially raise concerns if it:

• imposes discriminatory access conditions;
• uses infrastructure control to exclude rival AI providers;
• imposes unjustified commercial conditions; or
• leverages infrastructure dominance into downstream AI services.

Significance

The case provides the foundational concept that a dominant undertaking has a special responsibility not to undermine effective competition.

Case 2 — Commercial Solvents v Commission (1974)

Principle

A dominant undertaking's refusal to supply an essential input to downstream competitors can constitute an abuse of dominance.

Commercial Solvents controlled an important input and attempted to use that position to strengthen its position in a downstream market.

Application to cognitive infrastructure

Suppose a dominant AI-infrastructure company controls a critical computational resource and simultaneously competes in downstream AI services.

If it:

controls compute → refuses/limits supply → competes downstream

the conduct may raise a vertical foreclosure problem.

Potential examples include:

• restricting GPU/cloud capacity to competing AI developers;
• preferentially allocating scarce compute to its own applications;
• discriminatory API access;
• refusing necessary infrastructure interfaces.

Competition concern

The infrastructure provider potentially converts an upstream bottleneck into downstream market power.

Case 3 — Magill — RTE and ITP v Commission (1995)

Principle

The Court recognised that refusal to license certain intellectual-property information can, in exceptional circumstances, amount to an abuse of dominance.

The Court identified stringent circumstances involving:

1. indispensability;
2. prevention of a new product for which consumer demand exists;
3. exclusion of competition in a secondary market; and
4. absence of objective justification.

Cognitive infrastructure application

This is particularly relevant to AI datasets and proprietary information.

Suppose a dominant cognitive infrastructure company controls a uniquely valuable dataset required to develop a competing AI service.

A competition-law analysis could ask:

• Is the dataset genuinely indispensable?
• Are alternative datasets realistically available?
• Does withholding it eliminate effective competition?
• Would access facilitate a genuinely distinct downstream product?
• Is there an objective justification for refusal?

Importance

Magill demonstrates that access to information can become a competition issue when information is strategically indispensable to a downstream market.

Case 4 — Bronner v Mediaprint (1998)

Principle

The Court adopted a strict approach to refusal-to-deal and essential-facility claims.

A facility will not ordinarily be considered indispensable merely because obtaining an alternative is more expensive or inconvenient.

Cognitive infrastructure application

This is extremely important for cloud and AI infrastructure.

A smaller AI company cannot automatically claim:

"The dominant cloud provider's infrastructure is better, therefore I must receive access."

The legal inquiry is stronger:

• Is there a realistic alternative?
• Can the infrastructure reasonably be replicated?
• Would alternative infrastructure be economically or technically feasible?
• Does refusal eliminate effective competition?

Example

If an AI developer can migrate from Cloud A to Cloud B with reasonable technical adaptation, Cloud A may not constitute an indispensable facility.

But if Cloud A controls a unique technical interface, dataset or infrastructure capability that cannot realistically be replicated, the analysis becomes significantly more serious.

Case 5 — IMS Health v NDC Health (2004)

Principle

The Court reaffirmed the exceptional character of compulsory-access remedies involving intellectual property and dominant infrastructure.

The case concerned a proprietary system that competitors needed to operate effectively in a downstream market.

Relevance

The case is particularly useful for analysing:

• proprietary APIs;
• technical standards;
• data architectures;
• interoperability interfaces;
• proprietary AI infrastructure;
• developer access systems.

Key lesson

Interoperability importance does not automatically create a legal duty to disclose.

The infrastructure must satisfy demanding conditions before compulsory access becomes justified.

This is crucial because forcing infrastructure providers to open proprietary systems may affect:

• innovation incentives;
• security;
• intellectual-property rights;
• cybersecurity;
• investment incentives.

Case 6 — Microsoft Corp. v Commission (General Court, 2007)

Principle

The Microsoft case is one of the most important authorities for technology infrastructure and interoperability.

Microsoft was found to have abused its dominant position through conduct involving interoperability information and tying.

Cognitive infrastructure relevance

The case illustrates how control over one technological layer can be leveraged into neighbouring markets.

The structure resembles:

Dominant platform → control over interoperability → advantage in adjacent market

Modern equivalents could potentially involve:

• cloud → AI services;
• operating system → AI assistant;
• AI platform → application marketplace;
• cloud identity → enterprise software;
• AI model → downstream applications.

Example

If a dominant cloud provider makes its AI infrastructure substantially more interoperable with its own applications than with rival AI applications, competition authorities could examine whether the conduct produces exclusionary effects.

Case 7 — Google Shopping — Google and Alphabet v Commission (2024)

Although not an infrastructure case in the traditional physical sense, Google Shopping is highly relevant to modern cognitive infrastructure.

Principle

The EU courts upheld the Commission's finding that Google abused its dominant position in general search by favouring its own comparison-shopping service over competing comparison-shopping services.

Cognitive-infrastructure significance

The case illustrates self-preferencing through a strategically important gateway.

The competitive structure can be expressed as:

Infrastructure/gateway control → preferential treatment of own downstream service → foreclosure of rival downstream services

This concept can become relevant to AI ecosystems where a company simultaneously operates:

• cloud infrastructure;
• foundation models;
• AI applications;
• search;
• advertising;
• developer platforms.

The question becomes whether infrastructure control is being used to systematically favour the undertaking's own downstream products.

6. Gatekeeper Strategies That May Raise Competition Concerns

6.1 Self-preferencing

A cognitive infrastructure provider may favour its own downstream services by:

• allocating superior computing resources;
• giving its own models lower latency;
• providing preferential API access;
• giving its own applications better visibility;
• providing technical features unavailable to competitors.

The competition concern is stronger when competitors depend upon the infrastructure provider.

7. Tying and Bundling

A dominant infrastructure provider might require customers to purchase:

Cloud + AI model + cybersecurity + database + identity services

as a package.

Bundling may generate efficiencies, but competition concerns arise where customers are effectively prevented from obtaining competing components.

Relevant questions include:

• Are the products distinct?
• Is the undertaking dominant in the tying product?
• Is purchasing conditional?
• Are competitors foreclosed?
• Are efficiencies demonstrable?
• Can customers technically separate the services?

8. Interoperability Restrictions

Interoperability is particularly important in cognitive infrastructure.

Potential restrictions include:

• proprietary APIs;
• incompatible data formats;
• restrictions on model portability;
• restrictions on cloud migration;
• technical barriers to switching;
• restrictions on third-party model deployment.

A provider may therefore increase switching costs by creating:

Data lock-in + model lock-in + infrastructure lock-in.

The European Commission's current cloud investigation expressly considers interoperability obstacles, access to business-user data, tying/bundling and contractual conditions.

9. Data Lock-In

Data can function as a competitive infrastructure.

A provider may possess:

• customer data;
• operational data;
• telemetry;
• search data;
• training datasets;
• model-performance data.

If customers cannot easily export their data, competitors may face a significant barrier to entry.

The competition-law issue is therefore not simply:

"Who owns the data?"

but:

Does control over the data materially prevent customers or competitors from switching and competing?

10. Cloud Switching Costs

Cloud infrastructure can create several layers of dependence:

Technical dependence

Applications are built around proprietary APIs.

Financial dependence

Customers receive discounts for long-term commitments.

Data dependence

Data is expensive or technically difficult to migrate.

Human-capital dependence

Employees become specialised in one provider's ecosystem.

AI dependence

AI models are trained and deployed using proprietary cloud tools.

The combination may create substantial ecosystem lock-in.

11. Exclusive Dealing

A dominant infrastructure provider may enter contracts requiring customers to:

• purchase minimum quantities;
• use only its cloud;
• deploy AI models exclusively through its infrastructure;
• avoid competing infrastructure;
• obtain discounts conditional on exclusivity.

Competition authorities would need to examine whether such arrangements foreclose rivals and for how long.

12. Predatory or Strategic Pricing

Infrastructure providers may have incentives to price aggressively because they can recover losses elsewhere.

For example:

Cloud infrastructure → low price → customer acquisition → AI services → monetisation

This can create a potential cross-subsidisation concern.

However, low prices are not inherently anticompetitive. Authorities generally need evidence that pricing strategy is capable of excluding equally efficient competitors or otherwise violates applicable competition rules.

13. Discriminatory Access to Compute

AI development increasingly depends on access to scarce computational resources.

A dominant infrastructure provider could potentially discriminate between:

• internal AI projects;
• affiliated companies;
• independent AI developers;
• rival model providers.

Examples could include:

14. Vertical Foreclosure

The most important structural concern can be represented as:

Infrastructure

↓

Cloud / compute / data / operating system

↓

AI model

↓

AI application

↓

Consumer / enterprise

A vertically integrated company may control several stages.

The competition issue becomes whether the company can use control at an upstream stage to weaken competitors at downstream stages.

15. Ecosystem Leverage

Traditional competition analysis often examines a single relevant market.

Cognitive infrastructure requires a broader ecosystem perspective.

For example:

Cloud → compute → data → foundation model → API → application → distribution

Market power at one layer can reinforce market power at another.

This creates a feedback loop:

More users → more data → better AI → more users → more infrastructure demand → greater scale → stronger ecosystem

Such network and feedback effects can create substantial barriers to entry.

16. Essential-Facility Analysis

A useful legal framework is:

Step 1 — Identify the facility

What infrastructure is allegedly indispensable?

Step 2 — Define the relevant market

Examples:

• public cloud services;
• AI compute;
• GPU infrastructure;
• AI model hosting;
• AI APIs.

Step 3 — Establish dominance

Assess:

• market shares;
• switching costs;
• entry barriers;
• network effects;
• economies of scale.

Step 4 — Test indispensability

Are alternatives realistically available?

Step 5 — Examine refusal

Was access denied, restricted or made commercially unreasonable?

Step 6 — Examine competitive foreclosure

Did the conduct eliminate or substantially weaken competition?

Step 7 — Examine justification

Could the conduct be justified by:

• security;
• privacy;
• cybersecurity;
• intellectual property;
• capacity constraints;
• technical limitations?

17. Digital Gatekeeper Regulation

Traditional abuse-of-dominance law operates after dominance and potentially abusive conduct are established.

The EU Digital Markets Act represents a more ex-ante approach to designated gatekeepers.

The EU originally designated Alphabet, Amazon, Apple, ByteDance, Meta and Microsoft as DMA gatekeepers in 2023.

The significance for cognitive infrastructure is increasing because the European Commission has specifically examined whether cloud computing should fall within the gatekeeper framework. In June 2026, the Commission announced a preliminary view that AWS and Azure should be designated as gatekeepers for cloud computing services, subject to the companies' responses and final decisions.

This demonstrates an important regulatory shift:

Infrastructure itself can become a gateway worthy of gatekeeper regulation, rather than merely being treated as an ordinary input.

18. Cognitive Infrastructure and the Essential-Facility Doctrine

The doctrine can be summarised as follows:

19. Competition Issues in AI Compute Markets

AI compute has unusual characteristics.

Scarcity

High-end accelerators and specialised infrastructure may have limited availability.

Economies of scale

Large providers can spread infrastructure costs over enormous workloads.

Capital requirements

Building competitive infrastructure may require substantial investment.

Network effects

More developers can make an infrastructure platform more attractive.

Switching costs

Moving AI workloads can require substantial engineering effort.

Consequently, a concentrated AI-compute market may create infrastructure-level bottlenecks.

20. Remedies

Competition authorities may consider several remedies.

A. Access remedies

Require access to infrastructure on reasonable and non-discriminatory terms.

B. Interoperability

Require technical compatibility between competing systems.

C. Data portability

Allow users to transfer data between providers.

D. Non-discrimination

Prevent infrastructure providers from favouring their own downstream products.

E. Separation

In particularly serious circumstances, functional or structural separation may be considered.

F. Contractual remedies

Restrictions on:

• exclusivity;
• minimum commitments;
• termination fees;
• switching penalties.

G. Merger control

Authorities may scrutinise acquisitions of:

• AI startups;
• compute providers;
• model developers;
• data companies;
• cloud infrastructure businesses.

This is important because acquiring a potential future competitor can strengthen an already integrated cognitive ecosystem.

21. Six-Case-Law Comparison

22. Key Legal Tests

A competition-law assessment of a cognitive infrastructure gatekeeper should ask:

Market power

1. What is the relevant market?
2. Does the undertaking possess substantial market power?
3. Are there realistic alternatives?

Infrastructure

1. Is the infrastructure indispensable?
2. Can competitors reproduce it?
3. Are switching costs significant?

Conduct

1. Is there refusal to supply?
2. Is there discriminatory access?
3. Is there tying or bundling?
4. Is there self-preferencing?
5. Is there exclusive dealing?
6. Is interoperability being restricted?

Effects

1. Are competitors foreclosed?
2. Are entry barriers increased?
3. Is innovation reduced?
4. Are customers locked into the ecosystem?
5. Are prices or quality affected?

Justification

1. Is there a legitimate technical reason?
2. Is cybersecurity involved?
3. Is privacy protection involved?
4. Are capacity limitations genuine?
5. Are less restrictive alternatives available?

23. Conclusion

Cognitive infrastructure gatekeepers represent a convergence of traditional essential-facility doctrine, abuse-of-dominance law, interoperability regulation, platform regulation and merger control.

The central competition-law problem is not merely that an infrastructure provider has a large market share. It is that control over a critical technological gateway can allow the provider to influence who receives compute, data, APIs, interoperability, distribution and access to downstream AI markets.

The major authorities—Commercial Solvents, Magill, Bronner, IMS Health, Microsoft, United Brands and Google Shopping—provide different components of the legal framework.

The emerging regulatory model therefore focuses increasingly on:

Infrastructure control → gateway power → access conditions → interoperability → downstream foreclosure → ecosystem dominance.

For cognitive infrastructure, the most important future competition-law questions are likely to concern AI compute scarcity, cloud lock-in, proprietary datasets, AI-model access, interoperability, self-preferencing, vertical integration, exclusive infrastructure contracts and acquisitions that consolidate control across multiple layers of the AI stack. The European Commission's ongoing cloud work specifically identifies interoperability, data access, tying/bundling and contractual conditions as areas requiring examination.