Model Licensing Ecosystems and Gatekeeper Formation

1. Introduction

Model licensing ecosystems refer to the networks of contractual, technical and commercial arrangements through which artificial intelligence (AI) models are developed, distributed, integrated, modified and commercialised. These arrangements may include licences for model weights, source code, APIs, training data, inference services, fine-tuning, deployment, redistribution and downstream applications.

A gatekeeper emerges when a firm acquires the ability to control commercially important routes through which other businesses, developers or users access AI capabilities. Gatekeeping can arise even when the model itself is not the largest or most technologically advanced. For example, a firm may control access through restrictive API terms, exclusive distribution agreements, cloud-computing dependencies, certification requirements or contractual restrictions on interoperability.

The central competition-law concern is whether licensing enables legitimate protection of intellectual property, investment and safety, or instead creates barriers that exclude rivals, preserve market power and make customers dependent on a single ecosystem.

Three distinctions are essential:

Intellectual property protection versus market foreclosure: A model provider may legitimately protect its technology, but licensing restrictions can become problematic when they unnecessarily impede competition.

Technical superiority versus gatekeeper power: A superior model does not automatically make its developer a dominant undertaking. Market power may instead arise from control over distribution, compute, cloud infrastructure or essential complementary services.

Open licensing versus effective openness: A licence described as “open” may still impose material restrictions on commercial use, redistribution, modification or access to complementary resources. The precise terms and practical ability to use the model matter more than the label.

The legal analysis must therefore examine the complete licensing ecosystem rather than treating each model licence as an isolated contract.

2. Legal framework

A. United Kingdom

Competition Act 1998

Chapter I, section 2: Prohibits agreements, decisions and concerted practices that prevent, restrict or distort competition, subject to applicable exemptions.

Chapter II, section 18: Prohibits abuse of a dominant position.

Licensing arrangements may raise concerns where they impose exclusivity, restrict downstream competition, tie access to unrelated services, discriminate between business partners or prevent customers from switching.

Digital Markets, Competition and Consumers Act 2024 (DMCC Act)

The Act establishes a regime for firms designated as having strategic market status (SMS) in relation to a particular digital activity. The Competition and Markets Authority (CMA) may impose tailored conduct requirements and, where the statutory conditions are satisfied, pro-competition interventions.

For AI licensing, relevant issues may include restrictions on interoperability, unfair contractual terms, discriminatory access and the leveraging of control over one digital activity into adjacent markets. These obligations do not automatically apply to every AI model developer: the statutory designation and applicable requirements matter.

B. Germany

Gesetz gegen Wettbewerbsbeschränkungen (GWB)

Section 19: Prohibits abuse of a dominant market position.

Section 19a: Provides special powers concerning certain conduct by undertakings of paramount significance for competition across markets, subject to the statutory requirements and the Bundeskartellamt's determinations.

Section 20: Extends certain protections to specified forms of relative market power and dependency, even where conventional dominance is not established.

German law is particularly relevant when a model provider controls a critical input and businesses have no reasonable alternatives, such as when they depend on proprietary model APIs, specialised cloud infrastructure, or licences necessary to maintain compatibility with an established platform.

Section 19a is not a general AI licensing code. Its application depends on the undertaking and conduct falling within the statutory framework.

C. European Union

Article 101 TFEU addresses anticompetitive agreements, including certain exclusive licensing arrangements, market-allocation provisions and restrictions on independent commercial activity.

Article 102 TFEU prohibits abuse of a dominant position, including exclusionary conduct where the relevant legal conditions are met.

The EU framework also includes:

The Digital Markets Act (DMA), which imposes specified obligations on designated gatekeepers in relation to their core platform services.

The EU Technology Transfer Block Exemption Regulation (TTBER) and associated guidelines, which address certain technology-transfer agreements and licensing restrictions.

Merger control, where acquisitions of model developers, licensing businesses, cloud providers or complementary technologies could materially reduce competition.

Neither the DMA nor the TTBER automatically governs every AI model licence. The applicable rules depend on the type of agreement, parties, market position, designated services and restrictions in question.

3. How model licensing creates gatekeeper power

Upstream model and intellectual property

Weights, architecture, training assets, proprietary capabilities

Licensing and access control

API terms, exclusivity, usage limits, certification, pricing

Complementary infrastructure

Cloud, GPUs, identity systems, developer tools, distribution

Downstream dependency

Applications, enterprise workflows, agents and end users

Potential gatekeeper formation

High switching costs, restricted entry, reduced multi-homing and control over commercial access

This sequence is not inevitable. A model provider may remain subject to strong competition if customers can readily switch models, deploy alternative models locally, use multiple APIs or obtain equivalent services from competing providers.

Principal mechanisms of gatekeeper formation

1. Exclusive licensing

A model developer grants one cloud provider, distributor or application platform exclusive rights to host, distribute or commercialise a model. Exclusivity may support investment and quality assurance, but it can also deny rivals access to an important input or distribution channel.

2. API dependency and contractual lock-in

Developers build applications around proprietary APIs, function-calling formats, embeddings, safety systems and specialised model behaviour. Switching may require software redevelopment, testing, retraining and regulatory validation. The provider may then have the ability to raise prices or worsen terms without losing enough customers to make the change unprofitable.

3. Restrictions on competing models

Licences may prohibit benchmarking, comparative testing, fine-tuning, use of outputs to train competing systems, or deployment through alternative providers. Some restrictions can protect confidential information or legitimate investments. Their competitive effects depend on scope, duration, necessity and market context.

4. Bundling and tying

A provider may condition access to a model on the purchase of its cloud, identity, data-management, productivity or distribution services. This may be efficient when the services are technically complementary, but potentially exclusionary where customers cannot obtain the model independently and competing suppliers are disadvantaged.

5. Data and feedback advantages

A provider may combine model access with privileged access to usage data, feedback loops, evaluation infrastructure or customer relationships. If these advantages reinforce model quality and distribution at scale, competitors may face increasing difficulty entering or expanding.

6. Certification and interoperability control

A dominant provider may determine which models are certified, compatible or eligible for access to a platform. Discriminatory certification, delayed integration or technically unnecessary compatibility restrictions can become a gatekeeping mechanism.

4. Case laws and their application to model licensing

The following decisions provide established legal principles that can be applied by analogy to AI model licensing. They are not findings that the courts have already established AI-specific licensing liability.

Case 1: Magill — refusal to license intellectual property

Reference: Joined Cases C-241/91 P and C-242/91 P, RTE and ITP v Commission (1995).

Principle: The European Court of Justice recognised that, in exceptional circumstances, a refusal by dominant undertakings to license protected material may constitute an abuse of dominance. The circumstances included the prevention of a new product for which there was potential consumer demand, the absence of justification and the reservation of a secondary market to the rights holders.

Application to AI licensing: A dominant model provider's refusal to license model capabilities, interface specifications or other protected material may warrant scrutiny where the refusal prevents the emergence of a distinct downstream product and satisfies the demanding legal conditions.

Limitation: Magill does not create a general obligation to license AI models. Intellectual property rights and commercial freedom remain important, and the exceptional circumstances must be demonstrated.

Case 2: IMS Health — indispensable inputs and interoperability

Reference: Case C-418/01, IMS Health GmbH & Co. OHG v NDC Health GmbH & Co. KG (2004).

Principle: The Court clarified the exceptional conditions under which refusal to license intellectual property may be abusive. These included indispensability, the elimination of competition in a secondary market, prevention of a new product for which there was potential demand, and the absence of objective justification.

Application to AI licensing: Suppose a dominant provider controls a proprietary model interface that downstream businesses cannot realistically replicate, and refuses access needed to create competing services. IMS Health offers a framework for evaluating whether that refusal crosses the threshold from legitimate IP protection into abusive exclusion.

Limitation: A model being popular, technically superior or costly to reproduce does not by itself establish indispensability. Realistic alternatives and the actual effect on downstream competition must be assessed.

Case 3: Microsoft v Commission — interoperability and strategic exclusion

Reference: Case T-201/04, Microsoft Corp. v Commission (General Court, 2007).

Principle: The General Court upheld the Commission's findings concerning Microsoft's refusal to provide interoperability information to competing work-group server operating-system suppliers, as well as the associated remedial framework.

Application to AI licensing: This decision is relevant where a dominant AI ecosystem controls technical information, APIs, model interfaces or compatibility mechanisms needed by rival applications. A provider might use interface changes or discriminatory access terms to make rival models harder to integrate into established enterprise software.

The competition concern becomes stronger where the restriction protects an adjacent market from competition rather than serving a proportionate technical or security purpose.

Limitation: The decision does not establish that every API or model-weight disclosure is compulsory. The precise market circumstances and applicable legal conditions remain decisive.

Case 4: Bronner — the essential-facilities threshold

Reference: Case C-7/97, Oscar Bronner GmbH & Co. KG v Mediaprint (1998).

Principle: The Court set a demanding threshold for requiring a dominant undertaking to provide access to an infrastructure it controls. The facility must satisfy the applicable indispensability criteria, and refusal must be capable of eliminating all competition by the requester in the relevant market, subject to the judgment's specific legal requirements.

Application to AI licensing: A downstream developer's dependence on one advanced model does not automatically entitle it to a licence. The analysis should examine alternative models, open-weight substitutes, independent development, competing APIs and the feasibility of migration.

The case guards against turning competition law into a general obligation for successful AI developers to share their innovations with rivals.

Limitation: Not every licensing dispute falls within the essential-facilities doctrine; the legal framework varies with the conduct alleged.

Case 5: Huawei v ZTE — licensing, intellectual property and fair dealing

Reference: Case C-170/13, Huawei Technologies Co. Ltd v ZTE Corp. (2015).

Principle: The Court addressed the circumstances in which a holder of standard-essential patents subject to a FRAND commitment may seek an injunction against an alleged infringer. It identified a sequence of good-faith conduct relevant to balancing intellectual property rights and competition law.

Application to AI licensing: The case is useful by analogy where a model provider controls technology embedded in an industry standard or makes interoperability dependent on proprietary rights. Transparent terms, meaningful negotiation and non-discriminatory access may be relevant to assessing conduct.

For example, if a model interface becomes a practical industry standard, licensing terms could influence whether smaller developers can compete on equal terms.

Limitation: Huawei concerns standard-essential patents and FRAND commitments. Its negotiation sequence does not automatically apply to ordinary AI model licences.

Case 6: Google Android — tying, defaults and ecosystem leverage

Reference: Case T-604/18, Google and Alphabet v Commission (General Court, 2022).

Principle: The General Court largely upheld the Commission's decision concerning Google's Android-related restrictions, including certain tying and anti-fragmentation arrangements, while annulling part of the decision concerning portfolio-based revenue-sharing agreements and adjusting the fine.

Application to AI licensing: A model provider could potentially leverage dominance in an established platform to strengthen its position in AI services by tying access to a model to default placement, distribution arrangements or the use of complementary products.

The key question is whether contractual or technical restrictions limit competing providers' ability to reach users and whether the restrictions can be objectively justified.

Limitation: Integration and bundling are not inherently unlawful. Their effects, market position and justification must be evaluated.

Case 7: Google Shopping — leveraging market power into adjacent markets

Reference: Case C-48/22 P, Google and Alphabet v Commission (Court of Justice, 2024).

Principle: The Court of Justice upheld the General Court's judgment affirming the Commission's finding of abuse in Google's treatment of its own comparison-shopping service in general search results.

Application to AI licensing: A platform that controls both a widely used AI distribution channel and a competing downstream model or application may favour its own services through ranking, default settings, access restrictions or integration advantages.

For example, a model marketplace might systematically disadvantage competing models while promoting the operator's own model, potentially foreclosing downstream rivals.

Limitation: Self-preferencing is not automatically unlawful in every market. The applicable dominance and abuse requirements, as well as the evidence of exclusionary effects, remain important.

Case 8: Slovak Telekom — access restrictions and margin squeeze

Reference: Case C-165/19 P, Slovak Telekom v Commission (Court of Justice, 2021).

Principle: The Court upheld key aspects of the Commission's finding of abuse concerning restrictive access conditions and margin squeeze in telecommunications. The judgment addressed the relationship between refusal-to-supply case law and forms of abusive conduct that do not amount to a complete refusal of access.

Application to AI licensing: A model provider might technically grant access while imposing conditions that make competitive downstream operation commercially unviable. Examples could include excessive API charges, discriminatory usage limits or pricing structures that disadvantage independent application developers relative to the provider's own downstream services.

Limitation: Whether a pricing arrangement constitutes a margin squeeze or other abuse depends on the relevant market, cost and pricing evidence, and the applicable legal test. High prices alone are not sufficient.

Case 9: United Brands — unfair trading conditions and dominance

Reference: Case 27/76, United Brands Company v Commission (1978).

Principle: The Court examined abuse of dominance involving, among other matters, discriminatory and unfair commercial conduct. It also developed the established two-stage framework for assessing whether prices are unfairly high, although application of that test is demanding.

Application to AI licensing: Where a dominant model provider imposes materially different terms on similarly situated licensees, or charges exploitative prices in circumstances satisfying the legal test, Article 102 or the corresponding UK or German provisions may become relevant.

A meaningful assessment would compare contractual terms, commercial justifications, bargaining positions, available alternatives and the provider's market power.

Limitation: Different prices and contractual terms may reflect legitimate differences in volume, risk, service quality or investment. Competition law does not generally require identical terms for all licensees.

5. Applying competition law to different licensing models

These are potential theories of harm, not presumptions of illegality. Each requires assessment of the specific conduct, market conditions and legal elements.

6. How to establish gatekeeper formation

A competition authority should examine the following six dimensions.

1. Relevant market definition

Possible markets include foundation-model services, model API access, enterprise AI services, model-hosting infrastructure, model-development tools or specialised AI applications. These are hypotheses to be tested, not predetermined market definitions.

Authorities should investigate substitution across proprietary APIs, open-weight models, alternative cloud providers and in-house systems.

2. Market power and dependency

Relevant evidence includes market shares, the availability of credible alternatives, customer concentration, entry barriers, switching costs, control of key inputs and the ability to impose worse terms without losing sufficient business.

3. Contractual restrictions

Authorities should review exclusivity, minimum-purchase commitments, restrictions on multi-homing, anti-benchmarking provisions, output-use restrictions, termination rights, price structures and limits on redistribution.

4. Foreclosure and competitive effects

The authority should ask whether rival model providers or downstream application developers are prevented from entering, expanding or competing effectively. Evidence may include lost distribution opportunities, increased rival costs, reduced innovation and customer migration difficulties.

5. Objective justification

The provider may demonstrate that restrictions are reasonably necessary to protect security, confidential information, intellectual property, service reliability or legitimate investment. Authorities should examine whether less restrictive alternatives could achieve the same purpose.

6. Consumer and innovation effects

Relevant outcomes include prices, quality, privacy, reliability, choice, independent innovation and the ability of businesses to develop competing products. Short-term improvements in model performance should not obscure long-term exclusionary effects, but speculative innovation harms should not replace evidence.

7. Remedies and regulatory responses

Potential remedies should be proportionate to the identified harm.

Non-discriminatory licensing: Require fair and objectively justified access terms where the relevant legal requirements are met.

Interoperability measures: Address unjustified technical restrictions that prevent competing services from functioning with a dominant ecosystem.

Restrictions on exclusivity: Prohibit or limit exclusionary agreements where they materially foreclose competition.

Contractual transparency: Improve disclosure of material licence terms, usage restrictions, pricing and termination conditions.

Portability and switching: Reduce unnecessary migration costs and allow customers to transfer appropriate data, configurations or workflows.

Separation or structural remedies: Consider only where legally available, justified by the evidence and necessary to address persistent competition problems that less intrusive measures cannot adequately resolve.

Monitoring and compliance: Use proportionate reporting, audits and review mechanisms to test whether remedies work in practice.

In the UK, the CMA's powers under the DMCC Act depend on the applicable statutory conditions. In Germany, measures under GWB §§19 and 19a require the relevant legal basis. EU remedies under Articles 101 and 102 TFEU must address the established infringement and comply with proportionality requirements.

8. Critical analysis

Model licensing ecosystems create a difficult balance between encouraging innovation and preventing entrenched control.

First, mandatory access can weaken incentives to invest if it requires disclosure or licensing on terms that inadequately protect legitimate commercial interests. Conversely, unchecked exclusivity and ecosystem lock-in may allow an incumbent to preserve market power after its original technological advantage diminishes.

Second, open-weight models can reduce dependence but do not eliminate gatekeeping. Developers may still rely on proprietary cloud infrastructure, specialised chips, deployment platforms, safety tooling or distribution channels.

Third, licensing restrictions can compound across an AI supply chain. A developer may depend on a model provider, which depends on a cloud provider, while downstream applications depend on the developer's API. Each contractual relationship may appear reasonable in isolation even where their cumulative effects substantially increase entry barriers.

Fourth, rapid technological change complicates market definition and remedies. A model that is indispensable today may face credible substitutes tomorrow. Authorities should therefore examine both current competitive constraints and realistic developments over the relevant period.

Finally, competition-law intervention must remain evidence-based. The objective is not to punish successful model developers or require universal disclosure of proprietary technology. It is to preserve the conditions under which competing businesses can enter, innovate and reach customers.

9. Conclusion

Model licensing becomes a competition-law concern when contractual rights, technical controls and complementary infrastructure combine to give a firm durable power over access to AI capabilities or downstream markets.

The most useful authorities are Magill, IMS Health, Microsoft, Bronner, Huawei v ZTE, Google Android, Google Shopping, Slovak Telekom and United Brands. Collectively, they provide principles concerning exceptional licensing refusals, interoperability, tying, exclusionary access restrictions, discriminatory treatment and unfair commercial conditions.

Their application to AI remains fact-specific. The decisive question is whether a licensing ecosystem represents legitimate commercial exploitation of innovation or enables a dominant undertaking to exclude rivals, exploit dependency or extend its market power into adjacent activities.

For examination purposes, the central proposition is that gatekeeper formation is not established merely by owning a valuable AI model; it arises as a competition-law issue when control over model access or complementary infrastructure enables legally cognisable exclusionary or exploitative conduct.