AI Model Registry Monopolization Concerns

1. Introduction

An AI model registry is a centralized platform through which AI models are catalogued, versioned, authenticated, evaluated, distributed, licensed, deployed, or connected to downstream applications. A registry may contain foundation models, fine-tuned models, embeddings, multimodal models, safety models, model cards, evaluation scores, APIs, weights, datasets, and deployment metadata.

Examples of registry functions include:

• model discovery and search;
• model hosting and download;
• version control;
• model verification and authentication;
• benchmark and safety scoring;
• licensing information;
• API access;
• model deployment;
• compatibility certification;
• provenance and model-card information;
• access to fine-tuned or specialized models.

A monopolization concern arises when one registry becomes a critical gateway between model developers, cloud providers, application developers and users, allowing the registry operator to influence which models are visible, accessible, compatible, certified, monetized, or commercially successful.

China's current platform-antitrust framework is particularly relevant because SAMR's 2026 Internet Platform Anti-Monopoly Compliance Guidelines expressly identify data, algorithms, technology, capital and platform rules as sources of platform market power and identify refusal to deal, tying, unreasonable conditions and discriminatory treatment as potential abuses. The Guidelines specifically mention control over essential data, models and application platforms as potentially relevant to refusal-to-deal analysis.

2. Relevant Competition Markets

An AI registry can participate simultaneously in several related markets.

A. AI Model Registry Market

The narrowest market could consist of platforms providing:

• model discovery;
• model hosting;
• model versioning;
• model authentication;
• model distribution.

B. AI Model Marketplace

A broader market may include commercial transactions involving:

• model licenses;
• API access;
• inference services;
• fine-tuning;
• enterprise deployment.

C. Model-Distribution Infrastructure

A registry can become infrastructure connecting:

Model Developer → Registry → Cloud/Compute Provider → Application Developer → End User

Control over this intermediary layer can create significant competitive leverage.

D. AI Evaluation and Certification

If developers rely on one registry's:

• benchmark scores,
• safety certifications,
• compliance labels,
• reliability ratings,

the registry may obtain additional market power even without controlling the underlying models.

E. Model-to-Application Interoperability

A registry may control:

• APIs;
• SDKs;
• authentication;
• deployment formats;
• model adapters;
• metadata standards.

This can create switching costs and technical dependency.

3. How Monopolization Can Develop

3.1 Network Effects

The registry becomes more valuable as more participants use it.

More models → more developers → more users → more model providers → more models

This creates a self-reinforcing network effect.

A dominant registry may therefore become difficult for competitors to challenge even if alternative registry technology is technically available.

3.2 Model Discoverability Control

Search ranking can determine which AI models receive commercial exposure.

A registry operator could potentially:

• rank its own models first;
• demote competing models;
• manipulate benchmark presentation;
• provide preferential recommendations;
• hide competing models;
• place proprietary models in default deployment options.

This resembles competition concerns previously examined in digital-platform cases involving search, marketplaces and app ecosystems.

3.3 Self-Preferencing

Suppose a registry operates both:

1. an independent model marketplace; and
2. proprietary AI models.

It could theoretically give its own models:

• higher rankings;
• better search visibility;
• preferred API integration;
• cheaper deployment;
• superior documentation;
• default placement;
• preferential certification.

The registry would then simultaneously act as market infrastructure and competitor.

4. Refusal to Deal and Access Restrictions

A particularly important concern is refusal to provide access.

A dominant registry could:

• refuse to list competing models;
• remove competitors without objective criteria;
• deny API access;
• restrict model metadata;
• prevent downloads;
• block interoperability;
• refuse authentication;
• prevent migration to competing registries.

Under China's 2026 platform guidance, refusal to deal can include closing interfaces, interrupting data sharing, restricting traffic, and controlling essential data, models or application platforms without legitimate justification.

However, mere ownership of a registry does not automatically establish an antitrust duty to deal. Market power, indispensability, competitive effects and legitimate business justification remain important.

5. Tying and Bundling

A registry operator could potentially condition registry access upon purchase or use of another service.

For example:

"Models can be registered only if inference is purchased from our cloud platform."

Other possible bundles include:

• registry + cloud compute;
• registry + proprietary inference API;
• registry + storage;
• registry + cybersecurity;
• registry + monitoring;
• registry + proprietary AI assistant.

This may disadvantage independent cloud providers or inference platforms.

The competition question becomes whether the registry and the tied service constitute separate products and whether the arrangement forecloses competitors.

6. Exclusive Dealing

A dominant registry might offer developers:

• lower fees;
• enhanced visibility;
• premium certification;
• better API access;

in exchange for an agreement not to list models elsewhere.

This can be particularly problematic where developers need the registry to reach a large proportion of potential customers.

The economic effect may be:

Exclusive registry participation → reduced multi-homing → fewer models available elsewhere → weaker rival registry → increased registry dominance

7. Interoperability and API Lock-In

A registry can create technical dependency by controlling:

• proprietary APIs;
• authentication mechanisms;
• model metadata;
• deployment formats;
• SDKs;
• model adapters;
• evaluation protocols.

If developers incur substantial costs to move from Registry A to Registry B, the registry may obtain significant switching-cost power.

China's platform guidance specifically recognizes interfaces, data sharing, algorithms and platform rules as relevant competition considerations.

8. Data and Feedback-Loop Advantages

A dominant registry can potentially accumulate enormous quantities of:

• model usage statistics;
• downloads;
• benchmark results;
• developer behavior;
• model performance data;
• error reports;
• user feedback;
• fine-tuning information.

This can create a feedback loop:

More models → more usage → more data → better ranking/evaluation → more users → more models

A competing registry may therefore face an informational disadvantage even if it possesses comparable technical infrastructure.

9. Benchmark Manipulation

Registry-controlled evaluation can become a competition bottleneck.

A registry might control:

• benchmark selection;
• evaluation methodology;
• safety scores;
• performance rankings;
• latency measurements;
• accuracy classifications.

If developers and customers treat registry scores as industry standards, the registry could effectively influence market reputation.

The antitrust concern becomes stronger where the registry:

1. evaluates competing models;
2. owns competing models; and
3. uses evaluation methodology that systematically advantages its own models.

10. AI Model Registry and Essential-Facility Concerns

The essential-facility doctrine may become relevant where a registry becomes practically indispensable.

The analysis normally asks questions such as:

1. Is the registry controlled by a dominant undertaking?
2. Is access objectively necessary to compete?
3. Is duplication realistically possible?
4. Can alternative registries provide meaningful access?
5. Does refusal eliminate or substantially weaken competition?
6. Is there a legitimate justification for refusing access?

This does not mean every successful AI registry becomes an essential facility.

The legal threshold is generally high, particularly in U.S. antitrust law.

11. Six Major Case Laws

Case 1: United States v. Microsoft Corp., 253 F.3d 34 (D.C. Cir. 2001)

Facts

Microsoft possessed substantial market power in Intel-compatible PC operating systems and engaged in conduct concerning Internet Explorer and distribution channels.

The court examined Microsoft's contractual and technological restrictions on competing browsers.

Principle

The case demonstrates that monopoly power can be unlawfully maintained through conduct that restricts the ability of competing products to reach users.

Application to AI Registries

An AI registry could raise analogous concerns if it:

• prevents competing registries from accessing developers;
• imposes restrictive contractual conditions;
• technically disables interoperability;
• uses its platform position to exclude competing model-distribution services.

The important lesson is that control over distribution infrastructure can have competitive significance even where the infrastructure is not itself the ultimate product being sold.

12. Case 2: Google Android / Google and Alphabet v European Commission

The EU Google Android litigation concerned Google's contractual restrictions involving Android, app stores, search and competing Android forks.

In 2026, the Court of Justice's judgment continued to examine contractual restrictions, tying, exclusive pre-installation payments and obstruction of Android-fork development and distribution.

Principle

A platform operator can create competition problems by using control over one layer of an ecosystem to restrict competition at another layer.

AI Registry Application

An AI registry could potentially use:

Registry dominance → preferential model deployment → proprietary inference → exclusion of competing inference providers

or:

Registry access → mandatory proprietary API → reduced interoperability

The Android precedent is particularly relevant because AI ecosystems increasingly contain interconnected layers rather than one isolated product.

13. Case 3: Bronner v Mediaprint, C-7/97

Facts

The case concerned access to a newspaper home-delivery distribution system.

The European Court of Justice adopted a demanding approach to compulsory access.

Principle

Dominance does not automatically create an obligation to share infrastructure with competitors.

The facility generally needs to be sufficiently indispensable, and refusal must be capable of eliminating effective competition rather than merely making competition more difficult.

AI Registry Application

This is highly relevant to AI registries.

A model registry should not automatically be treated as an essential facility merely because it is large.

A competitor might instead be expected to develop:

• its own registry;
• alternative model-hosting infrastructure;
• independent APIs;
• alternative evaluation systems.

However, if a registry becomes practically indispensable because virtually all commercially important models, developers and users depend upon it, the Bronner framework becomes more significant.

14. Case 4: IMS Health GmbH & Co. KG v NDC Health, C-418/01

Facts

IMS Health controlled a data structure used by pharmaceutical companies for regional pharmaceutical sales information.

Competitors sought access to the structure.

Principle

The case established important criteria concerning refusal to license intellectual-property-related infrastructure where access may be indispensable for competition.

The Court emphasized exceptional circumstances before imposing compulsory access.

AI Registry Application

An AI registry could similarly develop proprietary:

• model metadata structures;
• model-identification systems;
• interoperability formats;
• evaluation architectures;
• proprietary model taxonomies.

If industry participants become dependent upon a registry's proprietary architecture, competitors may argue that access is necessary to compete.

The case therefore provides an important framework for distinguishing legitimate intellectual-property protection from potentially exclusionary control over indispensable infrastructure.

15. Case 5: Aspen Skiing Co. v Aspen Highlands Skiing Corp., 472 U.S. 585 (1985)

Facts

Aspen Skiing involved cooperation between competing ski operators and a later decision by the dominant operator to terminate a previously profitable cooperative arrangement.

Principle

The U.S. Supreme Court treated the withdrawal of cooperation as significant in the context of the dominant firm's previous willingness to deal and the absence of an apparent legitimate business justification.

AI Registry Application

Consider a registry that historically allowed independent AI developers to:

• access APIs;
• list models;
• share metadata;
• use interoperability tools.

If it suddenly withdraws access specifically after those developers become serious competitors, the historical pattern of cooperation may become important evidence.

The case therefore highlights the importance of examining changes in conduct, rather than merely asking whether a dominant firm refuses access.

16. Case 6: Verizon Communications Inc. v. Law Offices of Curtis V. Trinko, 540 U.S. 398 (2004)

Principle

The U.S. Supreme Court emphasized that antitrust law generally does not require dominant companies to assist competitors.

The Court was concerned that excessive compulsory-sharing obligations could reduce incentives to innovate and invest.

AI Registry Application

This creates an important counterweight to essential-facility arguments.

An AI registry operator could argue:

• it invested in infrastructure;
• it developed proprietary security systems;
• it bears hosting costs;
• unrestricted access could create cybersecurity risks;
• compulsory sharing could reduce innovation incentives.

Therefore, a claim that an AI registry is "monopolized" would require substantially more than simply showing that competitors would benefit from access.

17. Case 7: Google Search (Shopping) — European Commission / General Court

The Google Shopping litigation concerned the treatment of Google's comparison-shopping service in its general search results.

Principle

The case demonstrates how control over an important digital access point can affect the competitive position of downstream services.

AI Registry Application

A dominant AI registry could similarly function as a discoverability gateway.

If its ranking system preferentially promotes:

Registry's own AI model → competing models receive lower visibility

the competitive problem may not be the registry's ownership itself but the use of its gateway position to disadvantage competing models.

This makes ranking algorithms particularly important.

18. Case 8: SAMR v Alibaba (2021)

China's Alibaba enforcement is particularly relevant to AI registry analysis.

SAMR found Alibaba's "choose one from two" exclusivity conduct contrary to China's Anti-Monopoly Law. The enforcement action formed part of China's broader shift toward stronger scrutiny of digital-platform conduct.

Principle

Platform dominance can be reinforced through contractual restrictions that limit merchants' ability to use competing platforms.

AI Registry Application

A similar issue could arise where an AI registry tells model developers:

"To receive premium ranking, certification or commercial access, you must not list your model on competing registries."

The relevant economic concern would be foreclosure of rival registries.

19. Case 9: CNKI Antitrust Enforcement

China's enforcement against CNKI is also instructive because CNKI's market position involved extensive control over an important information-resource platform.

Relevance

An AI registry may similarly accumulate control over:

• model information;
• metadata;
• evaluation records;
• version histories;
• research materials;
• model documentation.

Where competitors depend upon such information, restrictions on access can become competition concerns.

The analogy is especially important where the registry evolves from merely being a catalogue into a critical information infrastructure.

20. China-Specific Legal Framework

China's Anti-Monopoly Law can address AI registry conduct through several categories.

Article 22-type abuse concerns

A dominant operator may face scrutiny for:

• unfair pricing;
• predatory pricing;
• refusal to deal;
• exclusive arrangements;
• tying;
• unreasonable conditions;
• discriminatory treatment.

China's platform-economy guidelines expressly recognize these categories.

Platform-specific factors

Relevant factors include:

• market share;
• ability to control the market;
• financial and technological conditions;
• network effects;
• switching costs;
• access to data;
• barriers to entry;
• dependence of business users.

21. 2026 Chinese Platform Guidance and AI Registries

The 2026 SAMR Internet Platform Anti-Monopoly Compliance Guidelines are particularly significant for this subject.

They state that dominant platform operators should avoid unjustified refusal to deal, including:

• closing interfaces;
• restricting traffic;
• interrupting data sharing;
• delaying cooperation;
• using discriminatory algorithms;
• controlling essential data, models or application platforms. 

The Guidelines also address tying, unreasonable conditions and differential treatment.

This means that, in a Chinese AI-registry scenario, the model itself can potentially become an important object of competition analysis, rather than merely the registry software.

22. Self-Preferencing Risks

Suppose Registry A owns:

• Model A;
• Registry A;
• AI inference service A;
• cloud service A.

It could theoretically create a vertical chain:

Model → Registry → Inference → Cloud → Application

Competition concerns may arise if the operator gives its own model preferential treatment at each stage.

Potential conduct includes:

23. Killer Acquisition and Registry Consolidation

A major future concern is acquisition of competing registries.

Suppose a dominant registry acquires:

• a model evaluation platform;
• an independent model marketplace;
• a model-security certification company;
• an AI model-hosting company.

Even if each target has relatively low turnover, the transaction may eliminate an important future competitive constraint.

This is particularly relevant in AI because innovation and competitive significance may not correspond neatly with current revenue.

24. Data Advantage and Competitive Feedback

A dominant registry may have access to information about:

• which models developers download;
• which models are abandoned;
• which models receive the most API calls;
• which models perform best;
• which fine-tuning techniques succeed;
• which industries are adopting particular models.

If the registry also develops its own AI models, it could potentially use those insights competitively.

The competition question would concern whether the registry is using information obtained from dependent businesses to advantage its vertically integrated AI products.

25. Interoperability Remedies

Potential regulatory remedies could include:

1. API interoperability

Require standardized interfaces for competing model providers.

2. Data portability

Allow developers to export:

• model metadata;
• usage data;
• configuration;
• evaluation records.

3. Non-discriminatory listing

Require objective criteria for model admission and ranking.

4. Transparent ranking

Require disclosure of significant ranking criteria where appropriate.

5. Multi-homing

Permit model developers to list simultaneously on competing registries.

6. Non-exclusive contracts

Restrict unnecessary exclusivity requirements.

7. Separation remedies

In serious cases, structural separation between:

• registry;
• model development;
• cloud infrastructure;
• inference services

could be considered.

26. Defences Available to a Registry Operator

A registry operator could rely upon legitimate justifications.

For example:

Security

Rejecting malicious or unsafe models may be necessary.

Privacy

A registry may restrict access to protect confidential information.

Intellectual property

Proprietary technology may legitimately be protected.

Quality control

A registry may impose technical standards to maintain reliability.

Cybersecurity

Unrestricted API access could create security risks.

Cost recovery

Fees may reflect genuine infrastructure costs.

China's 2026 guidance recognizes legitimate reasons including protection of intellectual property, trade secrets, personal information, data security and network security.

Thus, not every restriction is anticompetitive.

27. Key Competition Tests

For an AI Model Registry monopolization investigation, the following sequence is useful:

Step 1 — Define the relevant market

Is the relevant market:

• model registries;
• model marketplaces;
• model hosting;
• AI inference;
• model evaluation;
• broader cloud services?

Step 2 — Establish market power

Consider:

• market share;
• network effects;
• switching costs;
• developer dependence;
• data advantages;
• interoperability barriers.

Step 3 — Identify exclusionary conduct

Look for:

• refusal to deal;
• tying;
• exclusive dealing;
• self-preferencing;
• discriminatory access;
• API restrictions;
• manipulation of rankings.

Step 4 — Analyze foreclosure

Determine whether rivals are actually or potentially prevented from competing.

Step 5 — Assess efficiencies

Consider:

• security;
• quality;
• innovation;
• privacy;
• interoperability;
• cost savings.

Step 6 — Consider remedies

Possible remedies include:

• access obligations;
• interoperability;
• non-discrimination;
• portability;
• transparency;
• behavioral restrictions;
• structural separation.

28. Hypothetical Example

Assume ModelHub X controls 75% of commercially relevant model-registry activity.

It operates its own foundation model, X-1.

ModelHub X then:

1. places X-1 at the top of every search;
2. requires competing models to pay additional certification fees;
3. prevents certified models from appearing on competing registries;
4. restricts API access to competing inference providers;
5. bundles registry access with X's cloud service;
6. uses registry usage data to improve X-1;
7. lowers the ranking of models that use competing clouds.

The legal issues could include:

Market dominance → self-preferencing → discrimination → exclusivity → tying → interoperability restriction → data leveraging → foreclosure

The relevant precedents would include Microsoft, Google Android, Bronner, IMS Health, Aspen Skiing, Trinko, and China's Alibaba enforcement.

29. Consolidated Case-Law Matrix

30. Conclusion

AI model registry monopolization is fundamentally a gateway-control problem. The greatest competition risk arises when a registry moves beyond being a neutral catalogue and becomes the infrastructure through which developers must obtain visibility, certification, distribution, API access, evaluation, deployment and customers.

The most important antitrust theories are therefore:

1. abuse of dominance;
2. refusal to deal;
3. essential-facility concerns;
4. self-preferencing;
5. exclusive dealing;
6. tying and bundling;
7. discriminatory access;
8. API/interoperability restrictions;
9. data leveraging;
10. algorithmic ranking manipulation;
11. anticompetitive acquisitions; and
12. leveraging registry power into adjacent AI markets.

For China specifically, the issue is becoming particularly significant because the current platform-antitrust framework expressly addresses algorithms, data, platform rules, interfaces, data sharing and control over essential models or application platforms.

The central legal distinction is nevertheless important: a large or commercially successful AI registry is not automatically a monopoly, and possession of market power does not by itself establish unlawful monopolization. The decisive analysis concerns the relevant market, substantial market power, specific exclusionary conduct, competitive effects, indispensability where access is demanded, and legitimate technological or business justifications.