1. Introduction

Multi-model routing AI platform dominance arises when a platform connects users, developers or businesses to multiple artificial intelligence models and controls how requests are allocated among them. Instead of relying on a single model, the platform may route each request to different foundation models depending on cost, latency, accuracy, safety, computational availability or commercial priorities.

For example, a business might use one AI platform to access several models for legal research, coding, customer service and document analysis. The platform decides which model receives each request, controls the application programming interface (API), manages billing and usage data, and may determine which models receive the most traffic.

This architecture can improve efficiency and competition. However, it can also create a new form of intermediation power: the platform may not own the underlying AI models but can nevertheless control their access to customers, their commercial viability and the conditions under which competing model providers participate.

Competition law must therefore distinguish between legitimate technical optimisation and conduct that unfairly excludes rivals, exploits customers or entrenches market power.

The central legal question is:

When does an AI routing platform become an indispensable gateway whose control over model selection, visibility, pricing and access substantially weakens competition?

The analysis below focuses on UK and EU competition law, supplemented by relevant US precedents. The cases provide legal analogies; they should not be read as judgments that have already established liability for multi-model AI routing itself.

2. Meaning and operation of multi-model routing

Users and business applications

Developers, enterprises, consumers and automated agents

Multi-model routing platform

API gateway · model selection · billing · ranking · identity · telemetry

Model A

Premium reasoning

Model B

Low latency

Model C

Specialist tasks

Illustrative architecture: the router controls the path between customer demand and competing models.

A multi-model routing platform may perform several functions:

Dynamic model selection: choosing a model for each request.

Ranking and recommendation: presenting certain models as preferred, safer or more reliable.

Price intermediation: combining model charges, platform fees and subscription pricing.

Data aggregation: collecting prompts, outputs, latency measurements, customer preferences and failure rates.

Switching and portability: controlling whether users can migrate workflows, histories, evaluations and application configurations.

Capacity allocation: distributing scarce GPU capacity, inference requests and premium service access.

Policy enforcement: filtering models, limiting their availability or applying common safety and usage conditions.

These functions can make the router more than a neutral technical intermediary. Its economic significance depends on whether users and model providers can realistically bypass it.

3. Legal framework

A. United Kingdom

The principal legal provisions are:

Competition Act 1998, Chapter II prohibition: prohibits abuse of a dominant position in a market affecting trade within the United Kingdom.

Competition Act 1998, Chapter I prohibition: may apply where competing platforms or model providers coordinate routing, prices, capacity or other commercially sensitive conduct.

Digital Markets, Competition and Consumers Act 2024 (DMCC Act): provides a separate digital-markets regime under which the Competition and Markets Authority (CMA) can designate firms with strategic market status in respect of digital activities and impose conduct requirements, subject to the statutory tests and procedures.

Merger control: may apply when a dominant router acquires a competing model provider, routing technology, specialist inference service or strategically important data asset.

Dominance alone is not unlawful. A finding of abuse requires an assessment of the platform's market position, conduct, effects and applicable legal standards. The DMCC Act's strategic-market-status regime is also distinct from a traditional Chapter II infringement finding.

B. European Union

The key provisions are:

Article 102 TFEU: prohibits abuse of a dominant position affecting trade between EU Member States.

Article 101 TFEU: prohibits anticompetitive agreements and concerted practices, including certain exchanges of commercially sensitive information.

EU Digital Markets Act (DMA): imposes specified obligations on designated gatekeepers for their designated core platform services. It does not automatically cover every AI router or AI model.

EU Merger Regulation: may apply to acquisitions that meet the relevant jurisdictional thresholds, with additional referral mechanisms potentially relevant in some cases.

The legal analysis must identify the relevant market, establish dominance or the applicable regulatory designation, and then examine the particular conduct. An AI router does not become dominant merely because it offers access to multiple models.

4. Principal dominance risks

4.1 Self-preferencing in model selection

A routing platform may own or be affiliated with one of the models available through its interface. It could systematically route profitable or high-value requests to its own model while directing less valuable requests to independent competitors.

Potential methods include:

Giving affiliated models preferred default status.

Assigning competitors lower visibility in model-selection menus.

Applying stricter latency, reliability or safety thresholds to rival models.

Using asymmetric performance evaluations that favour affiliated models.

Making independent models appear more expensive through platform-specific fees.

Routing user requests away from a rival even when the rival performs better for the task.

Not every preference is anticompetitive. Routing based on demonstrated quality, security, availability or total cost can be legitimate. The concern arises when a dominant platform uses its control of the gateway to disadvantage competing models without adequate objective justification.

4.2 Exclusionary routing and traffic foreclosure

Model providers need sufficient usage to recover the costs of training, deployment, inference infrastructure and continuing development. A router that controls a large share of customer demand can influence which providers obtain that usage.

Consider a platform that supplies 75% of enterprise demand for a particular class of AI service. If it directs most requests to two affiliated models and substantially reduces traffic to independent alternatives, the latter may lose the revenue and operational scale required to compete.

The relevant issues include:

Whether the router is an important or unavoidable distribution channel.

Whether rival providers have effective alternative routes to customers.

Whether the conduct materially reduces rivals' ability or incentive to compete.

Whether the routing decisions are objectively justified.

Whether consumers experience reduced choice, higher prices or lower innovation.

The 75% figure is purely hypothetical; market shares must be established with evidence and an appropriately defined market.

4.3 Tying, bundling and exclusivity

A platform may bundle routing with cloud hosting, productivity software, enterprise subscriptions, identity management or proprietary development tools. Bundling can create genuine efficiencies, but it may also make independent models commercially unattractive.

Examples include:

Requiring customers to use the platform's preferred model to obtain discounted cloud capacity.

Making access to advanced routing features conditional on purchasing affiliated AI services.

Offering substantial discounts only where customers route nearly all requests through the platform.

Restricting customers from using alternative routers.

Requiring developers to adopt platform-specific APIs that make switching costly.

These arrangements can raise tying, bundling, exclusivity or loyalty-rebate concerns, depending on market power, contractual structure and likely competitive effects.

4.4 Data advantages and information asymmetry

The router may observe requests and performance across many competing models. This can generate a valuable informational advantage over individual model providers.

For instance, the router may know which models perform best for particular customer segments, where rivals are weak, how sensitive users are to price changes and which applications are considering switching.

A platform that owns a competing model might use such information to improve its own commercial position. That creates possible concerns about:

Preferential access to aggregated performance data.

Use of rivals' confidential information for competitive purposes.

Discriminatory access to evaluation tools.

Exploitative contract terms imposed on dependent providers.

Reduced incentives for independent providers to innovate or participate.

Competition law does not automatically prohibit a platform from collecting operational data. The critical question is whether its use or allocation of that information constitutes unlawful exclusion, coordination or another legally recognised abuse.

4.5 Algorithmic opacity and discriminatory routing

Routing algorithms can change rapidly, making exclusion harder to detect. A platform might claim that a model was selected because of latency, safety or reliability when the underlying decision was also influenced by ownership, commercial rebates or strategic incentives.

A robust investigation should examine:

Model-selection rules and their version history.

The weights assigned to price, latency, quality and safety.

Differences in evaluation standards between affiliated and independent models.

Routing outcomes for comparable requests.

Internal communications and commercial incentive structures.

The effects of routing changes on rivals' traffic and customers' costs.

A black-box algorithm is not itself evidence of abuse. But unexplained, persistent and asymmetric treatment can justify closer scrutiny when combined with evidence of dominance and exclusionary effects.

4.6 Multi-homing barriers and ecosystem lock-in

Users may technically have access to several models while remaining commercially dependent on a single router.

The source of lock-in may be the accumulation of proprietary prompt templates, evaluation datasets, agent workflows, monitoring tools, security policies, billing arrangements and application integrations.

Consequently, nominal model choice may not translate into effective competitive choice. Switching to another router could require expensive redevelopment, retraining, compliance reviews and testing.

This is especially significant where the router controls both access to models and the tools necessary to build applications around them.

5. At least eight relevant case laws

The following decisions provide the principal doctrinal foundations for analysing these risks. Their application to multi-model routing is analogical unless the judgment expressly concerns the same conduct.

Case 1: Google Shopping

Case: Google and Alphabet v European Commission, Case C-48/22 P (Court of Justice of the European Union, 10 September 2024).

Legal principle: A dominant platform's preferential treatment of its own service can constitute an abuse where its conduct departs from competition on the merits and is capable of producing anticompetitive exclusionary effects.

Application to AI routing: If a dominant router consistently promotes its affiliated model while disadvantaging rival models through rankings, defaults or routing rules, the Google Shopping reasoning becomes relevant. The investigation should examine how the platform allocates visibility and demand, rather than merely asking whether competing models remain technically accessible.

The analogy is particularly strong where the router controls a gateway that rivals need to reach customers.

Case 2: Microsoft

Case: Microsoft v Commission, Case T-201/04 (General Court, 17 September 2007).

Legal principle: The case addressed Microsoft's refusal to provide certain interoperability information and its tying of Windows Media Player to Windows. It illustrates how control over a strategically important platform can be leveraged into adjacent markets.

Application to AI routing: A dominant router might use control over its API, agent framework or enterprise integration layer to disadvantage rival models. For example, withholding essential interoperability information or tying access to an important routing service to an affiliated model could raise similar concerns.

However, a duty to provide access is not automatic. The legal test depends on the specific conduct, market circumstances and applicable doctrine.

Case 3: Android

Case: Google and Alphabet v Commission, Case C-738/22 P (Court of Justice of the European Union, 2 July 2026).

Legal principle: The judgment concerned Android-related tying, contractual restrictions, payments linked to exclusive pre-installation and restrictions affecting the development and distribution of Android forks.

Application to AI routing: The case is relevant where a platform uses commercial agreements or ecosystem dependencies to steer customers toward its preferred AI model. Restrictions on alternative model providers, competing routers or compatible application frameworks may create similar leveraging concerns.

The key issue is whether contractual arrangements reinforce market power and restrict effective competitive alternatives, rather than whether a platform simply integrates multiple products.

Case 4: Bronner

Case: Oscar Bronner GmbH & Co. KG v Mediaprint, Case C-7/97 (Court of Justice of the European Communities, 26 November 1998).

Legal principle: The judgment established a demanding test for certain refusal-to-deal claims, including whether access to the requested facility is indispensable and whether refusal is likely to eliminate all competition by the applicant, subject to the precise legal conditions.

Application to AI routing: An independent model provider might argue that a dominant router is indispensable because it controls access to enterprise customers. The Bronner framework warns that being commercially important or more convenient is not necessarily enough to establish legal indispensability.

Investigators must assess whether realistic alternative channels exist, including direct APIs, competing aggregators, cloud marketplaces and customer-operated routing systems.

Case 5: Slovak Telekom

Case: Slovak Telekom v European Commission, Case C-165/19 P (Court of Justice of the European Union, 25 March 2021).

Legal principle: The Court addressed access to telecommunications infrastructure and the circumstances in which the strict Bronner indispensability test applies. It distinguished certain forms of access restriction from a refusal to supply where the undertaking had already provided access under regulated or contractual conditions.

Application to AI routing: If a platform already offers rival models access to its routing infrastructure but imposes discriminatory technical conditions, the precise form of conduct matters. The dispute may concern unfair access terms or discriminatory treatment rather than a complete refusal to deal.

This distinction helps prevent every dispute about an API, routing interface or model marketplace from being incorrectly treated as an essential-facilities case.

Case 6: United Brands

Case: United Brands Company v Commission, Case 27/76 (Court of Justice of the European Communities, 14 February 1978).

Legal principle: The case remains a foundational authority on dominance and abusive conduct, including discriminatory commercial conditions and the exploitation of market power.

Application to AI routing: A dominant platform might impose materially different fees, access conditions or technical requirements on comparable model providers without a sufficient objective justification. It might also use its intermediary position to impose unfair terms on dependent providers.

The relevant inquiry is whether the treatment is discriminatory or otherwise abusive under the applicable legal standard. Different prices alone do not prove an infringement.

Case 7: Intel

Case: Intel v Commission, Case C-413/14 P (Court of Justice of the European Union, 6 September 2017).

Legal principle: The Court required proper consideration of Intel's arguments and evidence concerning the capacity of its rebates to restrict competition, including the as-efficient-competitor analysis where applicable.

Application to AI routing: A platform might offer rebates to enterprise customers in return for routing most or all requests through its ecosystem. An investigation should evaluate the actual economic incentives, the share of demand covered, the duration of the arrangements and the ability of rivals to compete.

The Intel framework cautions against treating every rebate as unlawful without analysing its context and potential exclusionary effects.

Case 8: United States v Microsoft

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

Legal principle: The judgment addressed Microsoft's monopolisation of the PC operating-system market and conduct that restricted competitive threats, including restrictions affecting rival browser distribution.

Application to AI routing: A dominant AI gateway could use defaults, contractual restrictions, technical integration or distribution advantages to protect its position against competing routing services or model providers.

The case is particularly relevant where an incumbent platform can influence user behaviour through default settings and control of the surrounding software ecosystem.

The American monopolisation framework differs from EU and UK law, so its reasoning should be used comparatively rather than treated as a substitute for the applicable domestic test.

6. Additional relevant case laws

Case 9: Google Android — General Court judgment

Case: Google and Alphabet v Commission, Case T-604/18 (General Court, 14 September 2022).

Legal principle: The case concerned restrictions associated with Google's Android ecosystem, including arrangements involving app distribution, pre-installation and payments to device manufacturers.

Application to AI routing: It illustrates how defaults and ecosystem-level contractual arrangements can reinforce a dominant platform's position. In AI, similar concerns may arise where a router is integrated into an operating system, cloud environment or enterprise software suite and gives its own model a distribution advantage that rivals cannot effectively counter.

Case 10: Ohio v American Express

Case: Ohio v American Express Co., 585 U.S. 529 (2018).

Legal principle: The US Supreme Court considered competition in a two-sided transaction platform and emphasised the importance of assessing interactions between the platform's different customer groups.

Application to AI routing: A multi-model router may serve both developers seeking model access and enterprises seeking AI capabilities. Pricing, subsidies and incentives on one side can affect demand and competition on the other. The case supports careful analysis of platform economics rather than automatically treating each side as an independent market.

It does not establish a universal rule that all AI routing services constitute a single two-sided market.

7. Comparative legal analysis

These are potential legal characterisations, not automatic findings of liability. The applicable legal test depends on the jurisdiction, market position, contractual arrangement and evidence of competitive effects.

8. Market definition and dominance assessment

A central challenge is deciding which market the platform occupies. There are several plausible approaches.

Market A — AI model access and aggregation

Competition among services that provide unified access to multiple foundation models.

Market B — AI routing and orchestration

Competition among systems that select models, optimise inference and manage multi-model workflows.

Market C — Enterprise AI infrastructure

Competition among broader cloud, inference, deployment and governance providers.

Market D — Task-specific AI services

Competition for specialised workloads, such as coding, document processing, research or customer support.

These are hypotheses to test, not predetermined market definitions. Depending on customer substitution, competitive constraints and the platform's functionality, routing may be a distinct market or part of a broader service.

The assessment should examine:

Substitution: Can customers switch to direct model APIs or another router?

Market shares: What share of relevant demand, revenue or traffic does the platform control?

Entry barriers: Do network effects, switching costs, capital requirements or data advantages protect its position?

Countervailing power: Can large enterprises negotiate favourable terms or build their own routers?

Multi-homing: Do customers use several routers in practice, and at what cost?

Innovation constraints: Can a new model provider obtain enough customers to become a viable alternative?

A high market share is relevant but does not, on its own, prove dominance or abuse.

9. Evidence and enforcement challenges

AI routing creates distinctive evidentiary difficulties because the decision that determines which competitor receives a request may occur automatically and change continuously.

A. Relevant evidence

Competition authorities may need to examine:

Source code, routing policies and model-selection configurations.

Historical logs showing requests, selected models and routing outcomes.

Internal performance benchmarks and the treatment of affiliated models.

Contracts, rebates, exclusivity provisions and cloud-credit arrangements.

Documents explaining commercial objectives behind routing changes.

Data-access rules and safeguards separating competitors' confidential information.

Customer switching records and evidence of alternative distribution channels.

Counterfactual simulations showing how demand would be allocated under neutral routing.

B. Causation

A rival model's decline in traffic does not necessarily prove foreclosure. It might have become less accurate, more expensive or less reliable.

A stronger case would demonstrate that similarly situated rival models were treated less favourably, that the difference was not adequately explained by objective factors, and that the treatment was capable of materially weakening competition.

Authorities should also distinguish short-term traffic losses from durable barriers to entry and innovation.

10. Remedies and regulatory safeguards

Remedies should target the demonstrated competitive harm while preserving legitimate technical optimisation.

A universal obligation to disclose source code or adopt identical routing rules would be disproportionate in many cases. Security, privacy, reliability and genuine quality differences may justify differentiated treatment. Remedies should therefore be tailored to the particular conduct and supported by evidence.

In the UK, the DMCC Act may provide additional ex ante tools where the statutory strategic-market-status framework applies. In the EU, the DMA can address specified conduct by designated gatekeepers within the scope of their obligations. Neither regime automatically subjects every multi-model AI platform to the same requirements.

11. Hypothetical case study

Suppose Nexus AI, a fictional platform, controls a substantial share of enterprise access to competing foundation models. It owns Model N and also routes requests to independent models A, B and C.

Nexus introduces a new routing algorithm. It claims to select models based on price, latency and task performance. However, internal testing suggests that Model B performs better for several important workloads at comparable cost, while the platform routes most of those requests to Model N.

At the same time, Nexus offers substantial cloud credits to enterprises that route at least 90% of their AI traffic through its platform. Independent routers cannot match the discount because they lack Nexus's cloud resources.

A competition investigation would ask:

Is Nexus dominant in a properly defined market, or subject to an applicable digital-markets designation?

Does its routing algorithm favour Model N beyond what quality, price or security differences justify?

Do the cloud credits effectively discourage customers from using competing routers?

Can Model B reach enterprise customers through viable alternative channels?

Does the conduct materially reduce competition, innovation or customer choice?

Are there proportionate technical or commercial justifications for the arrangements?

If the evidence established unjustified discriminatory routing and exclusionary incentives by a dominant undertaking, the conduct could potentially engage Chapter II of the UK Competition Act 1998 or Article 102 TFEU, subject to the relevant legal requirements. Separate digital-markets obligations might also apply if the platform and service fall within their scope.

If, by contrast, Nexus's routing decisions genuinely reflect superior performance and the credits produce verifiable efficiencies without unlawful foreclosure, the conduct may be lawful.

12. Conclusion

Multi-model routing creates an important competition-law problem: a platform can control the commercial pathway between customers and AI models without owning the underlying models themselves.

The principal risks are self-preferencing, exclusionary traffic allocation, tying, loyalty incentives, discriminatory access, misuse of commercially sensitive data and switching-cost-based lock-in. The most relevant case-law foundations include Google Shopping, Microsoft, Google Android, Bronner, Slovak Telekom, United Brands, Intel and United States v Microsoft, supplemented by Ohio v American Express for platform-market analysis.