Artificial intelligence and competitive dynamics in downstream markets

3. AI-related competition concerns in downstream markets

AI can potentially alter the basic dynamics of competition by increasing market transparency and accelerating competitors’ reaction times, with possible implications for collusion, exclusionary conduct, and ultimately for consumer welfare as firms can react faster and with more accuracy than in the past (OECD, 2023[90]; OECD, 2021[94]; Hagiu and Wright, 2025[65]). Many of these competition concerns will be “traditional”, where familiar theories of harm apply in new technological settings, but there will also be “emerging” concerns specific to how AI systems operate. The following sections will first address traditional enforcement issues that may arise when competitors use AI, before turning to AI-specific challenges such as attribution of liability.

As the deployment of AI systems is still comparatively new, their effects on competitive dynamics and potential anti-competitive behaviour are still emerging, with very few actual antitrust cases known at the time of writing. Hence this chapter is intended to provide an outline of likely competition concerns that NCAs may face, rather than an authoritative discussion of observed anti-competitive conduct.

While algorithmic conduct is not new in competition enforcement – particularly in relation to pricing algorithms, collusion risks, and hub-and-spoke arrangements – GenAI introduces new dimensions that may warrant distinct consideration. Traditional algorithms typically execute predefined rules or optimise within narrow parameters, whereas generative AI systems are capable of producing novel outputs, adapting to user inputs, and learning from feedback and user prompts in ways that may be less predictable and harder to audit (Mayer Brown, 2024[95]; OECD, 2023[90]; FTC, 2023[63]; Carugati, 2024[96]). This capacity for autonomous content generation and strategic adaptation may complicate the assessment of intent, attribution, transparency and predictability in enforcement settings. For example, while algorithmic collusion may involve observable pricing patterns or shared data inputs, GenAI conduct may manifest through dynamic content generation, evolving recommendation systems, personalised offers or autonomous decision-making that changes over time and influences market behaviour in less transparent ways.

The potential for firm- or sector-specific fine-tuning also introduces heterogeneity in how GenAI is deployed, which may affect market outcomes and complicate the assessment of competitive harm (Chapter 2). Firms can fine-tune models to proprietary data or sector-specific needs, potentially enhancing differentiation but also raising questions about access, interoperability, and control (Bostoen, 2025[97]; Bostoen and van der Veer, 2024[98]). Accordingly, some of the parameters of interest may be the degree of model customisation, the governance of model outputs, and the extent to which generative systems are accessed or rather embedded. The factors influence firms’ strategic behaviour and may also alter the dynamics of interactions in ways that differ from more traditional algorithmic systems (Hofmann and Lorenzoni, 2023[99]; Alvarez and Marsal, 2025[100]; CMA, 2024[67]).

3.1. Competition enforcement concerns

Competition concerns in AI-enhanced downstream markets include risks of horizontal co-ordination as well as unilateral conduct: issues that are typically addressed under existing competition law frameworks in most jurisdictions. While AI technologies may alter the mechanisms through which such conduct manifests, many of the underlying concerns are familiar from traditional market as well as, more recently, from digital markets. Some of these issues have been the subject of extensive analysis by NCAs and by the OECD Competition Committee. For instance, algorithmic pricing tools have raised new questions about collusion, including hub-and-spoke arrangements and tacit co-ordination, but these are often extensions of known theories of harm such as price-fixing and information exchange (OECD, 2017[101]; 2021[102]; 2021[94]; 2023[90]; 2025[103]). This section will briefly summarise some of the more familiar or potentially expected competition concerns that NCAs may encounter in downstream AI markets.

3.1.1. Horizontal co-ordination and algorithmic collusion

AI systems increase both transparency and speed, with ambiguous effects. On the one hand, the ability of AI to provide rapid access to competitor data can increase efficiency. On the other, AI can artificially increase market transparency; if combined with ability of algorithms to react instantaneously to competitors deviations, the use of AI can eliminate incentives to deviate from explicit or tacit arrangements and stabilise co-ordination (OECD, 2023[90]; OECD, 2025[103]). Algorithmic pricing tools, including those powered by GenAI, may, under certain conditions, facilitate co-ordination even without explicit communication between competitors (Autorité de la concurrence and Bundeskartellamt, 2019[104]; JFTC, 2025[105]; OECD, 2025[103]). Rather than businesspeople meeting in a smoke-filled room or on the golf course, now AI-powered algorithms are capable of aligning expectations and responses, to stabilise prices or market shares and reduce incentives or opportunities to deviate: enhanced observability and predictability of rivals’ behaviour can soften competition and sustain supra-competitive outcomes. By continuously monitoring market data, AI-driven algorithms may align prices or outputs across firms, stabilising oligopolistic equilibria and reducing incentives to deviate (OECD, 2025[103]; OECD, 2023[90]). The 2025 OECD report on algorithmic pricing in G7 jurisdictions distinguishes three settings – algorithm-facilitated explicit collusion, hub-and-spoke co-ordination, and autonomous algorithmic collusion – each with different implications for intent and liability.

Explicit collusion and hub-and-spoke

Explicit collusion facilitated by algorithms is where firms use AI systems or algorithmic tools to co-ordinate prices, market shares or other strategic variables in a manner that would traditionally be considered a cartel. While the underlying conduct remains subject to established legal standards, the use of algorithmic tools may alter how collusion is implemented or detected. For example, firms may use shared pricing algorithms, common optimisation software, or third-party platforms to align behaviour without direct communication (OECD, 2025[103]). Although the legal framework for assessing explicit collusion remains largely unchanged, enforcement agencies may face challenges in identifying the locus of decision-making, especially where AI-driven algorithmic tools are outsourced or embedded in broader service contracts. As such, competition authorities may need to scrutinise the role of intermediaries, the transparency of algorithmic logic, and the contractual relationships between firms and technology providers (OECD, 2025[103]).

The risks of algorithmic collusion are expected to be greater where firms use the same pricing software or rely on a common model provider (hub-and-spoke arrangements). The OECD (2025[103]) identifies several enforcement examples, including the by now well-known online posters cases (CMA’s Trod/GBE 2016)1 and the United States’ Amazon Marketplace posters (U.S. DoJ, 2015[106]), where competitors employed identical repricing software and pricing algorithms to maintain agreed margins on online marketplaces. In such settings, the algorithm serves as an instrument executing a cartel.

Hub-and-spoke co-ordination, where a common vendor or platform facilitates collusion on behalf of multiple rivals, can likewise establish a concerted practice if competitors are – or should be – aware that shared algorithms align their behaviour. The EC’s Guidelines on the applicability of Article 101 TFEU to horizontal co-operation agreements state that when competitors subscribe to the same third-party pricing tool, and the tool uses commercially sensitive information from competitors, this may result in an unlawful information exchange (European Union, 2023[107]). The JFTC identified similar distinctions and enforcement implications under the Japanese Antimonopoly Act. If firms using the same algorithm share a common understanding that it will co-ordinate prices – despite no direct communication between the competitors – this may constitute a violation of Japan’s Antimonopoly Act (JFTC, 2021[108]).

Tacit algorithmic co-ordination

AI pricing tools can reproduce traditional co-ordination in digital form. Even without explicit agreement, it is – so far hypothetically – possible that GenAI with self-learning or reactive algorithms may stabilise prices in oligopolistic markets. By continuously observing market data, AI-driven algorithms can infer rivals’ reactions and converge towards joint profit-maximising equilibria. However, some jurisdictions recognise that mere parallel behaviour without any kind of agreement or contact between competing companies may not constitute an infringement of antitrust laws (OECD, 2025[103]). In the EU, for instance, the EU Court has established that an agreement prohibited under Article 101 TFEU requires some form of communication and sense of mutual commitment (“a meeting of minds” or a “concurrence of wills”)2. Companies have the right to adapt themselves to the existing or anticipated conduct of their competitors, as long as they determine their policies independently. However, the question here is rather whether the firms’ AIs, acting independently with an objective of maximising profits, may lead to a co-ordinated outcome. In this case the analytical challenge is evidentiary – identifying intent and foreseeability.

3.1.2. Unilateral conduct

AI can magnify unilateral conduct by dominant firms, including exclusionary and exploitative behaviour. The theories of harm will still correspond to “classic” abuses of dominance: exclusion of rivals or exploitation of consumers and downstream markets through unfair prices, terms or discrimination. Here, it is the relationship with upstream FMs via APIs or plug-ins, that may facilitate the unilateral conduct.

Exclusionary conduct

Digital platforms act as curators of online options, with algorithms that determine the visibility of content and products. Because consumers tend to rely heavily on these rankings, platforms acquire significant influence over consumer choice. This power can be abused when the platform is both an intermediary and a competitor: ranking algorithms may systematically favour the platform’s own products or services, thereby excluding rivals (Bostoen, 2025[97]). Skewed rankings distort competition by marginalising competitors and misleading consumers, particularly where promoted results are not the most qualitative. AI-driven marketplaces using extensive platform data may be even better at tailoring these ranking to favour their own brands. Similar concerns have been raised over Amazon’s Buy Box and algorithmic control of third-party seller visibility.3

Dominant digital platforms may design AI systems – search, recommendation, or generative interfaces – to favour their own products or affiliated services, acting as “choice architects”, guiding firm and consumer behaviour. If such agents systematically favour in-house products, rivals may be excluded despite superior offerings, distorting downstream competition and consumer choice (Autorité de la concurrence, 2024[70]; CMA, 2024[67]; JFTC, 2025[105]). Some authors find that such optimisation can emerge endogenously when algorithms maximise engagement or conversion metrics, even absent explicit exclusionary intent (Hagiu and Wright, 2025[65]; Rafieian and Yoganarasimhan, 2023[25]). Where rivals’ visibility is reduced through algorithmic optimisation rather than explicit code, competition authorities face new evidentiary challenges in establishing intent and causality (Bostoen, 2025[97]). The legal assessment would have to be effects-based: if the self-learning system’s operation results in the systematic foreclosure of competing products or services, and cannot be objectively justified by relevance or quality criteria, it may constitute an abuse of dominance (Autorité de la concurrence, 2024[70]; Bostoen, 2025[97]; OECD, 2025[103]).

Dynamic and personalised pricing

The adoption of AI significantly expands firms’ ability to implement high-frequency dynamic pricing and personalised offers, including in sectors where such practices were previously impractical. By processing behavioural and transactional data at scale, AI can estimate willingness-to-pay with precision, enabling granular price discrimination across consumers (OECD, 2024[50]; Tucker, 2024[109]). While such practices can in theory increase efficiency, in practice they often reduce consumer welfare, and may foreclose rivals lacking equivalent data capacity (see for instance, (Ezrachi and Stucke, 2016[110]; OECD, 2021[94]; Tucker, 2024[109]).

Reinforcement-learning algorithms can continuously test consumers’ willingness to pay, capturing surplus or targeting discounts to deter rivals (OECD, 2025[103]). AI-enabled personalisation and dynamic pricing strategies may facilitate exploitative or exclusionary conduct, particularly when deployed by dominant firms with access to granular consumer data to raise firms’ predictive accuracy. By enabling firms to predict consumer preferences with precision, AI can transform pricing into a strategic tool for differentiation and ultimately rent extraction. In this sense, AI shifts pricing from a traditional efficiency mechanism into a strategic moat, enabling more sophisticated discriminatory practices that competitors without equivalent data cannot replicate (Stucke and Ezrachi, 2024[111]).

In downstream markets, this could manifest in airlines, insurance, or online retail, where AI systems infer consumers’ willingness to pay and set prices accordingly. For instance, while searching for a flight, prices across websites start adjusting. Such services are likely to increasingly rely on Agentic AI, potentially further eroding consumer surpluses. Shared models and datasets may further drive convergence in pricing and business strategies (JFTC, 2025[105]). However, competition law does not always treat price discrimination of final consumers as a relevant conduct.

Overall welfare effects may depend on market power and transparency (Rafieian and Yoganarasimhan, 2023[25]). The CMA warns that firms may use AI systems to personalise offers to customers, which could be used strategically to exclude competitors. For example, incumbents might target customers most likely to switch to a rival, making it harder for new entrants to gain traction, especially in concentrated markets, helping reinforce market power. The CMA also highlights that AI-enabled personalisation, including pricing, can be harmful if it unfairly targets vulnerable consumers or has unfair distributive effects. It notes that such practices are difficult to detect and may undermine consumer trust and choice (CMA, 2024[67]).

Predatory pricing through AI-driven systems

AI-enabled pricing systems may allow firms to adjust prices with greater speed and granularity, including across products, regions or consumer segments. In theory, machine-learning models could be used to identify competitive threats and respond with targeted price reductions, potentially below cost. In such AI-driven markets – such as online retail, advertising or ride-hailing – automated price-matching may compress reaction times and reduce incentives to undercut, potentially stabilising prices. Some studies also suggest that shared models or data sources could lead to similar pricing strategies across firms, though further analysis is needed to assess the competitive implications (JFTC, 2025[105]).

Such AI-based pricing systems can sustain below-cost pricing for longer periods than human-driven strategies, thanks to algorithmic optimisation and access to granular demand data. This raises concerns that AI-based predation could become both more effective and less detectable, as prices adapt instantaneously to rivals’ reactions (Harrington, 2018[112]; OECD, 2023[90]; 2025[103]). Predatory pricing is traditionally difficult to prove. The use of automated systems risks further complicate this. AI can better maintain precision targeting of the below-cost price that would limit losses during the predation phase and accelerate recoupment once rivals exit, potentially even making predation a rational strategy (Rafieian and Yoganarasimhan, 2023[25]). The Competition Commission of India also warns of the risk of algorithmic unilateral conduct that could resemble predatory pricing (CCI, 2025[113]).

Foreclosure and leveraging

Vertical integration between layers of the AI value chain or also within the AI “stack”, such as compute infrastructure, model development, and downstream deployment, can generate substantial efficiencies. Integration may lower transaction and co-ordination costs, improve interoperability and technical performance, and strengthen incentives to invest in safety, innovation, and quality across complementary layers (Hagiu and Wright, 2025[65]; CMA, 2024[67]).

However, vertical integration in AI markets may also enable firms to leverage market power across layers of the AI stack, with significant implications for downstream competition. Control over essential inputs – such as compute, proprietary data, and model interfaces such as APIs – can be used to privilege affiliated services, restrict interoperability, or impose discriminatory conditions (ACCC, 2025[93]; Autorité de la concurrence, 2024[70]; CMA, 2024[67]). These risks are amplified when downstream firms depend structurally on integrated providers for access to models, cloud infrastructure, or APIs, creating lock-in and raising switching costs, especially when AI copilots are bundled with cloud services and productivity software in downstream firms (OECD, 2021[102]; 2025[3]; Vipra and Korinek, 2023[62]; Vipra and Korinek, 2023[62]; Autorité de la concurrence and Bundeskartellamt, 2019[104]).

Among these inputs, data access, discussed in Chapter 2, may be a critical bottleneck affecting downstream firms’ ability to compete and innovate if there is vertical integration.4 Proprietary datasets are likely to confer durable advantages to their owners, and without portability or interoperability, there is a risk that this could prevent challengers from entering or scaling in some downstream markets (OECD, 2025[68]; Autorité de la concurrence, 2024[70]). Platforms’ control over behavioural signals may also be leveraged, also through tying and bundling practices, into adjacent markets such as advertising, search, and content, reinforcing incumbency (ACCC, 2025[93]).5 These dynamics reflect a broader theory of harm, i.e. input foreclosure, where a vertically integrated firm restricts rivals’ access to essential inputs – whether data, compute, or model interfaces – undermining their ability to offer substitutable services or innovate independently.

Enforcers are beginning to respond to this issue. Assistant Attorney General Gail Slater, in her Fordham address, warned that “Google cannot be permitted to leverage its dominance in general search to the GenAI product space,” emphasising the scale advantages of data and the likelihood that these dynamics will intensify with AI (U.S. DoJ, 2025[114]). Assessing vertical conduct in AI markets therefore requires distinguishing verifiable efficiencies from strategic integration that undermines contestability, consistent with the effects-based analysis adopted in recent enforcement practice (OECD, 2025[68]; CMA, 2024[115]; Krakowski, Luger and Raisch, 2022[26]). Refusals to grant access to essential AI inputs may constitute exclusionary conduct when such inputs are indispensable for effective competition and access to competitively critical data appears to be essential for AI competition:

The prospect of GenAI expands the competitive utility of data across every industry (..) Now every industry is a data industry. (…) The finite amount of publicly available data also means that there will be more and more of a premium on developing proprietary sources of data. Access to high-volume, high-quality, proprietary data is a significant competitive advantage that could prove extremely difficult for competitors to surmount. Strong demands for data may drive the forces of consolidation. For example, vertical integration may become more attractive, especially in industries where downstream businesses may have access to valuable and sensitive data like healthcare data. Assistant Attorney Gail Slater (U.S. DoJ, 2025[114])

Alongside data, compute capacity represents another vertical restraint or bottleneck. High-performance compute (HPC) and cloud hosting are concentrated in a small number of global providers (also known as hyperscalers), which may tie model access to their infrastructure services, raising switching costs (OECD, 2025[3]; OECD, 2025[68]). The CMA has identified these as choke points, warning that vertically integrated providers controlling both compute and applications may discriminate against rivals (CMA, 2024[67]). Such providers may try to attract complementary innovation from downstream users, while retaining control over their compute or APIs. This would replicate the dynamics observed in mobile operating systems (where app developers because dependent on the operating system). (Hagiu and Wright, 2025[65]) describe this as “cross-layer leveraging”, where dominance in one segment (e.g., cloud or model access) extends across the stack. This could lead to systemic lock-in that would reduce the ability to multi-home and raise switching costs for users and developers.

3.2. An emerging concern: the attribution of liability

While some of the competition concerns raised by AI will resemble traditional theories of harm and will probably be solvable with the existing competition toolbox, the rise of AI systems introduces novel challenges for competition enforcement, particularly in attributing conduct and establishing intent. Unlike traditional forms of co-ordination or exclusion, algorithmic and AI systems may produce anti-competitive outcomes autonomously, without direct human instruction or explicit agreement. Autonomous optimisation challenges traditional enforcement paradigms, requiring a shift in how responsibility is attributed.

AI systems can learn profit-maximising or exclusionary strategies through reinforcement learning or repeated interaction, even in the absence of deliberate co-ordination. This undermines conventional legal notions of intent and agreement. As the Autorité de la concurrence and Bundeskartellamt joint report on algorithms (2019[104]) noted, proving intent may be less relevant when competitive harm arises regardless of human intention. The report acknowledges the difficulty of attributing responsibility in cases where algorithms operate as opaque “black boxes,” noting that “the complexity and opacity of algorithms may make it difficult to determine whether a certain conduct is anticompetitive and whether it can be attributed to a specific undertaking” (p. 24).

Compounding these challenges is the issue of explainability and accountability. Agentic AI further complicates the ‘black box’ problematic by continually updating parameters and adapting behaviour (Belcak et al., 2025[116]; Desai and Riedl, 2025[89]). This could potentially have procedural consequences: a lack of explainability might impede effective enforcement. Some authors suggest that such opacity may justify ex-post audit rights or mandatory disclosure of documentation that spans all the elements of the AI system, including, but not limited to, the models, its training data and algorithms, rather than new substantive prohibitions (Hagiu and Wright, 2025[65]). Another possibility would be technical co-operation between authorities and developers to facilitate interpretability (OECD, 2025[103]). Opacity does not create new offences, but it can hamper the attribution of accountability. Enforced transparency and auditability may become essential to sustaining competition law enforcement in AI-driven markets (Hagiu and Wright, 2025[65]).

The OECD’s AI Principles underscore that AI systems should be designed so that (i) their logic, data and decision-making can be suitably explained (explainability) and (ii) affected parties have meaningful access to challenge, appeal or correct outcomes (contestability). The first ensures transparency; the second ensures that transparency is actionable. Without contestability, explanations risk being purely descriptive and not leading to meaningful oversight, accountability or market contestation (OECD.AI, 2025[117]).

From a competition-policy perspective, ensuring contestability means enabling suppliers, users, and third-party challengers to verify, question or override embedded algorithmic decisions that may raise barrier, lock-in or exclusion risks. Explainability alone is a necessary but not sufficient condition for contestability.

Hagiu and Wright (2025[65]) argue that competition authorities must adapt their frameworks to account for autonomous optimisation. The OECD has similarly recommended evaluating foreseeability and controllability – whether a firm could reasonably predict or prevent its algorithm’s exclusionary or collusive behaviour (OECD, 2023[90]; 2025[103]). Authorities may need to assess design choices, governance processes, and oversight mechanisms, rather than relying solely on evidence of explicit intent. The CMA’s 2021 report on algorithmic harms discusses how ineffective oversight of algorithmic systems can lead to consumer and competition harm. It highlights the need for regulators to develop audit tools and accountability frameworks, which could imply that firms may be held responsible for negligent deployment or lack of safeguards (CMA, 2021[118]).

The difficulties are likely to be further exacerbated with the rise of agentic AI, where autonomous agents may act on behalf of firms or individuals, or even independently, including in ways that resemble market co-ordination. This complicates attribution of harmful results, such as consumer welfare losses or exclusionary conduct. While agentic AI may foster pro-competitive dynamics in specific markets, it also raises forward-looking concerns around accountability, transparency, and the governance of interactions among autonomous systems – dynamics that are still not fully understood.

Agentic AI systems can raise complex liability questions, particularly when they act with a degree of autonomy that blurs the line between tool and actor. As these systems increasingly operate on behalf of firms or individuals – or even independently – the challenge lies in determining who is accountable for their outputs. Scholars highlight the need for governance mechanisms to mitigate risks linked to information asymmetry and discretion when such systems act on behalf of firms or individuals, or on their own volition (Kolt, 2024[119]; Desai and Riedl, 2025[89]). When agentic systems generate exclusionary or exploitative outcomes, the lack of direct human control complicates attribution, especially under legal frameworks that rely on intent or direct causality. This calls for a reassessment of liability standards, potentially shifting towards responsibility for design choices, oversight, and risk mitigation rather than traditional notions of intent.

Policy foresight exercises underline the potential systemic impact of agentic AI, but also its current limitations. The OECD’s AI Capability Indicators report notes, “Agentic systems typically perform below level 3, indicating significant limitations on AI’s ability to self-monitor and adaptively regulate its own reasoning” (OECD, 2025[120]). Nonetheless, the growing autonomy of agents could begin to extend their influence beyond narrow applications, reshaping online ecosystems and creating new intermediation layers between firms and consumers (Toner et al., 2024[121]). This suggests that, while agentic AI may foster contestability in specific markets, or even replace human agency in intermediation functions (such as travel agents), it also raises forward-looking concerns around accountability, transparency, and the governance of interactions among autonomous systems, which are still not fully understood.

3.3. Actions to support competition in AI-enabled downstream markets

To conclude the paper, this chapter discusses some of the levers available to policymakers and competition authorities to help ensure that AI-enabled markets remain competitive, drawing on OECD work and the experience of national competition authorities.

3.3.1. Enforcement actions

Enforcement remains a core tool for addressing anti-competitive conduct in AI-enabled downstream markets. While few cases have yet been concluded, several jurisdictions have initiated investigations or imposed remedies in adjacent markets with implications for AI.

Conduct cases

While there are no cases dealing specifically with AI, some NCAs have signalled the importance of monitoring how AI can be used by competitors not only for pro-competitive purposes but also to exploit consumers or exclude competitors. NCAs highlight the risks that can arise where AI is deployed in markets. These include, for instance, the risk that algorithms can allow competitors to share competitively sensitive information, fix prices, or collude on other terms or business strategies in violation of our competition laws; or the risk that algorithms may enable firms to undermine competition through unfair price discrimination or exclusion (CMA, EC, DoJ, FTC, 2024[122]). Similarly, the Competition Commission of India (CCI) in its recent market study also flagged algorithmic collusion and price discrimination as emerging risks in downstream sectors such as retail, logistics and e-commerce. Its 2025 market study identified self-learning algorithms and parallel pricing systems as potential facilitators of tacit co‑ordination, even in the absence of explicit agreements (CCI, 2025[113]). The Autorité de la concurrence and Bundeskartellamt (2019[104]) emphasised the need to scrutinise the effects of algorithmic systems, especially where they facilitate exclusionary or collusive outcomes without explicit co‑ordination. Similarly, the G7 Ministerial Declaration (2024[123]) called for adaptive enforcement tools capable of identifying discriminatory ranking, algorithmic collusion, or self-preferencing, even in the absence of direct human intervention.

In the United States, the Department of Justice’s remedies in the Google Search case required data sharing with AI competitors, recognising that search data are an essential input into generative AI models (U.S. DoJ, 2025[124]). Similarly, competition authorities in G7 jurisdictions have examined algorithmic pricing practices, including hub-and-spoke arrangements and tacit co-ordination, under traditional theories of harm (OECD, 2025[103]).

Enforcement may also be required to address bundling or tying of AI services with cloud infrastructure, particularly where access to compute is conditioned on the use of affiliated AI models (OECD, 2025[3]). The French Autorité de la concurrence has identified a range of risks associated with such practices in its sector inquiry into generative AI. Major digital firms, especially those vertically integrated across cloud infrastructure and AI model development, may leverage their control over compute to favour their own AI offerings, either through technical lock-in, preferential access, or exclusive agreements. The Autorité notes that such practices could constitute abuses of dominance, particularly where they restrict access to key inputs like GPUs, training data, or skilled labour, or where they involve tying AI services to cloud usage in ways that foreclose rivals or entrench incumbents (Autorité de la concurrence, 2024[70]). The CMA had similar findings, and proposed remedies to address switching barriers in cloud services, including obligations to improve interoperability, reduce egress fees, and prevent discriminatory treatment of third-party services. These remedies are particularly relevant in AI markets, where access to scalable compute is a prerequisite for model training and deployment (CMA, 2024[115]; CMA, 2023[46]).

Merger control

Merger control is also a critical tool for addressing structural risks in AI-enabled markets. Acquisitions can be a powerful way for digital incumbents to either become gatekeepers or strengthen their dominant position, possibly extending it to adjacent markets (Nicoletti, Vitale and Abate, 2023[125]). As highlighted in the CCI’s market study, acquisitions and partnerships involving AI firms may raise competition concerns, particularly where they involve exclusive access to data, compute, or intellectual property. Even minority investments can confer strategic influence over target firms, requiring careful scrutiny (CCI, 2025[113]). Similarly, exclusive cloud supply agreements, distribution arrangements, and IP licensing may create dependencies that foreclose rivals or entrench dominant positions (CCI, 2025[113]). Competition authorities have emphasised the need for case-by-case analysis to assess whether such transactions harm contestability or innovation. Existing merger control frameworks are normally equipped to address these risks, but the analytical framework might need to be broadened to incorporate assessments of data access, innovation incentives, and platform dependencies.

3.3.2. Competition advocacy and market monitoring

Competition enforcers need robust information on how AI-systems are reshaping competition dynamics. Given the rapid pace of AI adoption, market studies and monitoring tools are useful to assess emerging risks. These instruments allow authorities to gather evidence, consult stakeholders, and evaluate whether enforcement, regulatory adjustments or voluntary commitments are appropriate. The OECD and JFTC highlight the importance of proactive monitoring tools, such as sectoral mapping, adoption surveys and the use of AI-driven analytics to identify emerging patterns of market concentration, especially where conventional datasets lag behind market developments (JFTC, 2025[105]; OECD, 2024[50]).

To assess if AI adoption may give rise to structural risks, competition authorities could also consider launching market studies. These can explore, in greater depth, the implications of data control, interoperability restrictions or integration of AI assistants within digital ecosystems. Market studies are a flexible instrument: they allow authorities to gather evidence, consult stakeholders, and assess whether competition law enforcement, regulatory adjustments or voluntary commitments are most appropriate (OECD, 2018[126]). Several national competition authorities, including (but not limited to) the CMA, the French Autorité de la concurrence, the CCI, and the ACCC have already carried out such studies to shed light on competition issues in digital and AI-related markets, including foundation models, advertising, and digital ecosystems (CMA, 2023[92]; CMA, 2024[67]; ACCC, 2025[93]). These studies have identified structural risks such as data bottlenecks, interoperability restrictions, and integration of AI assistants into dominant platforms.

Competition advocacy also plays a critical role in shaping policy debates around AI and market design. Advocacy enables competition authorities to proactively highlight emerging risks and promote regulatory frameworks that support contestability without stifling innovation. OECD work has consistently emphasised the importance of interoperability standards, transparency obligations, and safeguards against discriminatory access to key AI inputs such as data, compute, and talent (OECD, 2023[127]; 2025[68]; 2021[128]). National authorities have also used advocacy to raise awareness and guide policy development. For example, the Portuguese Competition Authority has published a series of short papers on generative AI, warning of “competition risks regarding access and use of data,” the potential for “lock-in effects” from closed foundation models, and the impact of talent concentration and labour mobility restrictions on innovation and market entry (AdC, 2024[11]). These publications serve as early-stage policy interventions, aimed at shaping the conditions under which AI ecosystems evolve. More broadly, advocacy can take the form of public consultations, market studies, or guidance documents that inform legislative and regulatory processes, promote pro-competitive standards, and ensure that emerging AI markets remain open, interoperable, and innovation-friendly.

3.3.3. Regulation

Views differ across jurisdictions on the complementary role of regulation in ensuring competitive and open AI markets. In jurisdictions which have adopted ex ante regulation of digital market, the aim of the regulation was not to replace ex post enforcement, but ex ante regulatory frameworks can help provide clarity and predictability for market players, particularly in fast-moving digital environments where market power may become entrenched before traditional enforcement tools are able to intervene effectively. While this paper does not enter into the debate of whether AI should be subject to regulation, the experience of regulation of digital platforms has shown that ex ante measures can complement enforcement by addressing gaps where reactive tools have proven too slow or limited to prevent harm.

Measures that could prove useful to support competition in AI-enabled market would include licensing terms for accessing data, auditability of models and their architecture, and interoperability between systems. Such measures could potentially become critical to ensuring market contestability. On the contrary, restrictive licences, unilateral withdrawal of access, or opaque training practices can foreclose downstream rivals. The Autorité de la concurrence has warned that “restrictive licensing terms or opaque access conditions” for foundational AI inputs such as data and compute may lead to exclusionary effects, particularly when imposed by vertically integrated firms. Similarly, the CMA, in its cloud infrastructure market investigation, identified interoperability barriers and discriminatory licensing as key risks to competition, especially where access to compute is bundled with affiliated AI services. Potential instruments to remedy against such practices would include imposing transparency obligations in the AI stack or in-house models, standard-setting initiatives to facilitate interoperability, and finally enforcement against discriminatory licensing (Autorité de la concurrence, 2024[70]; CMA, 2024[67]; OECD, 2021[128]).

3.3.4. Sector and international co-operation

International and cross-sectoral co-operation is essential to address competition risks in AI-enabled downstream markets. The global scale of compute infrastructure and foundation model development means that national enforcement alone is unlikely to be sufficient to address concerns related to cross-border market power or multi-jurisdictional exclusionary conduct. Moreover, the complexity and opacity of AI systems – particularly those embedded in regulated sectors – can make it difficult for any single authority to detect and address anti-competitive outcomes in isolation.

In many sectors – such as health, finance, energy, and transport – sectoral regulators oversee access to critical infrastructures and datasets, including electronic health records, financial transaction data, smart grid information, and mobility platforms. These regulators define the technical and legal conditions under which such data can be accessed, shared, or reused. While these frameworks are typically designed to protect privacy, safety, or system integrity, they can also shape the competitive dynamics of AI markets. For example, if access to a key dataset is limited to a small number of incumbents, or if licensing terms are overly restrictive, new entrants may be unable to develop or deploy competitive AI solutions (OECD, 2019[129]). In such cases, the absence of co-ordination between competition and sectoral authorities may inadvertently reinforce market concentration. OECD analysis has highlighted the importance of aligning access regimes with competition principles, ensuring that safeguards do not become de facto barriers to entry (OECD, 2025[68]).

Sectoral regulators also influence market structure through their role in setting technical standards, certification requirements, and procurement criteria. These instruments can affect the interoperability, auditability, and deployability of AI systems. For instance, explainability or safety requirements may determine which models are eligible for use in a given sector, potentially favouring certain architectures or providers. Co-ordination with competition authorities can help ensure that such requirements are implemented in ways that support innovation and do not unintentionally favour vertically integrated firms. In this context, co-operation is not only about avoiding regulatory conflict but about jointly identifying and mitigating structural risks that may not be visible from a single institutional perspective.

International co-operation is equally important to align approaches to AI regulation, enforcement, and market monitoring. Annual international competition events and conferences, and NCAs’ bi-lateral or multilateral engagements with international partners have help sharing learning and experience, and set up joint initiatives to address common challenges. Co-operation can support the development of common standards, facilitate cross-border enforcement, and ensure that AI markets remain open and contestable. This includes collaboration on issues such as technical standards, data governance, and competition assessments of global AI providers.

Notes

← 1.  CMA Decision on the case No. 50223 12/08/2016.

← 2.  As established in Case T-41/96 Bayer, ECLI:EU:T:2000:242

← 3.  The Commission’s 2022 commitment decision in Amazon Buy Box (AT.40703), linked to Amazon Marketplace (AT.40462), addressed alleged self-preferencing in Amazon’s automated Buy Box and Prime eligibility systems, which the Commission found tended to favour Amazon Retail and Fulfilment by Amazon sellers. While no infringement was established, Amazon accepted binding commitments to ensure non-discriminatory Buy Box display (including a “Second Offer”), and a ban on using non-public seller data, with accountability extending explicitly to its automated systems.

← 4.  Other bottlenecks in the AI stack, such as cloud computing or access to infrastructure, are discussed in-depth in (OECD, 2025[4]) and (OECD, 2025[3]).

← 5.  A proprietary data ecosystem like LinkedIn’s professional graph of 1.2 billion members and 60 million firms, illustrates how control over strategic data assets can sustain incumbents’ positions and remain a source of competitive advantage (LinkedIn/Ipsos, 2025[55]).