Competition Notions in the World of AI

In this work, we want to test the hypothesis that the current market structure, as well as some technical aspects of AI (most notably how compute-intensive it is), have more to do with the concentration of compute resources - and the characteristics of the companies who hold them - than with inherent requirements of the technology. However, to better frame this question, we need to better understand some notions from competition analysis in the context of AI development.

Vertical Integration Dimensions for AI

A producer of goods or a supplier of services may either use them for its own purposes or supply them to the market. A firm that wishes to sell its products must therefore decide how to organize production, sales, and distribution. They may choose to carry out all of these activities themselves, a model commonly referred to as vertical integration, or it may rely on the products and services of other firms through vertical agreements. Both vertical integration and especially vertical agreements are generally less likely to harm competition, and can show efficiencies and lower prices to consumers, as each level of the value chain is often assumed to be subject to effective competition.

However, such arrangements can still be anticompetitive in certain circumstances. For example, in vertical agreements, there can be anti-competitive practices like foreclosure of rival suppliers or buyers by applying exclusivity deals, or practices such as 'resale price maintenance', where a supplier requires retailers not to sell below a fixed price. Vertically integrated groups can, for example, engage in self-preferencing, whereby a company prioritizes its own products or services over those of a third party.

Self-preferencing is a concern in the AI environment, where vertically integrated companies are often best placed to capture value.

In AI, most model companies in the US and EU are either cloud or compute providers themselves (e.g., Google, NVIDIA, Microsoft), have significant funding in the form of access to significant cloud credits or preferential terms (e.g., Mistral, OpenAI, Anthropic), or are partially owned by cloud or chip companies (TODO, more systematic + cite sources)

In app development, for example, vertically integrated firms compete directly with startups through general-purpose applications such as chatbots. This means that companies like Google or Microsoft could, for example, bundle their applications (Google's Gemini or Microsoft's Copilot) into their already widely used services like Google Workspace or Microsoft 365, potentially engaging in self-preferencing that gives their own products an unfair competitive advantage. Other concerns are, for example, the possibility for vertically integrated companies to have an unfair advantage by combining user data from multiple products that they offer to better train and sell their AI models; or to strategically limit access to their models to competitors in downstream markets. Importantly, anti-competitive effects may arise not from a single practice, but from the accumulation of several practices that may appear legitimate when considered individually. And especially in the AI value chain, vertically integrated companies compete with startups, but they also play a crucial role by providing infrastructure and acting as key access points to customers.

Economies of Scale in AI

In some markets, the larger the production of a particular product, the lower its price per unit produced, which is referred to as economies of scale. In particular, markets and activities that benefit from strong economies of scale tend towards concentration. However, the concentration of resources or market share is not necessarily seen as a sign of anti-competitive behavior. Abusive behavior would arise when a monopolist, or a firm with sufficient market power to act as one, is able to shape market outcomes in line with its own interests. Such conduct may include practices like pricing below cost in order to drive competitors out of the market or deter entry, with the aim of subsequently raising prices once competition has been weakened. This strategy is commonly referred to as predatory pricing. This is possible only for companies with an existing significant market share, since a firm with no market power or sufficient resources cannot engage in this type of behaviour.

In the case of compute for AI, one way in which economies of scale show up is by facilitating better utilization of compute resources. The running costs of AI compute are primarily proportional to the numbers of GPUs purchased, and to a lesser extent to their "on-time". Maximizing utilization is the most straightforward way to reduce this cost per user. This can be achieved through either managing a sufficiently diverse user base to see consistent requests through time, or by developing use cases where model requests can be staggered through time. Models that run on a single GPU vs models that require multiple nodes may be affected differently. Another aspect to economies of scale is the ability to reduce prices through bulk purchases or reservation of compute time, which facilitates the efficient management of when resources are used.

The greater a firm's market power in a given market, the more effectively it can engage in anticompetitive conduct. This is particularly relevant for AI supplier startups that depend on hyperscalers for critical infrastructure such as compute and cloud services. These startups may face multiple forms of dependency that expose them to potential lock-in practices. For example, hyperscalers can offer cloud credits for free or at heavily discounted rates much more easily than mid-tier cloud providers, creating dependency. They may also impose data egress fees or limit interoperability, making it difficult for AI suppliers to move data or workloads between cloud providers. This would restrict their access to essential resources and limit their ability to switch providers. While such behaviors have not been widely observed in the market to date and can be directly illegal in certain jurisdictions, they remain important considerations in assessing the AI value chain.

Common Ownership Structures

Vertically integrated firms such as Amazon, Google, Microsoft, and, to some extent, Nvidia are expanding across multiple stages of the AI value chain through a combination of internal development, partnerships, and strategic investments. As a result, so-called "coopetition" (strategic alliances between actual or potential competitors formed to achieve mutual benefits, often for specific projects) has become a dominant business model in the industry.

And a related phenomenon is common ownership, which arises when competing firms are partially owned by the same entity. In the AI sector, there have been multiple partnerships and investments involving the same major firms, increasing the overall level of interdependence within the ecosystem and potentially affecting competitive dynamics.

From a merger control perspective, there is a clear trend of large technology companies entering into partnerships and making investments in smaller AI developers. These arrangements are often pro-competitive, as they provide access to essential inputs, such as compute, data, and capital, and enable the development of advanced AI systems.

Partnerships and investments can, in some cases, raise competition concerns similar to those associated with mergers, particularly where they lead to anticompetitive effects such as reduced rivalry or foreclosure, contributing to entrenched market positions from the bigger firms. For this reason, some partnerships have been scrutinized by competition authorities, notably the Microsoft - OpenAI partnership, which has been reviewed at least by the European Commission and UK competition authority. Microsoft's investment in OpenAI over several years, has created a mutually beneficial relationship: OpenAI gains access to Microsoft's extensive computing infrastructure, while Microsoft gains access to OpenAI's technologies. Nevertheless, partnerships of this nature must not serve as a means for one party to acquire de facto control over the other without appropriate regulatory scrutiny. Where such control arises, the arrangement may fall within the scope of merger review.

Commonly Discussed Externalities

The question of externalities is particularly important when analyzing AI markets, as many of the costs of the technology are borne neither by the provider nor by the consumer. These include harm to the information ecosystem (AI slop everywhere, automating abuse, etc.), and particularly in the case of compute harms to the environment and pressure on the energy and water infrastructure, including in areas that are already distressed.

Competition analysis has traditionally focused on price, output, and innovation effects, and has generally not taken environmental externalities into account. Environmental harms have instead been addressed through sector-specific regulations. However, non-price dimensions are increasingly recognised as relevant parameters of competition in digital markets, and this separation may begin to erode (such as the case for privacy concerns, see below).

Scholars have argued that standard competition tools, such as pricing and cost analyses, could be adapted to reflect the fact that market prices do not always internalise environmental costs. In this view, environmental externalities borne by society, rather than by firms, may distort competitive outcomes. For example, a dominant firm may be able to undercut rivals through predatory pricing, loyalty rebates, or margin squeeze practices because it relies on polluting inputs or greenhouse-gas-intensive technologies, thereby externalising environmental harm and gaining an artificial competitive advantage. Comparable concerns arise where firms dispose of waste or pollutants in violation of environmental standards, reducing costs relative to competitors that comply with the law. In such cases, the anticompetitive conduct and the environmental infringement may be closely intertwined (more on (1)).

In AI markets, dominant and vertically integrated firms often have the ability to secure exclusive or preferential access to critical inputs such as energy and water, for example through long-term nuclear or other energy supply agreements for data centers. At the same time, many of the environmental externalities associated with these activities are borne by society rather than internalised by the firms themselves. Limited transparency regarding energy use, water consumption, and emissions can further contribute to the externalisation of environmental harm, allowing such firms to gain a competitive advantage over rivals. While these practices are not necessarily unlawful, particularly where regulatory frameworks have not yet fully adapted to the realities of AI infrastructure, they nonetheless raise significant societal concerns and could be relevant considerations for competition analysis.

This line of reasoning parallels the evolving treatment of privacy in competition law. In at least one case so far (Meta v. Bundeskartellamt, referred to the CJEU), it was found that breaches of data protection rules may be relevant in assessing abuses of dominance, insofar as they affect competitive conditions and consumer choice (more on this: (2)). Regulatory standards, while not replacing competition law, can thus inform competition analysis and, in some circumstances, serve as relevant benchmarks or proxies. A similar approach could be envisaged for environmental regulation, particularly where environmental externalities are so high stake.