This article is part of the special issue New Measures of Competition in the Digital Economy, prepared in collaboration with the International Center for Law & Economics (ICLE).
Abstract
Market definition and market share analysis are central elements of most antitrust assessments. In markets for AI models, however, challenges arise. Because of the variety of ways in which models are deployed, it is difficult to obtain data that cover the whole market and are commensurable. This article investigates how far market share analysis carries here. It specifically looks into turnover, token volume, model downloads, and web traffic. The core thesis is that, where comprehensive market share quantification fails, traditionally distinct analytical steps, namely market delineation, share quantification, and market power assessment, collapse into one unitary analysis of competition. This shows that, owing to insufficient data for market share measurements, the approach shifts from form-based to effects-based, and that further empirical research, especially using survey techniques, is required.1
I. Introduction
The present article deals with the analysis of market shares in the supply of AI models and inference services from an antitrust perspective as well as with the antecedent issue of market definition. The core thesis is that a comprehensive market share quantification can fail owing to insufficient and incommensurable data. As a consequence, traditionally distinct analytical steps, namely market delineation, share quantification, and market power assessment, may converge towards one unitary analysis of competition. An EU Competition Policy Brief of 2024 has suggested that, for market share quantification, alternative metrics besides turnover might become relevant.2 The present article assesses turnover and two of the Policy Brief’s suggestions, namely token volumes and model downloads. It adds web traffic as a fourth. It evaluates the results and deals with data gaps. Combined data analysis can be challenging if the available data do not contain information about the data to be estimated. Based on these findings, the article draws conclusions on what this means for the analytical structure of market power assessments in the field of AI models. Substitutability between models can be assessed on a combination of different parameters and yield insights into competitive forces. Industry surveys might become more important to close data gaps and monitor competitive developments.
The article starts with a primer on how models and model-based inference are deployed (II.). It then homes in on market definition (III.), followed by gauging market volume and market shares (IV.), before it concludes (V.).
II. AI Models and Inference: Generation and Distribution – a Primer
From a functional perspective, the notion of a model can be used to represent the stack of components that allows a user prompt (input) to be turned into a response (output) by way of inference. The heart of the model is the model weights, which have been created during training. Naturally, the model was trained on data. Model weights, accordingly, are “the set of learned parameters that overlay the model architecture to produce an output from a given input.”3 The model architecture, on the other hand, is the code which specifies the network structure. This is a piece of code which is not produced by the model but by the author of the model before training starts. When the user makes a prompt vis-à-vis a trained model, another piece of software becomes involved. This is the inference code, which now allows a reaction to be produced according to the model weights. This leads to the following definition as provided by the Open Source Initiative: “An AI model consists of the model architecture, model parameters (including weights) and inference code for running the model.”4
From the perspective of the demand side, one can now distinguish between the demand for the model as a technological product on the one hand, and the demand for the inference service based on a model on the other. Both types of demand can ultimately serve the same purpose from the customer’s perspective, but technically they can be seen as different types of goods.
Take, as an example, a law book publisher offering a legal AI chatbot to clients. Here, the publisher owns relevant special data (the law library) which it wants to use for offering Retrieval Augmented Generation (RAG) legal reasoning to its own subscribers (lawyers and universities).5 The missing element, from the publisher’s perspective, is the inference performance.6 The publisher can now purchase inference as a service from a firm which runs a model in its own data center. This can be a firm which has developed and trained the model, such as Anthropic or OpenAI, where the customer, viz. the legal publisher, accesses the inference service via an API.7 Those model developers usually do not disclose their model weights (hence the name “closed-weight model”). All that is offered is the inference service (in this case to the commercial customer via API, not via a consumer-facing platform). Inference service can also come from a provider who relies on a model it has obtained from a third party that has developed and trained it (third-party model inference service provider).8 Alternatively, the publisher can obtain a model and run it autonomously. Repositories such as Hugging Face offer a range of models that can be downloaded (open-weight models9). In this instance, the model is acquired, not the inference service. Here, one can further distinguish between cases in which the commercial user (viz. publisher) has its own data center where the model is deployed, and cases where the commercial user rents compute in a third-party data center while operating the model there autonomously, i.e. purchasing compute, not inference.
III. Market Definition
1. Model Download vs. Inference Service
When it comes to market definition, the initial challenge is to decide whether buying inference services on the one hand and downloading a model to perform inference by oneself on the other are part of the same relevant market from the customer10 perspective or not. Literature in AI science and economics often treats the rendering of inference services on the one hand and the possibility of downloading models on the other hand as functional substitutes.11
From the angle of antitrust market definition, however, such an equivalence is not self-evident. Inference requires compute. Computation is performed on servers. Running such servers requires expertise and energy. A copy of a model, therefore, is incapable of producing any inference. Rather, the model operator (e.g. the legal publisher who has downloaded the model) needs to build its own server, or to rent compute on a third-party server to run inference on the downloaded model. Technically speaking, the user now self-supplies the inference service. The outcome, ultimately, is the same as when buying inference services, e.g. from a closed-weight model inference service provider. Yet the way to reach that aim is different. It is buying a service versus buying the components necessary to perform the service in-house.
What the article encounters here is a well-known phenomenon in market delineation. The question is whether self-supply of a service can count as market volume for the merchant market. This could make sense where such self-supply constitutes a viable option for the demand side. Is driving in one’s own car a substitute for taking the train, and does the volume of kilometers covered by people in their own cars on a certain route count as volume on an overarching transportation market together with travelling by train? In some instances, the Commission has shown openness towards considering the self-execution of a service as a substitute within the framework of market definition.12
In the context of AI models and services, the following might matter. Market definition is not an end in itself. It is supposed to define those competitive forces that work directly on the firms which are involved in the relevant conduct.13 If the relevant options are viable alternatives from the customer perspective, this speaks in favor of substitutability. At least for larger firms with a need for inference, downloading a model and running it in-house might constitute a viable substitute for using external inference services. Model quality will oftentimes not be an obstacle. The quality of open-weight models is catching up with that of closed-weight models, only lagging behind by some months.14
Yet even if one came to the conclusion that model downloads for self-supplying the inference versus buying inference services were two different markets15, it would still be necessary to recognize the possibility for a user to self-host a model. Even if it is less convenient and more complicated than buying the inference service, depending on potential price increases for such service, self-supply can be a realistic option which limits the inference service providers’ market power, albeit, and merely in arguendo, not to an intensity that would justify its inclusion in the relevant market. The market-power mitigating effect of self-supply, even if it stops short of being considered as market volume in a given case, is well recognized in antitrust practice.16
2. More Granular Analysis of Substitutability
Within closed-weight and open-weight models, further distinctions might be necessary in terms of market definition. While large foundation models can perform a range of tasks including coding, some specialist models have more limited capabilities. Depending on the purpose, the substitutability can therefore vary between different models and be asymmetric. Specialized models can be more token efficient and, thus, cheaper for certain tasks, thereby producing competitive pressure on some large language models for certain user segments, as reported recently on TypeSafe AI’s Jev. Jev17 is marketed as a cheaper and more specialized alternative for software developers to some large language models.18
When the demand for a specialized model means that other specialized models are no substitutes, they can still be compressed into the same relevant market if the concept of supply-side substitutability applies. This requires that developers of different kinds of specialized models can offer new or adapted models tailored to the specific demand. Even if the launch of new substitutes is not possible with the required speed for them to constitute part of the existing market, the more distant possibility of such a new model release can qualify as potential competition. Potential competition, albeit not sufficiently direct for it to feature within the concept of the relevant market, is a recognized factor that can offset supply-side market power to a certain degree.19
IV. Market Volume and Market Shares
1. Data Sources
a) Turnover
Quantifying the supplier’s share in the relevant market calls for a unitary metric. For closed-weight inference services and for third-party model inference services, turnover can be a viable gauge. An EU Commission Policy Brief of September 2024, however, shows some reluctance towards relying on revenue, citing the “nascent state” of the industry and a lack of full monetization of “many AI foundation models”.20 Yet it is questionable whether turnover should really be discarded as a valuable measurement. The Policy Brief dates from 2024, with the AI industry developing quickly. Even where monetization is not full, an equal rate of monetization can be enough to use turnover as a proxy for market share. Beyond that, Anthropic and OpenAI are currently preparing public offerings.21 Once listed, their revenues will be auditable and publicly reported, which will make those data more reliable. Against this backdrop, the proposition that turnover is an unsuitable gauge for the analysis of market shares for inference services would require further scrutiny in 2026.
Irrespective of the above, however, revenue does not necessarily reflect the exact use volume of a model. Closed-weight models can be relatively more expensive per token than inference done on some open-weight models.22 That can lead to a bias when measuring use volume based on turnover estimates. Most importantly, however, when it comes to measuring inference done on self-hosted open-weight models, turnover as a metric is usually not available. This leads to a data gap if one wants to compare the use volume of self-hosted models with inference services purchased from providers (closed-weight or third-party open-weight). The article will return to that problem later when it deals with data gaps.
b) Tokens
aa) The Indicative Value of Tokens
The Commission’s Policy Brief considers alternatives to turnover, such as tokens or model downloads.23 The present article will test tokens and model downloads – as two prominent examples mentioned in the Policy Brief – and will add web traffic.24
A token is the smallest unit into which a model segments input and output. Often, tokens are only fragments of words. Generally, the more tokens processed, the heavier the workload. Token volume within a certain time has been suggested to measure a model’s usage in practice.25 Potential sources for token volume data are aggregator platforms. They route requests from model users to inference service providers. Such a platform is OpenRouter. It is able to measure model use based on the requests passing through its infrastructure.26 OpenRouter data have been used to estimate the relative usage of open-weight and closed-weight models as well as model substitutability.27
bb) Limited Market Coverage
However, the data such platforms provide are only partial. Platforms such as OpenRouter capture two channels, namely user traffic with closed-weight models and with third-party model inference providers. This allows for a comparison between closed-weight traffic and the traffic regarding some open-weight models as operated by these third-party inference service providers. What such an aggregator platform cannot capture is the token volume relating to models that are downloaded and run autonomously by the users on their own or rented servers.28 Also, it does not capture traffic routed through competing platforms.29 OpenRouter only accounts for a small fraction of global LLM inference.30 This shows how the platform’s share in global token volume is very limited. Moreover, OpenRouter states that “[r]equests that a user or app keeps private are excluded before aggregation”.31 Also, some users connect with the AI provider directly, for example for data security reasons, without directing traffic through any aggregator platform.32
cc) Differences in Token Efficiency
Another difficulty arises, which is independent of the coverage problem. How much text a token stands for can vary between models or model iterations. For example, Anthropic’s Claude Opus 4.7 produced between 32% and 45% more tokens than the previous version for the same amount of text, depending on prompt length.33 Where a model’s tokenizer fragments text more finely, the model will therefore appear to be used more extensively even where the actual usage is the same. Token count for the same task can also vary across languages.34 OpenRouter points out these limitations of token data.35
dd) User Characteristics on Aggregator Platforms
A better understanding of the relationship between closed-weight and open-weight models, as represented on an aggregator platform such as OpenRouter, may require examining the characteristics of the platform’s users in order to extrapolate figures to the market as a whole. Some authors raise the possibility that users active on OpenRouter are “more inclined toward open models than the general population”.36 If this is the case, the ratio between closed-weight and open-weight utilization in the broader market would be shifted more towards closed-weight than the platform data show. The authors make that comment in support of their central finding, namely that open-weight models are “underutilized relative to what their price-performance characteristics would predict”.37 Their estimates, therefore, operate as lower bounds for an assessment of the broader market, since the self-selection of user types on OpenRouter works against their own results. While the present article takes no stance on the authors’ results, the analysis shows how token data derived from aggregator platforms call for a cautious interpretation when drawing conclusions about the market as a whole.
c) Model Downloads
The number of model downloads, as measured for example on Hugging Face, has been used as a proxy for market relevance in empirical studies.38 Some authors caution that counts should be adjusted to avoid distortions caused by automated retrieval. They propose narrowing counts to downloads of models that have been “created within a year of the download”.39 These authors argue that such a “Rolling Window Filter” metric “isolates more authentic popularity and usage trends, before the download signal is overcome by automatic processes that don’t actually use the models.”40 They hold that “[t]ypically, more performant or efficient versions of models are available well within one year”.41
Yet the download number does not directly say anything about the use volume of this model.42 Download is a one-time event whereas usage is a flow, so that download numbers are not commensurable with usage as measured by token volume. The Rolling Window Filter does not change that. The authors acknowledge that Hugging Face download metrics are “at best proxies for real usage”.43 These metrics “best serve the analysis of broader trends and relative popularity over time, rather than precise numerical designations of economic concentration.”44
d) Website Traffic
Other metrics might be considered, but they come with limitations of their own. Website traffic, as measured on consumer-facing sites of model operators, such as OpenAI for ChatGPT or Anthropic for Claude.ai, may provide complementary evidence of the relative reach of such services. However, website traffic does not directly represent underlying model usage.45 Visiting a website does not provide information about user prompts and the concomitant inference work. Also, there is a coverage gap. The approach leaves out access via mobile apps and desktop apps46, which can be a substantial portion. Moreover, this approach only measures traffic where model operators have a consumer-facing web presence. For many open-weight models that is not the case. Self-hosted models are not represented in the data either.
Alternatively, one might look into web traffic referred by AI model services onward to other websites.47 This measure will not reflect model usage, but the relevance of a model for outbound traffic. It might still indicate the economic footprint of inference services and can therefore be part of a broader body of data to be looked at when measuring market relevance. This approach, however, is limited too since model use without any outbound traffic is not covered, which is possibly a significant chunk.
2. Dealing with Incomplete and Incommensurable Data
a) The Challenges of Data Combination
How to deal with data gaps? The data sources – turnover, token volume on aggregator platforms, downloads from repositories, traffic on consumer-facing websites of model operators, and traffic routed to third-party websites – each have missing observations for some user segments. It could therefore make sense to exploit the information contained in the different measures jointly. It might be feasible to specify a joint distribution which describes the measures and their relationships and to estimate its parameters from the available data.48 Problems, however, arise. The first is that the approach builds on the assumption that the relationships estimated from observed cases also apply to missing entries. The assumption is difficult to support here since the data gaps arise for reasons which are not reflected in the observed cases. Some users self-host open-weight models for reasons of confidentiality or high inference load, which both bear on the token volume to be estimated. Token volumes observed in other use cases, such as those for closed-weight model service providers, do not reflect these circumstances. The second problem is that the different data are not expressed in a common unit of measure. Turnover is expressed in units of money, while token volume is not. What is missing in the raw data is a linking structure that would make these data commensurable.
b) Surveys
To overcome the challenges resulting from the fragmentation and heterogeneity of the data, user surveys could grow into significance.49 Surveys are a well-established tool in empirical economics to measure business and consumer adoption as well as time use. Some of the problems coming with tokens as a unit of measure may persist even if a survey is used.50 Nevertheless, surveys could address the data fragmentation problem in a comprehensive way that would also enable the gathering of qualitative evidence on critical aspects of the market. Such a survey could ask about hosting arrangements, usage intensity, and even model identity, where possible. The advantage of a survey would be that it could see the whole market, thereby covering data gaps that arise in the other data sets discussed above. In particular, it would capture self-hosted model deployments, which largely do not show up in aggregator data. Cross-country comparisons could help identify differences in adoption, provider choice, and competitive conditions. Further data could be harvested on perceived switching costs and customer lock-in.
c) Switching Costs
Finally, the competitive analysis on relevant markets for models will need to put additional weight on assessing qualitative evidence for determining the degree of outside options available for users. One such factor is switching costs. Customers can face higher switching costs, for example, when moving from an inference service provider to a self-hosted open-weight model. Former customers will then need to build the technological expertise, develop or contract the computing capacity and energy to replicate the service. Switching between inference services, and switching between self-hosted open-weight models, might imply lower barriers therefore resulting in a closer competitive relationship within each supply group, i.e. model services versus self-hosted open-weight models.51
d) Event Studies
Additional information on substitutability might come from event studies, for instance on new model releases and how users react. Fradkin, for example, has analyzed substitutability from this perspective based on token volume data from OpenRouter, looking into whether model releases have attracted new users or displaced existing demand from competing models.52 Insights into the relationship between closed-weight and open-weight models might be obtained if open-weight download numbers reacted to price changes by certain closed-weight model service providers. Even though download numbers do not say anything about the use volume of a downloaded model, they indicate the direction in which customers move and whether they have viable outside options. They do not substitute for market share data but can yield information on the competitive forces acting on closed-weight model service providers.
e) Direct Assessment of Competitive Forces
This is not the place to elaborate further on these additional paths that could be followed to get a more comprehensive understanding of markets for models. The overarching conclusion is that market share as an isolated type of data will hardly do justice to the competitive landscape. Market share reflects the current footprint which a model has gained at a certain moment. While such information, where obtainable, is a valuable source for market power assessment, other criteria gain in importance. Where data gaps hinder a full description of the market in quantitative terms, the analysis must focus on the competitive interaction between models, on how deployment concepts constrain each other, and on how accessible markets are, including the role of intermediary platforms. The underlying data may not cover the entire user base, so that absolute market positions will be difficult to state. Yet the data can provide information on relative movements, such as the impact which a model release has on token volume on an aggregator platform. Such fragmented data are still useful for finding out about model substitutability and the degree of competitive constraint which models exert on each other. Thus, the distinctions between market delineation, market share quantification, and market power assessment become brittle.
Consequently, the assessment of markets for models shows how an effects-based approach, comprising quantitative and qualitative evidence, can become necessary where data fragmentation hinders a quantification of the overall market. It demonstrates the limits of form-based approaches.53 This is not an extension of the well-known argument that effects-based approaches can yield more granular information. It responds to the insight that in some markets, the data necessary for resorting to form-based criteria, like market shares, are insufficiently robust or even absent.
V. Conclusions
This article has shown how far market share quantification carries, the challenges into which it can run, and what follows from that. Two core problems arise: data do not cover the entire market, and available data lack commensurability. Against this backdrop, a combination of different metrics in conjunction with qualitative evidence on outside options can be used to get a better understanding of market power. Where market share quantification fails, the distinction between traditionally separate analytic steps, namely market delineation, share quantification, and market power assessment, gradually dissolves. This demonstrates why an effects-based approach can become necessary owing to insufficient data for form-based measurements. The focus shifts to relative movements, such as how model releases impact on token volumes among certain models and user groups. This helps to understand substitutability and the competitive constraints between certain models. The article also highlighted the need for further empirical research, especially using survey techniques, as discussed.
In the network
Articles linked to this one by shared themes.
- 1Measuring Dynamic Competition in Digital Markets: Artificial Intelligence, Generative AI and the Evolution of Market Power
Reiko Aoki, Fall 2026 - 2The Future-Proof Fantasy of AI Regulation
Thibault Schrepel, Fall 2025 - 3Principal-Agent Dynamics and Digital (Platform) Economics in the Age of Agentic AI
Volker Stocker and William Lehr, Fall 2025 - 4Crafting the Regulatory Ecosystem for AI: Competition 2.0
Nuno Cunha Rodrigues, Summer 2025