�ԹϺ�����

Executive summary

The rapidly evolving market for artificial-intelligence services is apparently thriving and very competitive, with a growing number of AI start-ups and ever larger and more capable AI models. Investors are pouring large amounts of money into start-ups and into a few big tech firms that have the large financial resources and computing capacity that constitute key inputs to the production and running of AI-driven services. However, exponentially growing AI model training costs create a market entry barrier for AI start-ups, forcing them to cooperate with big tech firms to get access to computing infrastructure and end users. At the same time, their AI models compete with those of the big tech firms. Competition authorities investigate these co-opetition agreements, looking for market-distorting clauses. Beyond computing costs, other competition choke points in the AI supply chain include access to dedicated AI processor chips and training data. Strict enforcement of copyright on data may further tighten an already scarce supply of affordable training data. 

Policymakers have few, if any, effective tools to deal with these potential competition bottlenecks in AI industries. Co-opetition agreements are necessary to enable AI start-ups to access hyperscale and costly computing infrastructure. At the same time, the agreements give  big tech privileged access to the latest AI models. Competition authorities can try to minimise market-distorting provisions in these agreements but there are technical and economic limits to the extent of fragmentation in the AI value chain that can be meaningfully imposed on big tech firms. Licensing fees for copyright-protected data further tighten the already scarce supply of affordable data. It is hard to implement a fair market-clearing mechanism for dedicated AI processors.

If model training costs continue to grow exponentially, as appears to be the case for the foreseeable future, the entire competition policy setting for AI industries may need a revision, with collaboration, including between big tech firms, dominating competition to keep the pace of AI innovation going.

The artificial intelligence industry is booming, powering stock market valuations ¹ Lewis Krauskopf, ‘Echoes of dotcom bubble haunt AI-driven US stock market’, Reuters, 2 July 2024, . . From 2022 to 2023, investment in generative AI in the United States by big tech companies and private venture capital in AI start-ups increased from less than €1 billion to over €20 billion (Madiega and Ilnicki, 2024). In the EU, it increased from almost zero to nearly €4 billion over the same period. 

However, a series of high-profile collaboration agreements between AI start-ups and big-tech firms, combined with emerging bottlenecks in AI input markets, have drawn the attention of competition authorities. There is suspicion that big tech companies are jostling to strengthen their market positions in the AI industry and that they use these agreements to stifle competition and increase the dependency of, and restrict the room for manoeuvre for, start-ups. Investigations into these agreements have been started in France, Portugal, Hungary, the United Kingdom and the US ² Matt O’Brian, ‘FTC opens investigation into Big Tech’s partnerships with leading AI startups’, Los Angeles Times, 25 January 2024, . . The US Department of Justice in June 2024 announced “urgent” scrutiny of big tech’s control of AI ³ Stephen Morris, Javier Espinoza and Stefania Palma, ‘US antitrust enforcer says ‘urgent’ scrutiny needed over Big Tech’s control of AI’, Financial Times, 6 June 2024, . , in particular choke points in the supply chain, including access to computing power, data and dedicated AI processors. Moreover, acquisitions of entire teams of AI engineers from other firms, such as Microsoft’s hiring of Inflection AI staff in early 2024 ⁴ Paul Kunert, ‘Microsoft hiring Inflection team triggers interest from EU’s antitrust chief’, The Register, 5 April 2024, . , are perceived as a way to circumvent merger regulation.

Such competition checks are considered urgent because competition authorities do not want to be caught short a second time by big-tech firms, as happened over the past decade when a few online platforms grew very fast and managed to carve out dominant market positions. Classic slow-grinding competition policy procedures were unable to catch up with them. A major justification for the EU Digital Markets Act (DMA, Regulation (EU) 2022/1925) – the main EU competition policy tool for very large digital platforms – was precisely to create a fast ex-ante instrument so that policymakers would no longer have to wait for a competition problem to occur before they could intervene. 

The French Autorité de la Concurrence (2024) issued an opinion in June 2024 on competition bottlenecks in the AI value chain. The UK Competition and Markets Authority (2024) noted that the growing presence of a few big-tech firms that underpin AI by providing computing resources, expertise and monetisation channels, might shape AI-related markets to the detriment of fair, open and effective competition. The European Commission as the European Union’s competition authority is reviewing collaboration agreements and called at the start of 2024 for contributions on competition in generative AI ⁵ See European Commission press release of 9 January 2024, ‘Commission launches calls for contributions on competition in virtual worlds and generative AI’, . . AI systems are already integrated in services operated by the hard-to-avoid ‘gatekeeper’ platforms that are monitored under the DMA.

This Policy Brief examines competition-reducing market-entry barriers in each segment in the AI value chain, from upstream markets for model training inputs, intermediate markets that match AI model developers with deployers of AI-driven services and downstream markets that match deployers and end users. To structure the debate, it proposes a collaboration and competition or “co-opetition” rationale (Brandenburger and Nalebuff, 1996) for collaboration agreements between AI start-ups and big-tech firms. It examines possible policy responses to these bottlenecks and shows how, in some parts of the value chain, there are no clear-cut solutions because competition authorities are caught between conflicting policy objectives. In particular, exponentially growing AI model training costs constitute an increasingly high market entry barrier that cannot be addressed by imposing constraints on the role of big tech firms or increasing public sector AI investments.

2.1 Competition between AI models

At first sight, the overall market for AI models, which respond to prompts by producing original outputs – for example images or text – based on analysis of vast datasets ⁶ Examples include OpenAI’s ChatGPT, Microsoft’s Co-Pilot and Google’s Gemini. , looks very competitive. The supply of AI models and the number of firms developing them is increasing fast. The AI Index Report (Maslej et al, 2024) lists more than five hundred “notable” AI models ⁷ Epoch, a widely respected AI research organisation, uses the term “notable machine learning models” to designate noteworthy models handpicked as being particularly influential within the AI/machine learning ecosystem. In contrast, foundation models are exceptionally large AI models trained on massive datasets, capable of performing a multitude of downstream tasks. Examples of foundation models include GPT-4, Claude 3 and Gemini. While many foundation models may qualify as notable models, not all notable models are foundation models. See Epoch AI, ‘Notable AI Models’, 19 June 2024, . , including hundreds of very large foundation models (FMs), or models able to perform a range of tasks that are trained on very broad sets of data and can be added to in order to create new applications (Figure 1).

Figure 1: Foundation models by access type

 ^{Source: Maslej et al (2024).} 

Large language models, such as OpenAI’s ChatGPT ⁸ For more explanation about ChatGPT, see for example Amanda Hechler, ‘What is ChatGPT’, TechTarget, June 2024, . , are a subset of FMs. They can handle text, images and audio, for example providing images and audio output based on descriptive text input. Hundreds of thousands of specialised AI model applications run like apps in an app store on top of FMs. By the end of April 2024, the OpenAI GPT applications store contained 159,000 specialised GPT applications created on top of the baseline ChatGPT AI model ⁹ Daniel Baek, ‘GPT Store Statistics & Facts: Contains 159.000 of the 3 million created GPTs’, SEO.AI, 24 April 2024, . , which provides the equivalent of an ‘operating system’ for these AI apps, similar to the way Google Android is the operating system for apps running on an Android smartphone.

However, a first important qualification to this overall view of the market for AI models is that models should be distinguished according to their availability to users. About two thirds of all FMs are fully open source, available publicly for free commercial and non-commercial use. Open models reduce AI market entry barriers and increase competition between users, giving them a wide range of models to choose from. Depending on the degree of openness, model users can create their own versions of the model (Solaiman, 2023). Once they are released, developers lose control of their open models, which may have consequences for safety and responsible use. Anyone can modify the models for harmful purposes, such as producing fake news, disturbing videos or racist speech.

However, the performance of open models is somewhat below that of closed models (Maslej et al, 2024, p 146). The best FM models are not released for open use ¹⁰ Meta’s Llama model is an exception, according to Maslej et al (2024). . That reduces the competitiveness of firms that deploy freely available open models, compared to firms that have access to closed but better performing models – if the underlying AI model is less good, the apps built on top of it will also be less good. This finding contradicts the often-heard claim – copied from the open software movement – that open models accelerate AI-driven innovation because many parties can use them.

2.2 Fixed computing costs as a barrier to market entry

A crude measure of the capability of AI models is the number of tokens they can handle. Tokens are created by slicing training data into fragments – for example parts of words or sentences – enabling analysis at a very granular level. Since the launch of the first Google Transformer model (Vashwani et al, 2017), AI models have grown exponentially, from millions to trillions of tokens in the latest models (Figure 2). The processing of more tokens require more computing capacity and training data, and so computing capacity requirements and the associated costs have increased accordingly ¹¹ Computing requirements have increased from a few thousand to a hundred billion petaflops – a unit of measure for the calculating the speed of a computer equal to one quadrillion floating-point operations per second. Floating point operations per second is a measure of computer performance.  (Maslej et al, 2024, p 49-51).

The red line in Figure 2 shows that models are still on a linear growth path on the logarithmic scale of computing needs and costs. There are no signs yet of diminishing returns to scale, not even across this wide range of orders of magnitude of increase in inputs. Since models are pre-trained, prior to use, training is a fixed cost, independent of the intensity of use of the model. High fixed training costs are the main AI market entry barrier. This favours large firms with the financial capacity to cover these costs. 

It explains why AI start-ups seek collaboration agreements with big-tech firms. 

Figure 2: Estimate training costs and computing requirements for selected AI models 

^{Source: Maslej et al (2024). Note: * see footnote 11.}

Model training requires expensive dedicated AI processors. Nvidia is currently the market leader but competition is picking up, from established chip producers like AMD and Intel, and from big-tech firms building their own AI processors. Other input market segments in the chips supply chain are dominated by a small number of firms. For example Netherlands-based ASML is a market leader in lithography machines for chip production.

To train and run the largest FMs, hundreds of thousands of dedicated AI processors need to be assembled in cloud server farms. Only the largest big-tech players – essentially Google, Amazon, Microsoft and Meta, and Apple and Nvidia, collectively known as the GAMMANs – have the required cloud infrastructure and computing capacity to cater to the training needs of the largest AI models. Meta is reportedly developing a state-of-the art AI computing system, which with 350,000 Nvidia H100 processors ¹² Katie Paul , Stephen Nellis and Max A. Cherney, ‘Exclusive: Meta to deploy in-house custom chips this year to power AI drive – memo’, Reuters, 1 February 2024, . , each costing around $30,000, is estimated to cost more than $10 billion. Extrapolating the red line in Figure 2 by a few years, it is easy to see that the trillion-dollar AI processing farm would soon be reached. This market entry barrier is completely out of reach of public funding; only the very largest firms can aspire to this, and even they may have to collaborate in future.

Several authors suggest that an oligopolistic market for AI FMs is inevitable because the tight control that GAMMANs have over essential assets confines smaller players to subordinate roles. Azoulay et al (2024) argued that sharing resources is the only way to avoid market concentration, unless a “rogue” GAMMAN were to use openness as a competition strategy. Korinek and Vipra (2024) also expected the market for FMs to be quite oligopolistic because of significant economies of scale and scope in deployment.

High training cost market entry barriers explain why AI start-ups seek to negotiate collaboration agreements with the GAMMANs, trading access to computing infrastructure for GAMMAN access to their latest models (Figure 3). Alternatively, AI start-ups can drop away from the red technology frontier line and side-track to smaller AI models that can be fined-tuned and perform well on specific tasks, or they can move into the booming market for ‘grounding’ of large models to create applications with proprietary data (grounding refers to connecting AI model output to verifiable sources of information and specific proprietary data sources ¹³ Grounding thus links model responses to user queries more closely to trusted data and reduces the chances of model hallucination, or invention of content. See for example Google Cloud documentation on grounding, available at . ).

Figure 3: Collaboration agreements between big tech and smaller AI start-ups

^{Source: CMA (2024).} 

2.3 Access to model training data

FMs and LLMs are pre-trained on very large text datasets taken from books, documents, Wikipedia and webpages. FMs are essentially statistical models that predict sequences of tokens. They need many occurrences of sequences of tokens in order to make robust predictions. That, in turn, requires large volumes of data. Likewise, FMs that produce images and audio require large volume inputs of audio-visual material. 

AI developers have already reached the limits of available high-quality human-edited text data to train their models (Maslej et al, 2024, p 52). The supply of lower-quality text from social media, or voice-to-text conversion is less constrained and could be sufficient until the early 2030s on current model size trends. However, low-quality input reduces the quality of model outputs. Using synthetic training data may cause model ‘collapse’, meaning a dramatic drop in model output quality. This will become an important issue as more digital media content is generated synthetically by AI models. It would trigger a negative synthetic data feedback loop in AI model training. Picture and audio data are sufficiently ubiquitous to keep audio-visual model training going for another two decades.

Many if not most of the text, audio and picture AI training datasets are subject to copyright. Authors can in principle charge license fees for use of their works. That would trigger an additional cost-induced shrinkage in the supply of training data (Gans, 2024). It would also increase the cost of training of models and reduce competition between model developers. The EU AI Act (finalised in May 2024) requires model developers to respect EU copyright law as set out in the Copyright Directive (Directive (EU) 2019/790), in particular Article 4, which grants an exception to copyright for commercial research but allows the copyright holder to opt-out of this exception. The AI Act requires AI model developers to be transparent about the data they use, including with regard to this opt-out. Many copyright holders and their collecting societies have become aware of uses of their creative content for AI training purposes and are introducing explicit opt-outs ¹⁴ See for example Brad Spitz ‘AI data mining: French music collecting society Sacem opts out’, Kluwer Copyright Blog, 25 January 2024, . .

In the US, the fair-use and transformative-use exceptions to copyright may apply to AI training data. But this is still subject to legal uncertainty, with several court cases pending. AI investors run the risk of punitive statutory damages if courts decide against the fair-use exception. To avoid this, the largest AI firms have signed data-licensing deals with large media companies. For example, OpenAI signed with the New York Times, the Bertelsmann media group and the Reddit news platform. These deals give access to high-quality human-edited text for model training and to recent information that can be ‘grounded’ into models without having to go through costly re-training. But even the largest AI firms are unlikely to sign licensing deals with all copyright holders. Strict enforcement of copyright law will increase the price of access to training data. That is likely to reduce the volume of copyright-protected training data that model developers will access, thus weakening model performance. Smaller AI developers and start-ups may not have the financial resources to pay for copyright licenses and may be pushed out of the market altogether. Smaller models with less data and tokens can still be trained, but will not be at the frontier of AI model performance.   

Global differences in copyright regimes may distort the geographic level playing field for AI developers. They may choose to move model training to countries with more favourable copyright regimes. The EU AI Act requires that all models deployed in the EU respect EU copyright law. This ‘Brussels effect’ may force AI firms based outside the EU to comply with EU copyright law, even though copyright law is essentially territorial. Alternatively, a reversal of the Brussels effect might keep the best models out of the EU market, at the expense of EU consumer welfare and businesses productivity ¹⁵ For more details on the AI and copyright debate, see Martens (2024). .

2.4 Access to intermediate model deployers and end users

AI model developers need business channels to generate revenue to cover the costs of training and running their models. Some start-ups try to build their own business models from scratch. For example, OpenAI reached more than 100 million users within a year of the launch of ChatGPT. It has created a GPT app store which downstream model deployers have populated with hundreds of thousands of specialised ChatGPT applications. Application developers have access to the open ChatGPT model but pay an app-store entry fee. OpenAI also charges subscription fees for users of the professional version of ChatGPT. The app store generates network effects that make ChatGPT even more attractive: more users attract more application developers, and vice versa. Responsibility for compliance with AI Act standards shifts from model developers to deployers.

However, starting a business model from scratch is hard for less-successful AI start-ups with weaker or no network effects. An easier route to revenue is to collaborate with the GAMMANs and embed AI models into their well-established business models. For example, Microsoft is embedding its own and third-party AI models into all its productivity software and the Bing search engine. It charges premium prices for access to some of its AI-driven services. Google is doing the same for its search engine and other services. Meta has substantially increased its advertising revenue with the help of AI systems ¹⁶ Jo Arazi, ‘Meta’s Q1 Ad Revenue Soars 27% Amid AI-Driven Growth’, West Island Community Blog, 12 July 2024, available at . .

This results in “co-opetition” agreements (Brandenburger and Nalebuff, 1996) between GAMMANs and AI start-ups. Start-ups collaborate with GAMMANs to embed their AI models into existing GAMMAN user-facing services at the downstream end of the value chain, in return for reverse collaboration at the upstream end of the value chain, where GAMMANs grant start-ups access to computing infrastructure and possibly training data. In the middle of the value chain however, collaboration is replaced by competition: start-up AI models compete with the GAMMAN’s own AI models. And the parties must bargain over how to divide the value chain they create together. GAMMANs may be vertically integrated along the entire AI value chain while start-ups cover mostly the input and intermediate parts of the value chain. 

Figure 4: The collaboration-competition model between AI start-ups and GAMMANs

^{Source: Bruegel.}

This section examines competition policy concerns that may arise in this competitive symbiosis between GAMMANs and start-ups, and possible tools to deal with these concerns.

3.1 Bottlenecks in upstream AI model inputs markets 

Hardware inputs

GAMMANs have their own hyperscale hardware facilities. Hardware is a rival good; it can only be used by one party at the time. It is difficult to find a market-clearing mechanism that ensures open and fair access for all contenders to available AI processor chips and cloud computing capacity. First-come-first-served is not a viable strategy because the first could buy the entire production capacity and re-sell it later. Price auctions are also hard to conceive. Quota would be difficult to allocate on a fair basis: how much to whom? The essential facilities doctrine (Graef, 2019) under Article 102 TFEU is not applicable because there is no single unique source of hardware inputs ¹⁷ An essential facility is an asset or infrastructure to which a third party needs access in order to offer its own product or service on a market. A facility is essential if no reasonable alternatives are available and duplication of the facility is not feasible because of legal, economic or technical obstacles. .

This makes it difficult to issue a practical policy recommendation. An optimist might assume that hardware bottlenecks are just temporary transition problems as new AI processors from competing manufacturers, and additional cloud computing capacity from new market entrants, enter the market. There is no guarantee however that this bottleneck will disappear soon. The UK CMA is prudent and advocates “vigilance” in this segment of the AI value chain. The French competition authority (Autorité de la Concurrence, 2024) suggests public sector investment in EU supercomputing capacity – in other words, government subsidies. Given the exponential growth in AI computing requirements (Figure 2), it is unlikely that the public sector has the financial capacity to offer a viable alternative. Also, providing public computing capacity does not solve the fair-allocation problem. On the contrary, political preferences might affect allocation decisions.

Model training data

With exponentially growing AI model size, and no diminishing returns in sight yet, maximum access to data is a necessary condition to maximise model performance. The good news is that, unlike hardware inputs, training data is non-rival and can be used by many at the same time. The bad news is that many sources of high-quality training data are subject to copyright and licensing fees. Strict copyright enforcement will increase the cost and shrink the supply of data. Model quality will decline, especially for smaller start-ups that cannot afford copyright licensing fees, and in smaller language zones where the supply of data is intrinsically limited. This reduces competition and innovation (Azoulay et al, 2024; Korinek and Vipra, 2024). It may also distort the geographical level playing field if copyright regimes differ between countries. 

Competition authorities are aware of these potential negative effects. France’s Autorité de la Concurrence (2024) recognises that provisions in the EU AI Act might disproportionately affect start-ups. It recommends collective licensing to reduce copyright licensing transaction costs, and price differentiation according to the value of the data. It is not clear how that would work in practice. In our view, competition and innovation would be best served with the elimination of the opt-out clause for AI training data from Article 4 in the EU Copyright Directive. If US courts would confirm the application of fair use and transformative use for AI model training data, that would have a similar impact. It would not affect traditional revenue channels for copyright holders, while they would benefit from productivity increases through better AI models.

3.2 Bottlenecks in downstream markets for model deployers and users

Co-opetition agreements between start-ups and GAMMANs point to potential competition bottlenecks in downstream AI markets on two levels: the market for models between AI start-up model developers and GAMMAN deployers, and the AI services market between GAMMAN deployers and end users. If start-ups directly deploy their models to end users, or via their own application stores, there is no competitive bottleneck because the market for start-ups is very competitive. Except for Nvidia, all GAMMANs have been designated as gatekeepers in their core platform services (CPS) for business users and end users under the EU DMA ¹⁸ See . . Most of these CPS already run on embedded AI models. The DMA is in principle technology-agnostic: relevant DMA obligations for these CPS apply, irrespective of the technology used to deliver the service.

Some obligations in the DMA’s Article 6 are relevant for co-opetition relationships between independent AI model developers that deploy their services through established GAMMAN business models and intermediation platforms. For example, Microsoft Windows, Apple iOS and Google Android were designated as CPS operating systems. They already include third-party AI model-driven services applications. Under the DMA obligations, these CPS should allow the installation of third-party software and the un-installation of gatekeeper software (Articles 6(3) and 6(4)), enable access to and interoperability of the same hardware and software features (Article 6(7)), permit effective portability of end-user and business-user data (Articles 6(9) and 6(10)) and not give preferential treatment to their own and third-party services (Article 6(5)).

Would this imply that these three CPS should allow the (un)installation of other AI models than those they selected, ensure that these models are inoperable with their AI-driven services, allow data portability between their CPS and third-party model providers and give no preferential treatment to their own AI models, compared to any of these third-party models? For example, Google, Microsoft and Apple already bundle and tie several AI models with their AI-driven services. Should end users have a say in the choice of models? Uncertainty about the application of DMA obligations may explain why Apple decided to withhold some AI applications from the EU market ¹⁹ Stephen Warwick, ‘EU says Apple withholding Apple Intelligence from the EU market is anti-competitive in frankly hilarious turn of events’, iMore, 28 June 2024, . . Google Search has been designated as a CPS. DMA Article 6(11) mandates sharing of search engine ranking, query, click and view data. Would that apply to queries and outputs of AI-driven services embedded in the search engine?

Some start-ups grow so fast that they may become DMA gatekeepers in their own right. For example, OpenAI ChatGPT is coming close to fulfilling the quantitative DMA threshold criteria to become a gatekeeper, and the ChatGPT app store is close to being a core platform service: over 100 million users, estimated capital value over $80 billion and hundreds of thousands of business users that develop specialised ChatGPT applications. If designated, would DMA app store obligations apply and would they be relevant to an AI model app store, as compared to a smartphone app store for which the obligations were designed? ChatGPT could be considered as the ‘operating system’ on which specialised AI model applications run. As long as access to the app store remains open, not subject to discrimination or preferential treatment of OpenAI’s own apps, that should satisfy the DMA objective of contestable markets.

The Autorité de la Concurrence (2024) recommended that competition policymakers should pay particular attention to this developer-deployer market segment for AI models-as-a-service (MaaS). The UK CMA (2024) recommended that co-opetition agreements should not contain vertical restraints, exclusive deals or tying or bundling, either in upstream or downstream markets in the AI value chain. For these recommendations to work in practice, the technical layers of this intermediate developer-deployer market side, as well as the deployer-end user market side, should be unbundled. This is technologically complex and economically difficult. DMA authorities are currently trying to cope with the challenges of the designated CPS and are nowhere near to contemplating further unbundling of these CPS. Competition authorities can explore vertical restraints and exclusive deals but have no practical means to deal with them yet. That leaves a lot of regulatory uncertainty and investor hesitations hanging over the EU market for AI model services. 

Exponentially growing model training costs, with no end in sight, are the biggest AI model market entry barrier, at least at the technology frontier. For firms with smaller models below the technology frontier, competition is likely to be intense, suggesting a more limited role for competition policy. AI start-ups that want to stay at the technology frontier need to sign co-opetition agreements with GAMMANs to overcome the training cost barrier. Competition authorities are looking at these deals with suspicion. They fear that co-opetition agreements may become Trojan horses for GAMMANs to exercise leverage over, and reduce competition from, AI start-ups. But several investigations by competition authorities have so far not produced any smoking guns.

These investigations have to some extent diverted attention away from the main market entry barrier – model training costs – and directed it to potential competition problems in different market segments in the AI supply chain, while suggesting ways to keep these segments open and contestable. The UK CMA (2024) offers guidance on rules of behaviour that it expects the GAMMANs to follow in their business dealings with AI start-ups. This includes open access to hardware and data, making a variety of business models available to AI businesses and consumers, partnerships that do not reduce the ability of others to compete, no self-preferencing of AI services and models, and no tying or bundling. The Autorité de la Concurrence (2024) proposed similar principles in its opinion.

These principles are also in line with the obligations that the EU DMA imposes on a set of designated gatekeeper platforms that largely overlaps with the GAMMANs. However, they may be hard to implement in practice as they may require a considerable degree of technically complex and economically costly unbundling and untying of AI-driven services and the underlying AI models that power these services. This may do more harm than good to consumers and business users. Co-opetition agreements are necessary to enable AI start-ups to access computing infrastructure. They come with privileged access for big tech to the latest AI models. Licensing fees for copyright-protected data further tighten the already scarce supply of affordable data.

Model training infrastructure and AI processor chips are rival physical inputs that can only be used by one party at the time. It is hard to see how a fair quota or price-based market-clearing access mechanism could be put in place. Training datasets are non-rival inputs that can be used by many parties at the same time. Public data pooling could be a fair access mechanism (Azoulay et al, 2024; Korinek and Vipra, 2024) but would run counter to the private interests of copyright holders.

If model training costs continue to grow exponentially, as appears to be the case for the foreseeable future, the competition policy setting for AI industries may need a revision, with collaboration dominating competition to keep the pace of AI innovation going. It may eventually require collaboration between the GAMMANs. Computing infrastructure and data may become an essential facility, with agreed rules to govern shared access.

It may also be that the current generation of FMs and LLMs is superseded in the next years by new types with different cost structures and characteristics. Rapid advances, for example in adversarial networks in which several AI models compete and generate feedback between themselves, or in expanding memory capacity and query context data, so that AI models can learn from user feedback and instructions, may still change the AI landscape.

The EU does not have home-based GAMMANs with sufficient computing capacity, financial resources and business models for FM training and use. Private equity and venture capital might contribute financial resources but physical infrastructure and business outlets remain a bottleneck. Intervention through public finance and infrastructure is not an option given the order of magnitude of financing required. Keeping access to AI inputs and outputs markets open, as far as technically possible and economically meaningful, may be the best option for EU policymakers if they want European AI start-ups to remain competitive in an industry dominated by very large US-based GAMMANs.

This also applies to the strict EU copyright regime, which will inflate model training costs and shrink the supply of data, at the expense of EU AI start-ups. The EU AI Act imposes potentially high compliance costs that may push up the market entry barrier for model developers and deployers, particularly for start-ups. The EU AI Office, established under the AI Act, still needs to draft many implementation modalities for the AI Act. It should try to do so in a way that keeps markets open and does not impose disproportionate compliance costs on smaller AI developers. But the AI Act does not give the AI Office sufficient leeway to address the underlying competition bottlenecks in the AI supply chain.

Autorité de la Concurrence (2024) ‘Avis sur le fonctionnement concurrentiel du secteur de l’intelligence artificielle générative’, Avis 24-A-05, available at 

Azoulay, P., J. Krieger and A. Nagaraj (2024) ‘Old Moats for New Models: Openness, Control, and Competition in Generative AI’, Working Paper 32474, National Bureau of Economic Research, available at 

Brandenburger, A. and B. Nalebuff (1996) Co-opetition: A Revolution Mindset that Combines Competition and Cooperation, Crown Publishers, UK

CMA (2023) Proposed principles to guide competitive AI markets and protect consumers, UK Competition and Markets Authority, available at 

CMA (2024) ‘CMA AI strategic update’, UK Competition and Markets Authority, 29 April, available at 

Gans, J. (2024) ‘Market power in Artificial Intelligence’, Working Paper 32270, National Bureau of Economic Research, available at 

Graef, I. (2019) ‘Rethinking the Essential Facilities Doctrine for the EU Digital Economy’, TILEC Discussion Paper DP2019-028, available at 

Korinek, A. and J. Vipra (2024) ‘Concentrating Intelligence: Scaling Laws and Market Structure in Generative AI’, mimeo, available at 

Madiega T. and R. Ilnicki (2024) ‘AI investment: EU and global indicators’, At a Glance Digital Issues in Focus, European Parliamentary Research Service, available at 

Martens, B. (2024) ‘Economic arguments in favour of reducing copyright protections on Generative AI model inputs and outputs’, Working Paper 09/2024, Bruegel, available at /working-paper/economic-arguments-favour-reducing-copyright-protection-generative-ai-inputs-and

Maslej, N., L. Fattorini, R. Perrault, V. Parli, A. Reuel, E. Brynjolfsson … J. Clark (2024) The AI Index 2024 Annual Report, AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, available at 

Schaeffer, R., B. Miranda and S. Koyejo (2023) ‘Are Emergent Abilities of Large Language Models a Mirage?’ 37th Conference on Neural Information Processing Systems (NeurIPS 2023), available at 

Solaiman, I. (2023) ‘The Gradient of Generative AI Release: Methods and Considerations’, mimeo, available at 

Vaswani, A., N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, L. Kaiser and I. Polosukhin (2017) ‘Attention Is All You Need’, mimeo, available at 

Zhao, X., S. Ouyang, Z. Yu, M. Wu and L. Li (2022) ‘Pretrained language models can be fully zero-shot learners’, mimeo, available at