TILEC

TILEC Discussion Paper

DP 2024-02

Regulating Competition in Generative AI: A Matter of

Trajectory, Timing and Tools
By

Friso Bostoen
Anouk van der Veer

March 2024

ISSN 1572-4042
SSRN: https://ssrn.com/abstract=4756641

Electroniccopy
Electronic copy available
available at:
at: https://ssrn.com/abstract=4756641
https://ssrn.com/abstract=2956308
 Regulating Competition in Generative AI: A Matter of Trajectory, Timing and Tools

Friso Bostoen* and Anouk van der Veer†

In January 2024, the European Parliament called on the European Commission (EC) to consider
adding generative AI (GenAI) services to the list of ‘core platform services’ of the Digital Markets
Act (DMA). Others argue that the EU need not wait for new regulation: competition law’s abuse
of dominance provision can already be used to tackle concentration in the industry. To help shape
its views, the EC launched a call for contributions on competition (or the lack thereof) in GenAI
markets.
This is serious regulatory zeal for a technology that only became widely used at the end of 2022.
But the present must be understood in light of the past. The previous platform shift, in the 2000s,
saw the rise of dominant players in key web 2.0 markets like search and app distribution. The
prevailing idea is that regulatory intervention came too late to keep those markets effectively
competitive. Hence, it is better to intervene early now.
But is the analogy justified? Is GenAI a platform technology that requires immediate regulatory
scrutiny to stay competitive? To find out, we address three sub-questions. The first concerns the
trajectory of GenAI: is the market destined to evolve into a few ecosystems ruled by dominant
platforms? (Section I) The second concerns the timing of regulation: given the new and quickly
evolving nature of GenAI, what is the ‘right’ time to act? (Section II) The third concerns the
regulatory tools: can we put existing competition law to work, or do the market dynamics call for
a tailored instrument? (Section III)

I. The Trajectory of Generative AI

ChatGPT is synonymous with GenAI in the same way that Google has become a synonym for web
search. Having acquired 100M users in just two months, it is considered the fastest-growing
consumer product of all time (for comparison, it took Facebook four and a half years to acquire as
many users).3 But, in an evolution that is characteristic of digital markets, GenAI did not stop at
being a product—instead, it evolved into a platform. The models have been growing larger and

*
Assistant Professor of Competition Law & Digital Regulation, Tilburg University. Correspondence:
[email protected]. Note that the article builds on the authors’ previous work, in particular Anouk van
der Veer and Friso Bostoen, ‘Two Views on Regulating Competition in Generative AI’ (Network Law Review, 20
February 2024) <https://www.networklawreview.org/veer-bostoen-generative-ai/>. The article is updated until 1
March 2024.
†
Ph.D. Researcher, European University Institute. Correspondence: [email protected].
3
Jon Porter, ‘ChatGPT continues to be one of the fastest-growing services ever’ (The Verge, 6 November 2023)
<https://www.theverge.com/2023/11/6/23948386/>.

Electronic copy available at: https://ssrn.com/abstract=4756641

larger, but that trend may be starting to reverse. In this section, we look at both developments: the
platformization of GenAI, and the trends in model size.

a. From Products to Platforms

It is helpful to think about the GenAI industry in layers. ChatGPT is just the tip of the iceberg.
Such consumer-facing applications rely on GenAI models, in particular Large Language Models
(LLMs)/Foundation Models (FMs). For example, ChatGPT—at least the paid version—is built on
OpenAI’s latest model, GPT–4. These models need to be trained, which requires two key inputs:
data and compute.4 OpenAI’s GPT models were trained using compute provided by Microsoft’s
Azure cloud infrastructure, which is, in turn, powered by Nvidia’s microchips (in particular, its
graphics processing units or GPUs).5 The training data was collected from the open internet,
licensed from third parties, and provided by users.6
One can thus distinguish three GenAI layers: infrastructure, in particular compute; models; and
applications, which is how models are ‘deployed’ to end-users. The model layer, in particular,
shows platform characteristics, as independent developers can build their own products on top of
this technological foundation. To facilitate this, OpenAI puts a range of application programming
interfaces (APIs) at the disposal of developers, which can ‘call’ on the APIs in their own products.7
A GPT Store has been set up for developers to sell their apps to users.8 Meta goes a step further:
it open-sourced its LLM, Llama, so that developers can customize the model itself.9
The GenAI market has structured itself not just around platforms but into ecosystems. ‘Ecosystem’
is a strategic management concept that refers to ‘a group of interacting firms that depend on each
other’s activities’, usually for the purpose of innovation.10 The terminology puts the focus on the
complex relations between players, as well as the environment in which interactions take place.
Ecosystems tend to be orchestrated by a core platform, such as those found in the model layer. But
what is notable in GenAI ecosystems is the outsized role played by infrastructure providers.
The outsized role of providers of infrastructure, in particular cloud computing, is explained in part
by its concentrated nature. Microsoft Azure and Amazon Web Services (AWS) are the market

4
A third key input is talented personnel.
5
John Roach, ‘How Microsoft’s bet on Azure Unlocked an AI Revolution’ (Microsoft News, 13 March 2023)
<https://news.microsoft.com/source/features/ai/how-microsofts-bet-on-azure-unlocked-an-ai-revolution/>. More
generally, see ‘Computational Power and AI’ (AI Now Institute, 27 September 2023)
<https://ainowinstitute.org/publication/policy/compute-and-ai>.
6
Michael Schade, ‘How ChatGPT and Our Language Models Are Developed’ (OpenAI)
<https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-language-models-are-developed>.
7
See < https://platform.openai.com/>.
8
See <https://chat.openai.com/gpts>.
9
See <https://llama.meta.com/>. Nuancing the open-source nature of Llama, see Charlie Hull, ‘Is Llama 2 open
source? No – and perhaps we need a new definition of open…’ (OpenSource Connections, 19 July 2023)
<https://opensourceconnections.com/blog/2023/07/19/is-llama-2-open-source-no-and-perhaps-we-need-a-new-
definition-of-open/>.
10
Michael Jacobides, Carmelo Cennamo and Annabelle Gawer, ‘Towards a Theory of Ecosystems’ (2018) 39
Strategic Management Journal 2255.

Electronic copy available at: https://ssrn.com/abstract=4756641

leaders, with a combined 60–80% share of the market, with Google Cloud Platform (GCP) a distant
third.11 The chip industry is potentially even more concentrated: Nvidia is said to capture 90% of
the AI GPU market.12 Investors report that their portfolio companies spend up to 80% of their
capital on compute resources, which underlines the importance of cloud infrastructure. 13 This is
aggravated by the fact that compute is currently constrained, mainly due to the underlying need
for chips.14 For these reasons, cloud providers play a structuring role in GenAI ecosystems.

Figure 1. Non-exhaustive overview of GenAI ecosystems15

GenAI was hailed as a ‘platform shift’.16 The last platform shift occurred in the 2000s, which saw
the rise of mobile (smartphones) along with the maturing of the web (search, social networking, e-

11
The exact market shares vary per country (and in reality, there is unlikely to be a single ‘cloud’ market in antitrust
terms), see Authority for Consumers & Markets, ‘Cloud Services’ (Market Study, 2022) ACM/21/050317, 33–35;
Autorité de la concurrence, Opinion 23-A-08 on Competition in the Cloud Sector, 29 June 2023, paras 296–304;
Ofcom, Cloud Services Market Study (Final Report, 2023), para 11.30 and Annex 1.
12
‘NVIDIA has 90% of the AI GPU Market Share’ (techovedas, 21 January 2024) <https://techovedas.com/nvidia-
has-90-of-the-ai-gpu-market-share-1-5-to-2-million-ai-gpus-to-be-sold-by-nvidia-in-2024/>. Note that Google has
developed its own Tensor Processing Units (TPUs), which are optimized for AI training and can compete with
GPUs, see <https://cloud.google.com/tpu>. Another significant breakthrough is Groq’s Language Processing Unit
(LPU), see <https://wow.groq.com/>.
13
Guido Appenzeller, Matt Bornstein and Martin Casado, ‘Navigating the High Cost of AI Compute’ (Andreessen
Horowitz, 27 April 2023) <https://a16z.com/navigating-the-high-cost-of-ai-compute/>.
14
The main cloud players therefore invest strategically to safeguard access, as when Microsoft invested in the cloud
infrastructure startup CoreWeave, which has preferential access to chips from NVIDIA (another of its investors), see
Mark Haranas, ‘Microsoft’s CoreWeave Deal “Adds AI Pressure” To AWS, Google’ (CRN, 2 June 2023)
<https://www.crn.com/news/cloud/microsoft-s-coreweave-deal-adds-ai-pressure-to-aws-google>.
15
The three layers (horizontal) group firms active at the levels of infrastructure, models and applications. The firms
developing models—or platforms to collaborate on models, as with Hugging Face—are placed in the ecosystem
(vertical) of their preferred cloud provider. The overview is based on public information (e.g., press releases), which
may be incomplete.
16
Ben Thompson, ‘Windows and the AI Platform Shift’ (Stratechery, 24 May 2023)
<https://stratechery.com/2023/windows-and-the-ai-platform-shift/>.

Electronic copy available at: https://ssrn.com/abstract=4756641

commerce). This meant the rise of new firms like Google, Facebook (Meta) and Amazon, while
older firms like Microsoft declined in relevance.17 What is notable about the shift to GenAI is that
the firms that rose to incumbency during previous shifts are also leading this one. This is due to
their access to key inputs. As discussed above, Microsoft, Amazon and Google own the main cloud
infrastructures. In addition, they have access to proprietary sets of training data; Google even has
superior access to the open web.18
It is no surprise, then, that entrants in the GenAI space partner with infrastructure providers.
OpenAI benefits from Microsoft’s compute infrastructure via a $13B investment, for example.19
Anthropic has received multi-billion investments from both Amazon and Google.20 Of course, the
infrastructure providers also have their own models (e.g., Google’s Gemini/DeepMind). For now,
there seems to be plenty of competition at the model level. Benchmarking is difficult, but every
other week, the ranking of top models appears to change. At the time of writing, the French startup
Mistral just announced that its latest model outperformed every other model except for GPT–4.21

b. Model Size: Two Roads Diverged into the Woods

Competition may seem fierce now, but where is the industry headed? So far, there has been a trend
towards ever-larger models, as performance increases with size.22 With about 1.8 trillion
parameters, GPT–4 is over ten times larger than GPT–3.23 Naturally, this trend raises concerns for
competition authorities.24 If there are large or even extreme returns to scale, as we saw with web

17
Apple, another company heralding the platform shift, was not exactly new but was certainly not (or no longer) an
incumbent from the previous platform shift to desktop PCs.
18
More websites allow its crawlers, which it has long used to build the web index for its search engine. Now,
crawlers are used to collect training data, but websites can block them. Only 24% of news sites block Google’s AI
crawler, compared to 48% blocking OpenAI’s crawler. See Richard Fletcher, ‘How many news websites block AI
crawlers?’ (Reuters Institute, 22 February 2024) <https://reutersinstitute.politics.ox.ac.uk/how-many-news-websites-
block-ai-crawlers>.
19
Kyle Wiggers, ‘Microsoft invests billions more dollars in OpenAI, extends partnership’ (TechCrunch, 23 January
2023) <https://techcrunch.com/2023/01/23/microsoft-invests-billions-more-dollars-in-openai-extends-partnership/>.
20
See <https://www.crunchbase.com/organization/anthropic/signals_and_news>.
21
Mistral AI Team, ‘Au large’ (press release, 26 February 2024) <https://mistral.ai/news/mistral-large/>.
22
Jared Kaplan et al., ‘Scaling Laws for Neural Language Models’ (2020) arXiv:2001.08361.
23
Maximilian Schreiner, ‘GPT–4 architecture, datasets, costs and more leaked’ (The Decoder, 11 July 2023)
<https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/>
24
The Portuguese competition authority focuses on the risks of the increasingly large models, see Autorida de da
Concorrência, ‘Competition and Generative Artificial Intelligence’ (Issues Paper, 2023). Other authorities are still
examining the market, see Autorité de la concurrence, ‘Generative artificial intelligence: the Autorité starts inquiries
ex officio and launches a public consultation’ (press release, 8 February 2024)
<https://www.autoritedelaconcurrence.fr/en/press-release/generative-artificial-intelligence-autorite-starts-inquiries-
ex-officio-and-launches>; FTC, ‘FTC Launches Inquiry into Generative AI Investments and Partnerships’ (press
release, 25 January 2024) <https://www.ftc.gov/news-events/news/press-releases/2024/01/ftc-launches-inquiry-
generative-ai-investments-partnerships>; EC, ‘Commission launches calls for contributions on competition in virtual
worlds and generative AI’ (press release, 9 January 2024) IP/24/85; Hungarian Competition Authority, ‘GVH
launches market analysis on the impact of artificial intelligence’ (press release, 4 January 2024)
<https://www.gvh.hu/en/press_room/press_releases/press-releases-2024/gvh-launches-market-analysis-on-the-
impact-of-artificial-intelligence>.

Electronic copy available at: https://ssrn.com/abstract=4756641

2.0 platforms, the market is bound for concentration. But authorities also noted that this is not an
inevitability.25
In fact, there are signs of a reversal in model size. Smaller GenAI models are developed that rely
on less trainable parameters but perform similarly or even better.26 For example, DeepMind’s
Chinchilla27 and Meta’s Llama28 have fewer parameters but outperform GPT–3.29 Llama was, in
turn, outperformed by the five-times-smaller Phi–2 model of Microsoft.30 Stanford researchers
managed to build Alpaca, a model on par with GPT–3.5, at a price of less than $600.31 The
aforementioned Mistral model also offers an example: its compact 5GB size allows it to run on
ordinary computers with about 16GB of RAM, while GPT–4 cannot run on consumer hardware.32
GenAI executives are seeing the writing on the wall. Microsoft’s Gates predicts that GenAI models
are reaching a performance plateau, so that GPT–5 will not be much better than GPT–4.33
OpenAI’s Altman believes models can still improve but believes these improvements will not
come from increasing their size.34 Such public statements should be taken with a grain of salt, as
these parties have an interest in downplaying the importance of size so as not to attract regulatory
scrutiny. However, an internal Google memo corroborates this story, setting out how the moat of
large models is not as large as it may seem.35
In short, while there has been a trend towards ever-larger models, it is far from certain that this
trend will continue. While it is early days, it is thus fair to say that the GenAI market does not
necessarily exhibit the same tendency towards concentration as web 2.0 markets. Recognizing it
is early in GenAI’s trajectory brings us to the next issue: timing.

25
Competition & Markets Authority, ‘AI Foundation Models’ (Initial Report, 2023) paras 3.73–3.78.
26
See, e.g., Yao Fu et al., ‘Specializing Smaller Language Models Towards Multi-Step Reasoning’ (2023)
arXiv:2301.12726 and Ian McKenzie et al., ‘Inverse Scaling: When Bigger Isn’t Better’ (2023) arXiv:2306.09479.
27
Jordan Hoffmann et al., ‘Training Compute-Optimal Large Language Models’ (2022) arXiv:2203.15556.
28
Hugo Touvron, ‘LLaMA: Open and Efficient Foundation Language Models’ (2023) arXiv:2302.13971.
29
Massive Multi-task Language Understanding (MMLU) is a benchmark for AI models. Whereas GPT–3’s
accuracy is 53.9%, LLaMA’s is 57.8% and Chinchilla’s is 67.5%, see <https://paperswithcode.com/sota/multi-task-
language-understanding-on-mmlu>.
30
Phi-2 outperforms LLaMA on different benchmarks (though not all of them), see Mojan Javaheripi and Sébasatien
Bubeck, ‘Phi-2: The surprising power of small language models’ (Microsoft Research Blog, 12 December 2023)
<https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/>.
31
Rohan Taori et al., ‘Alpaca: A Strong, Replicable Instruction-Following Model’ (2023)
<https://crfm.stanford.edu/2023/03/13/alpaca.html>.
32
See <https://docs.mistral.ai/models/>.
33
Matthias Bastian, ‘Bill Gates does not expect GPT–5 to be much better than GPT–4’ (21 October 2023, The
Decoder, 21 October 2023) <https://the-decoder.com/bill-gates-does-not-expect-gpt-5-to-be-much-better-than-gpt-
4/>
34
Will Knight, ‘OpenAI’s CEO Says the Age of Giant AI Models Is Already Over’ (Wired, 17 April 2023)
<https://www.wired.com/story/openai-ceo-sam-altman-the-age-of-giant-ai-models-is-already-over/>
35
The memo is available at <https://www.semianalysis.com/p/google-we-have-no-moat-and-neither>. The memo
includes statements like ‘Giant models are slowing us down’ and ‘Focusing on maintaining some of the largest
models on the planet actually puts us at a disadvantage’.

Electronic copy available at: https://ssrn.com/abstract=4756641

II. Regulatory Timing: Facing the Collingridge Dilemma
How does one regulate a technology—and, by extension, a market—that is in flux? That difficulty
is illustrated by the Collingridge Dilemma.36 Collingridge described how the consequences of a
technology cannot be predicted early in its lifecycle—hence, the need for regulation is unclear.
But by the time undesirable consequences manifest themselves, the technology has become part
of the economic and social fabric, so that controlling it is extremely difficult—you can no longer
put the genie back in the bottle. In other words, regulation is doomed to come either too early or
too late.
In making the decision between regulating early or late, the spectre of the past looms large. In their
calls for intervention, antitrust enforcers refer to the trajectory of the last platform shift. They
describe how it started with revolutionary technologies disrupting markets, but ended with
concentrated markets where firms solidified their dominance with anticompetitive practices.
Antitrust intervention came too late—if it came at all. Federal Trade Commission (FTC) Chair
Khan warns that ‘public officials have a responsibility to ensure this hard-learned history doesn’t
repeat itself.’37 EU Commissioner Vestager ties the timing of regulation even more closely to the
trajectory of the technology:
If the development of AI mirrors this history [of disruption followed by competition problems], then
the right response is to phase in competition control gradually, in line with market growth. To respond
faster than we did for Web 2.0, but to still give some time for the benefits of disruption to fully play
out.38
But she is doubtful history will repeat itself. GenAI development is not originating from tinkerers’
garages or university campuses as it did previously; instead, ‘disruption from AI will come from
within the nest of existing tech ecosystems.’39 As pointed out above, this is due to the nature of
GenAI models, which are trained using huge amounts of compute and data—a significant barrier
to the entry of upstarts (certainly those without Big Tech partner). In short, the trajectory of GenAI
may justify earlier intervention.
To dig a little deeper, let us consider the costs of early vs late intervention. As intervention in web
2.0 came (relatively) late, we have a better view of its downsides. The obvious one, as observed
by Collingridge, is that remedying undesirable consequences becomes exceedingly difficult. We
saw this clearly in the ineffectiveness of antitrust remedies.40 The problem is explained in part by
the particular nature of digital markets: due to economic characteristics such as network effects,
digital markets tend towards concentration. This is not problematic in itself, as increasing scale

36
David Collingridge, The Social Control of Technology (St. Martin’s Press 1980).
37
Lina Khan, ‘We Must Regulate A.I. Here’s How.’ (The New York Times, 3 May 2023)
<https://www.nytimes.com/2023/05/03/opinion/ai-lina-khan-ftc-technology.html>.
38
Margrethe Vestager, ‘Making Artificial Intelligence Available to All—How to Avoid Big Tech’s Monopoly on
AI’ (speech, 19 February 2024) SPEECH/24/931.
39
ibid.
40
See Friso Bostoen and David van Wamel, ‘Antitrust Remedies: From Caution to Creativity’ (2024) 14 Journal of
European Competition Law & Practice 540.

Electronic copy available at: https://ssrn.com/abstract=4756641

benefits users. However, when firms then solidify their dominant position through anticompetitive
conduct, it is difficult, if not impossible, to restore competition, as the remedy sagas of Microsoft
and Google have shown.41
Early intervention also has its downsides. New technologies like GenAI have the potential to
disrupt markets.42 Regulation can arrest the process of dynamic competition. This can harm users
who are denied the fruits of dynamic competition, whether they come in the form of a new product,
a reshuffled market structure, or further innovation driven by that competition.43 By contrast,
regulation may benefit first movers who would like to see their early lead cemented. From that
perspective, OpenAI’s calls for regulation44 can be seen as an effort to influence policymaking
with the goal of solidifying its first-past-the-post position.45 Asymmetrical regulation, which
applies to market leaders but not challengers, can partially address the risk of setting in stone a
particular market structure.
Importantly, GenAI models are not the only ones that learn from data—antitrust enforcers and
policymakers do too. And that is perhaps the strongest argument against regulating (too) early:
without deeper knowledge of its object, regulation is likely to miss the mark. But just like GenAI
models, authorities are learning faster and faster. While it took a decade or longer before web 2.0
platforms faced scrutiny,46 GenAI has been under investigation almost since day one. Sensibly,
scrutiny tends to take the form of market investigations and other information-gathering

41
On Microsoft, see Daniel Gore and Ashwin van Rooijen, ‘Ex Post Assessment of European Competition Policy:
The Microsoft Cases’ in Assimakis Komninos and Nicolas Petit (eds), Ex Post Evaluation of Competition Cases
(Kluwer 2021) 17–44; Google Shopping has not benefitted from an objective retrospective, but complaints continue
six years after the decision.
42
For the classical work, see Clayton Christensen, The Innovator’s Dilemma: When New Technologies Cause Great
Firms to Fail (Harvard Business Review Press 1997). However, note, first, that GenAI is not necessarily starting at
the low end of the market as foreseen by Christensen, due to the high cost of development and deployment
compared to the technologies it is challenging (e.g., search). Second, it can be debated whether GenAI is a disruptive
or rather a sustaining innovation.
43
Microsoft, which quickly integrated GenAI features into its Bing search engine, did not win significant share from
Google Search (less than 1 percentage point), see Jackie Davalos, ‘Microsoft’s Bing Market Share Barely Budged
With ChatGPT Add-On’ (Bloomberg, 18 January 2024). But the effects of innovation are more likely to show in the
decreased use of search engines altogether.
44
See, e.g., Cecilia Kang, ‘OpenAI’s Sam Altman Urges A.I. Regulation in Senate Hearing’ (The New York Times,
16 May 2023) <https://www.nytimes.com/2023/05/16/technology/openai-altman-artificial-intelligence-
regulation.html>.
45
For example, requiring government licenses to train large-scale models, as suggested by Altman, could slow down
progress to the benefit of the current market leader(s). More generally, any (symmetrical) regulation comes with
compliance costs that larger firms can bear more easily. On the compliance costs of financial regulation, see Drew
Dahl, Andrew Meyer and Michelle Neely, ‘Scale Matters: Community Banks and Compliance Costs’ (2016) The
Regional Economist July 1.
46
The EC’s Google investigation marked the start of the platform antitrust, see EC, ‘Commission probes allegations
of antitrust violations by Google’ (press release, 30 November 2010) IP/10/1624. Important reports followed some
years later, see notably Autorité de la concurrence and Competition & Markets Authority, ‘The Economics of Open
and Closed Systems’ (Report, 2014).

Electronic copy available at: https://ssrn.com/abstract=4756641

exercises.47 Having established a solid understanding of the industry’s dynamics, authorities can
then start taking a closer at potentially anticompetitive conduct.
At the same time, not every layer of the GenAI industry presents a mystery. Authorities started
studying the cloud computing market years before GenAI entered the scene.48 They did not stop
there. The French competition authority followed up its market investigation with unannounced
inspections at the offices of an unnamed graphics card manufacturer (read: Nvidia).49 To ward off
an EU probe, Microsoft changed its cloud licensing conditions, making it easier for customers to
switch providers.50 EU policymakers even felt comfortable enough to adopt the Data Act, which
also regulates cloud providers, including by facilitating customer switching/multi-homing.51
As the interventions in cloud markets show, a similar issue can be tackled by different instruments.
This throws up the question of regulatory tools: which type of intervention, if any, does the GenAI
market call for?

III. Regulatory Tools: Of Hammers and Nails

‘If the only tool you have is a hammer, everything looks like a nail.’ This quote is sometimes
(ab)used to criticize the EC for using antitrust enforcement to solve problems that are arguably not
competition problems. Meanwhile, the EC’s toolbox has expanded: it is now also the main enforcer
of the DMA. This new tool certainly shows overlap with the abuse of dominance provision, but its
enforcement mechanism is very different as it takes an ex ante rather than ex post approach.
Which tool is best for ‘the job’, i.e., safeguarding a competitive GenAI ecosystem? In answering
this question, timing matters. Competition law is a flexible tool; its general provisions can
immediately be applied to new technologies and market dynamics. By contrast, sectoral tools like
the DMA must define their scope and conduct rules in greater detail. This is hard to do before
competitive issues have clearly been identified, in particular via antitrust enforcement. Once in
place, however, sectoral tools can be applied within a shorter timeframe than competition law,
where investigations take years. Moreover, there may be issues of competition that escape
competition law, which can justify additional regulation.
Let us start by surveying the applicability of competition law.

47
See the consultations and investigations referenced above, footnotes 24–25.
48
See the reports referenced above, footnote 11.
49
Autorité de la concurrence, ‘Le rapporteur général de l'Autorité de la concurrence indique qu’une opération de
visite et saisie inopinée a été réalisée dans le secteur des cartes graphiques’ (press release, 27 September 2023)
<https://www.autoritedelaconcurrence.fr/fr/communiques-de-presse/le-rapporteur-general-de-lautorite-de-la-
concurrence-indique-quune-operation>.
50
Foo Yun Chee, ‘Microsoft offers to change cloud practices to ward off EU antitrust probe’ (Reuters, 28 March
2023) <https://www.reuters.com/technology/microsoft-offers-change-cloud-computing-practices-after-rivals-
complaint-source-2023-03-28/>.
51
Regulation (EU) 2023/2854 of the European Parliament and of the Council on harmonised rules on fair access to
and use of data (Data Act).

Electronic copy available at: https://ssrn.com/abstract=4756641

a. Competition law
For the GenAI industry, competition law’s abuse of dominance and merger control provisions are
most relevant.
Before finding any abuse, the authority must establish a dominant position. This appears difficult
in the GenAI model layer, which boasts a variety of players. Even if we consider only large-scale
models, there appears to be, at most, an oligopoly. The same goes for the infrastructure layer. In
cloud computing, there are three large players without any one of them being clearly dominant.
Based on the market shares quoted above, the firm most likely to qualify as dominant is Nvidia.
As of yet, the contours of abusive conduct are also unclear. Extrapolating from the current market
structure and dynamics, however, the following types of abuse can be envisaged.
• Input foreclosure. Compute is one of two key inputs of GenAI models. The main providers
of compute also have their own models. Therefore, they may have an incentive to foreclose
the developers of competing models from this input. But their ability to foreclose may be
limited, since there are at least two other major infrastructure providers. Nevertheless, model
developers may want to pre-empt such a move by partnering with an infrastructure provider,
which they have been doing widely.
• Exclusivity. The other key input for GenAI models is data. Given the extensive datasets that
GenAI models are trained on, however, no one firm holds a significant enough amount of
data to foreclose. Indeed, models are trained on the open web, or at least a chunk of it (e.g.,
the open CommonCrawl index). A competitive advantage may be gained by adding valuable,
proprietary datasets to the training data. Hence, a model developer with deep pockets may
try to lock up those datasets through exclusivity agreements.52
• Leveraging. The platformization of GenAI models opens the door to leveraging, in
particular from the platform to the application layer. A model provider could, for example,
favor its own apps over those of independent developers that rely on its model. This could
take a variety of forms, from demoting in ranking to degrading API access.
• ‘Open early, closed late’. Open-source models play an important role in the GenAI
landscape by allowing developers with fewer means to nevertheless build on powerful
models. Having reaped the rewards of openness, a provider could start closing down its
model. Such an ‘open early, closed late’ strategy need not concern models; it could also
relate to other inputs in the GenAI chain.53

52
Data from Reddit, for example, is said to be particularly useful for model training. Google inked a deal with the
company, though there is no mention of exclusivity, see Anna Tong, Echo Wang and Martin Coulter, ‘Reddit in AI
content licensing deal with Google’ (Reuters, 22 February 2024) <https://www.reuters.com/technology/reddit-ai-
content-licensing-deal-with-google-sources-say-2024-02-22/>.
53
For example, Microsoft has threatened to cut off access to its web index to firms that use it to compete with its
own AI chatbots, see Leah Nylen and Dina Bass, ‘Microsoft Threatens Data Restrictions in Rival AI Search’
(Bloomberg, 25 March 2023) <https://www.bloomberg.com/news/articles/2023-03-25/microsoft-threatens-to-
restrict-bing-data-from-rival-ai-search-tools>.

Electronic copy available at: https://ssrn.com/abstract=4756641

Merger control is designed to prevent the creation of a dominant position via acquisition. It thus
plays an important proactive role. The GenAI model market has not seen large acquisitions.54 It
has seen a wave of partnerships, the most eye-catching being Microsoft–OpenAI. But the EU
Merger Regulation only applies when there is a change of ‘control’, meaning one firm acquires
decisive influence over the other.55 The Microsoft–OpenAI partnership appears structured to avoid
such influence (Microsoft has a minority share of 49%), though the leadership struggle at OpenAI
demonstrated Microsoft’s significant behind-the-scenes influence.56 As a result, several
competition authorities are probing the partnership.57 Their vigilance is certainly welcome, as there
is a risk that the disruption brought by GenAI is ‘co-opted’ by incumbents.58

b. Digital Markets Act

GenAI models are not on the DMA’s list of ‘core platform services’ (CPSs) and are therefore not
subject to its rules.59 Other layers of the GenAI chain are, to varying degrees, subject to the DMA.
GenAI applications, such as chatbots, may be subject to the DMA as ‘online search engine’ or
‘virtual assistant’. As for the infrastructure layer, ‘cloud computing services’ are on the list CPSs.
But at the time of writing, the EC had not designated any cloud providers as gatekeepers. Their
omission was notable,60 and is likely due to definitional and user number difficulties.61 Cloud
providers may still be designated following a market investigation, which some countries are
reportedly pushing for.62 At the same time, they are already regulated by the aforementioned Data
Act.

54
The infrastructure layer has seen such acquisitions, for instance of Arm by Nvidia. The transaction was called off
after regulatory scrutiny and opposition, see Federal Trade Commission, ‘Statement Regarding Termination of
Nvidia Corp.’s Attempted Acquisition of Arm Ltd.’ (press release, 14 February 2022) <https://www.ftc.gov/news-
events/news/press-releases/2022/02/statement-regarding-termination-nvidia-corps-attempted-acquisition-arm-ltd>.
55
Council Regulation (EC) No 139/2004 of 20 January 2004 on the control of concentrations between undertakings
(EC Merger Regulation), article 3.
56
See Charles Duhigg, ‘The Inside Story of Microsoft’s Partnership with OpenAI’ (The New Yorker, 1 December
2023) <https://www.newyorker.com/magazine/2023/12/11/the-inside-story-of-microsofts-partnership-with-openai>.
57
Competition & Markets Authority, ‘CMA seeks views on Microsoft’s partnership with OpenAI’ (press release, 8
December 2023) <https://www.gov.uk/government/news/cma-seeks-views-on-microsofts-partnership-with-openai>
and Federal Trade Commission, ‘FTC Launches Inquiry into Generative AI Investments and Partnerships’ (press
release, 25 January 2024) <https://www.ftc.gov/news-events/news/press-releases/2024/01/ftc-launches-inquiry-
generative-ai-investments-partnerships>.
58
See Mark Lemley and Matthew Wansley, ‘Coopting Disruption’ (2024) <https://ssrn.com/abstract=4713845>.
59
Regulation (EU) 2022/1925 of the European Parliament and of the Council on contestable and fair markets in the
digital sector (Digital Markets Act), articles 2(2) and 3(1).
60
The only other CPS category without designation is virtual assistants, see EC, ‘Digital Markets Act: Commission
designates six gatekeepers’ (press release, 6 September 2023) IP/23/4328.
61
On those difficulties, see Friso Bostoen ‘Understanding the Digital Markets Act’ (2023) 68 Antitrust Bulletin 263,
275–277.
62
Luca Bertuzzi, ‘Are EU regulators ready for concentration in the AI market?’ (Euractiv, 3 November 2023)
<https://www.euractiv.com/section/artificial-intelligence/news/are-eu-regulators-ready-for-concentration-in-the-ai-
market/>.

10

Electronic copy available at: https://ssrn.com/abstract=4756641

The European Parliament has called for a market investigation to examine whether GenAI services
need to be added to the list of CPSs.63 Is the regulation of GenAI under the DMA, beyond its
current coverage, justified? Conceptually, if the GenAI model market turns out to be a natural
monopoly (with one large provider) or evolves into an oligopoly (with, say, 2–4 providers), there
may be a case for sectoral regulation, as antitrust law is ill-suited to either situation.64 But the
current market dynamics do not clearly point towards that outcome.
This relates to the timing question discussed above. New markets are not an easy or even suitable
target for sectoral regulation. This is in part because such regulation demands a deep understanding
of the market dynamics and competition issues, which requires (enforcement) experience and,
therefore, time to acquire. In that sense, Scott Morton pointed out that the DMA’s rules ‘only apply
to fairly old technology, like search and operating systems and social networks, that have been
around for 15 years or more, so we understand them reasonably well.’ The DMA itself has a built-
in time lag: before being designated a gatekeeper, a firm must enjoy an ‘entrenched and durable
position’, which means its CPS must have had 45 million EU users for at least three years.65

IV. Conclusion
We sought to contribute to the ongoing debate on competition in GenAI by, in the first place,
asking the right questions. When it comes to regulation, we believe those questions relate to the
trajectory of GenAI, the timing of intervention, and the regulatory tools relied on. Answering these
questions is more difficult. First, GenAI has followed a clear platformization trajectory. It is
unclear, however, whether that platformization will also end in significant concentration. The trend
towards ever-larger models points in that direction, but there are signs of a reversal. Second, in
terms of timing, the choice is essentially between regulating too early and too late. As enforcers
have erred towards the latter in web 2.0 markets, they appear determined to get in early now. But
intervention requires understanding, which is still limited. One strategy to solve this dilemma is to
progressively dial up enforcement, starting with information-gathering initiatives. Third, the
available tools range from competition law to more sectoral regulation like the DMA. Each has
their benefits (and costs), and the choice between them ties in with timing. The DMA itself only
came after widespread antitrust scrutiny, and we do not recommend skipping that step now.

63
European Parliament, Resolution of 16 January 2024 on competition policy – annual report 2023
(2023/2077(INI)).
64
Oversimplifying: In case of natural monopoly, the monopoly is efficient and should be regulated as such—not
through the introduction of competition. In case of oligopoly, no single firm may be dominant, and they may not
even be ‘collectively dominant’, but their conduct may nevertheless be anticompetitive.
65
Digital Markets Act, article 3(2)(c).

11

Electronic copy available at: https://ssrn.com/abstract=4756641