Issue 3 | September 2024 ISBN 978-92-68-16497-6, ISSN: 2315-3113

Competition Policy Brief

Competition in Generative AI and These technologies are
redefining the way we all
Virtual Worlds interact, work and consume In a nutshell
content and are widely Generative AI and virtual
Klaus Kowalski, Cristina Volpin, and Zsolt Zombori* presumed to enhance worlds technologies are set
automation and improve
to have a profound impact on
Introduction productivity. It is also
many industries. While they
Generative artificial intelligence (AI) refers to neural networks expected that together,
will bring many positive
that can generate high-quality text, images, and other forms of generative AI and virtual
changes, they could also give
content based on the data they were trained on. Differently from worlds will create
rise to competition concerns.
traditional AI, generative AI models can process inputs to produce immersive and intelligent
new content by predicting the likelihood of data typically interactions, which will Some of these concerns may
appearing together. 1 have an impact on many arise in connection with key
sectors, among others inputs to these technologies,
Virtual worlds are persistent, immersive environments, based on manufacturing, retail,
such as data, AI accelerator
technologies including 3D and extended reality (XR), which make finance, education, energy
chips, computing
it possible to blend physical and digital worlds in real-time, for a and healthcare. The
infrastructure, cloud capacity
variety of purposes such as designing, making simulations, unprecedented and fast
and technical expertise.
collaborating, learning, socialising, carrying out transactions or transformation these
Others may relate to the
providing entertainment. 2 technologies are likely to
deployment and distribution
bring raises many
of these technologies.
questions, including in
* European Commission – Directorate-General for Competition – relation to competition
Directorate for Information Technology, Communication and Media. The The European Commission
authors are grateful to Brice Allibert, Inge Bernaerts, Friedrich Wenzel policy and enforcement.
remains alert to potential
Bulst, Kassiani Christodoulou, Thomas Kramler, Luca Manigrassi, Linsey
anticompetitive practices and
McCallum, Neale McDonald, Emily O’Reilly, Carlota Reyners Fontana, DG Competition is deeply
Annemarie Ter Heegde, Paolo Tomassi, Joao Vareda and Marc Zedler, is committed to keeping
committed to
for their precious comments and contribution to the preparation of this these sectors competitive
brief. understanding how these
1 transformative and contestable via antitrust,
See Regulation (EU) 2024/1689 of the European Parliament and of the
Council of 13 June 2024 laying down harmonised rules on artificial technologies will reshape merger control and the DMA.
intelligence and amending (Artificial Intelligence Act) OJ L, 2024/1689, the EU´s economy and
12 July 2024, recitals 99 and 105; European Commission’s Living
Guidelines on the Responsible Use of Generative AI in Research, ERA potentially improve
Forum Stakeholders’ document, available at https://research-and- productivity and competitiveness and is keen to ensure that
innovation.ec.europa.eu/document/download/2b6cf7e5-36ac-41cb- citizens, small and large businesses can enjoy the benefits that
aab5-0d32050143dc_en?filename=ec_rtd_ai-guidelines.pdf, p. 3. See
also G7 Competition Authorities and Policymakers’ Summit Digital
competitive generative AI and virtual worlds markets can bring, in
Competition Communiqué, 8 November 2023, available at terms of price, choice, innovation and quality. With that goal in
https://www.bundeskartellamt.de/SharedDocs/Publikation/EN/Others/G7 mind, the Commission launched two calls for contributions on
_2023_Communique.pdf?__blob=publicationFile&v=2. competition in virtual worlds and generative AI, open from
2
See Commission Staff Working Document, An EU initiative on Web 4.0
and virtual worlds: a head start in the next technological transition,
available at https://digital-strategy.ec.europa.eu/en/library/staff-working
-document-information-insights-and-market-trends-web-40-and-
virtual-worlds, pp. 3 and 87.

The content of this article does not necessarily reflect the KD-AK-24-003-EN-N, doi 10.2763/679899
official position of the European Commission. Responsibility ©European Union, 2024
for the information and views expressed lies entirely with Reproduction is authorised provided the source is
the authors. acknowledged. More publications on:
https://ec.europa.eu/competition-policy/publications_en and
http://bookshop.europa.eu
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

9 January to 11 March 2024. 3 Interested stakeholders were Hungarian, Portuguese, and the UK competition authorities, as
invited to share their experience and provide feedback on well as the US Federal Trade Commission.
competition in these sectors and their insights on how
competition law can help ensure that these sectors remain While the role of competition enforcement in preserving
competitive. DG Competition received around 120 contributions competitive generative AI and virtual worlds markets is clearly
on generative AI. Most of the contributions came from companies important, it should be noted that the way in which market
or business representatives. The remaining contributions were dynamics and competition will unfold in relation to these
submitted by academics, policy groups, competition authorities technologies is susceptible to be affected by many other factors,
and regulators, civil society, other government representatives including regulation on policy aspects different from competition,
and law firms. On virtual worlds, DG Competition received just such as, for instance, AI safety, data and copyright law. Some of
over 50 submissions from businesses, academia, national these aspects are discussed below in the section on Other factors
competition authorities, regulatory bodies, associations, and promoting competition in generative AI related markets.
citizens.
Generative AI
As a follow-up to the calls for contributions, DG Competition
organised a workshop on 28 June 2024, which brought together Market dynamics and potential barriers to entry and
different perspectives emerging from the contributions and expansion
facilitated exchanges and the sharing of insights on the This policy brief focuses on competition dynamics and potential
complexities of competition dynamics within virtual worlds and concerns in generative AI related markets. It does not discuss AI
generative AI as well as the challenges, opportunities, and as an enforcement tool or AI as a possible tool to facilitate
regulatory implications arising from the evolving landscape of anticompetitive conduct in other markets, as in the context, for
these digital innovations. 4 The workshop featured three panel instance, of algorithmic collusion.
discussions, each addressing key themes and issues pertinent to
competition in virtual worlds and generative AI. The panel The generative AI sector, which encompasses several markets at
discussions are available on the YouTube channel of the upstream and downstream levels of the provision of
DG Competition. generative AI models, 6 is currently described by players and
stakeholders as very dynamic, with a rich R&D activity and many
As regards generative AI, DG Competition has also engaged in a players of different sizes along the supply chain of AI foundation
thorough analysis of several investments and partnerships model development and deployment.
between large digital players and generative AI developers, both
from the merger control and the antitrust viewpoint, including by The responses to the call for contributions and the on-going
sending requests for information to relevant players. 5 It also sent market investigations identified some market tendencies, which
requests for information to better understand whether will shape generative AI related markets in the future, as well as
agreements between large digital players and original equipment several potential bottlenecks, i.e. parts of the generative AI
manufacturers (OEMs) for the pre-installation of AI foundation production or supply chain that can be more vulnerable to
models on new devices may raise anticompetitive concerns. anticompetitive practices and negatively affect the overall well-
functioning, performance, and efficiency of generative AI related
This policy brief is based on the responses to the calls for markets. These bottlenecks may be due to resource constraints,
contributions on competition in virtual worlds and generative AI, technological barriers (such as a lack of interoperability) or
the follow-up workshop, and, in relation to generative AI, also on market access issues, which may reduce the presence of active
interviews with key stakeholders and parallel market competitors. Depending on the context, some of these
investigations. It has benefitted from fruitful collaboration and bottlenecks may qualify as barriers to entry and expansion or
exchange with other competition authorities engaged in the lead to an anticompetitive practice.
analysis of these issues around the world, including the French,

3
See DG COMP’s Calls for Contributions on Competition in Virtual Worlds
and Generative AI, 9 January 2024, available at https://competition-
policy.ec.europa.eu/about/europes-digital-future_en and https:// 6
Based on the feedback from the call for contributions and other
competition-policy.ec.europa.eu/system/files/2024-01/20240109_call- information collected, for the purpose of this policy paper, the
for-contributions_virtual-worlds_and_generative-AI.pdf. generative AI sector should be understood as the value chain of
4
See DG COMP’s Workshop on Competition in Virtual Worlds and generative AI models, which may indicatively include, among others,
Generative AI, 28 June 2024, available at https://competition-policy. and without prejudice to the market definition assessment performed
ec.europa.eu/about/reaching-out/virtual-worlds-and-generative-ai_en. in concrete cases, the following markets: chips manufacturing,
5
See DG COMP’s Press Release of 9 January 2024, IP/24/85, provision of cloud infrastructure, data licensing, supply of specific types
Commission launches calls for contributions on competition in virtual of AI workforce, the supply of productivity software, supply of specific
worlds and generative AI, available at https://ec.europa.eu/commission/ chatbot services, supply of specific mobile phone digital assistant
presscorner/detail/en/IP_24_85. services, etc.

2
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

Market tendencies Tendency towards vertical integration or establishing

There are several emerging tendencies characterising generative partnerships to access distribution channels
AI related markets that seem to be prevailing at the time of Linked to the above, often, partnerships also offer AI foundation
writing and may be relevant from a competition perspective: 7 model developers a more direct access to customers and
consumers via, for example, integration of the AI foundation
Tendency towards vertical integration or establishing model functionality into an established product of a larger digital
partnerships to access input resources player, or more targeted access to customers or consumers. From
The first noticeable tendency in this sector is linked to the the competition viewpoint, this may be efficient if it gives smaller
presence of established vertically integrated players, typically players wider outreach and direct access to the distribution
offering, in addition to proprietary AI foundation models, also network of the larger players. It may also, however, raise
cloud or data centre services at the upstream level, and AI concerns in relation to the risks of abuse of dominance by
systems and applications to customers and final consumers at established large players aiming at foreclosing competitors, for
the downstream level. This is the case, for example, of Google, instance, through control over distribution channels for generative
Amazon and Microsoft. These players combine large financial AI applications or services. A similar approach may also be the
resources with access to inputs that are key for offering AI rationale for mergers and acquisitions leading to the integration
services and the direct knowledge of ongoing customers’ needs, of players into an ecosystem. 10
for example, in terms of IT services, privacy needs, data storage
requirements or fine-tuning demands. 8 Tendency towards more efficient, smaller models
Another important tendency that is likely to shape the generative
Innovative small AI foundation model developers also often AI sector in the coming years is the race to produce smaller AI
choose to secure privileged access to cloud computing power by foundation models that can efficiently run on mobile devices
establishing partnerships with existing digital players. 9 These such as tablets and smartphones, locally and without internet
partnerships are aimed at providing them with access to these connection (i.e. offline). This tendency promises to be
important inputs. From the competition viewpoint, this may be transformative for the industry because, if confirmed, it would
beneficial when it allows smaller and more innovative players to mark the beginning of a parallel tendency to the so far strongly
enter and grow in the market by securing access to important predominant principle of ‘scaling laws’ which characterises the
inputs. It may, however, also raise concerns in relation to risks of industry, according to which the performance of an AI foundation
concentration of key inputs in the hands of few players and model is a function of the model size (which depends on the
creation of dependencies between players, which may lead to number of parameters or “weights” of the model) and pre-
foreclosure issues and have negative impacts on competition for training dataset. Due to the importance of economies of scale
consumers. and scope in this sector, it has been considered until recently that
only AI foundation model developers with a critical mass can
effectively recoup initial investments and be competitive in the
market. 11
7
See responses to Questions 9 and 10 of DG COMP’s Calls for
Contributions on Competition in Virtual Worlds and Generative AI,
Running smaller and more efficient AI foundation models on
9 January 2024, available at https://competition-policy.ec.europa.eu/ mobile devices is becoming a commercial reality, 12 with varying
about/europes-digital-future_en. levels of performance and integration depending on the
8
See on this point also French competition authority, Opinion 24-A-05
on the competitive functioning of the generative artificial intelligence
10
sector, 28 June 2024, available at https://www.autoritedelaconcurrence According to the European Commission’s Notice on the definition of the
.fr/en/opinion/competitive-functioning-generative-artificial-intelligence- relevant market for the purposes of Union competition law, 8 February
sector, para. 220; the Portuguese competition authority, Issues Paper 2024, C(2023) 6789 final, https://eur-lex.europa.eu/legal-content/
on competition and generative AI, 6 November 2023, available at EN/TXT/PDF/?uri=PI_COM:C(2023)6789&qid=1726475579651, p. 40,
https://www.concorrencia.pt/sites/default/files/documentos/Issues%20 “(Digital) ecosystems can, in certain circumstances, be thought of as
Paper%20-%20Competition%20and%20Generative%20Artificial%20 consisting of a primary core product and several secondary (digital)
Intelligence.pdf, pp. 16 and 25; and the UK CMA’s Update Report on AI products whose consumption is connected to the core product, for
Foundation Models, 11 April 2024, available at https://assets.publishing. instance, by technological links or interoperability”.
11
service.gov.uk/media/6617ef792b2963dfa2d1ea6d/Update_Paper.pdf, See also the French Opinion 24-A-05, fn. 8, p. 63; the Portuguese
p. 8. Issues Paper, fn. 8, pp. 16 and 25; and the UK CMA’s Update Report,
9
Examples include, for instance, OpenAI exclusive computing partnership fn. 8, p. 8. See also Azeem Azhar, Substack, AI’s USD 100 billion
with Microsoft to build new Azure AI supercomputing technologies, question: the scaling ceiling, 13 July 2024, available at https://www.
https://news.microsoft.com/2019/07/22/openai-forms-exclusive- exponentialview.co/p/can-scaling-scale?utm_source=substack&utm_
computing-partnership-with-microsoft-to-build-new-azure-ai-super medium=email.
12
computing-technologies/; Amazon and Anthropic strategic collaboration See, for instance, the agreement between Samsung Galaxy and Google
to advance generative AI, https://www.aboutamazon.com/news/ Cloud to deploy Google’s Gemini Pro and Imagen 2 on Vertex AI via the
company-news/amazon-aws-anthropic-ai; Microsoft and Mistral AI cloud to the Samsung Galaxy S24 Series, Press Release, Samsung,
partnership to accelerate AI innovation and introduce Mistral Large first 18 January 2024, Samsung and Google Cloud Join Forces to Bring
on Azure, https://azure.microsoft.com/en-us/blog/microsoft-and-mistral- Generative AI to Samsung Galaxy S24 Series, https://news.
ai-announce-new-partnership-to-accelerate-ai-innovation-and- samsung.com/global/samsung-and-google-cloud-join-forces-to-bring-
introduce-mistral-large-first-on-azure/. generative-ai-to-samsung-galaxy-s24-series.

3
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

characteristics of the hardware. Such development may bring sector, in particular for startups, unless they benefit from
significant improvements in terms of speed, privacy and data substantial funding by large digital platforms. Finally, while
protection, but may also raise concerns linked to exclusivity useful, it is also generally considered that synthetic data (i.e.
agreements and default pre-installation of specific foundation artificial data generated by large language models) cannot be a
models on popular device brands. complete substitute for high-quality real-world data. The more
the datasets are complete, accurate, relevant, and unique, the
Tendency towards the parallel development of open source more they are considered valuable. 15
and proprietary models
A further tendency that is emerging is the development of AI Second, the cost and scarcity of the specialised chips supporting
foundation models that present varying degrees of openness. AI neural networks, such as GPUs, TPUs and other AI accelerators,
This includes (i) fully open models (where access to the source may also represent an important bottleneck. Waiting times for
code and training weights and data is provided); (ii) partially open buying an H100 chip (used for AI applications) from the biggest
models (where access to model-trained weights is provided but manufacturer at present (Nvidia), for example, were nearly 12
not to the source code); (iii) closed source models (accessible via months at end of 2023. This waiting time is now reduced to 3-4
APIs, i.e. access is given to the output but not to the source code); months, but costs remain quite high (reported to be up to 30 or
and (iv) fully closed (where no access to source code, weights or 40 thousand US dollars per unit, and possibly more for the faster
output is provided). GB200 NVL72), 16 especially if one considers that thousands of
GPUs are required to train and operate AI foundation models. 17
Some players focus on open-source models (such as Meta with
Llama), while others (such as Mistral AI, Open AI or Google) Computing capacity, in the form of large clusters of chips, can be
develop both an open-source and a proprietary version of their made available in the form of physical data centres or on the
models. This has a significant impact on the market dynamics, cloud (or a mix of these two options). Despite the availability of
potentially reducing the barriers to entry in some of the many compute power providers, costs remain very high also for
generative AI related markets and promoting innovation and cloud capacity, which also consume a very substantial amount of
choice. energy. Several partnerships concluded in this industry have
precisely the objective of providing promising startups with
Potential barriers to entry in generative AI related markets compute capacity at cost in exchange for access to the startups’
The consultation and the on-going market investigations revealed AI technology and intellectual property. For instance, press
that the key components for the development and deployment of reports describe partnerships between Microsoft and OpenAI,
generative AI systems include data, AI accelerator chips, Microsoft and Mistral AI, as well as Amazon and Anthropic as
computing infrastructure, cloud capacity and technical concluded for this purpose. 18 Some companies are also racing to
expertise. 13 As mentioned above, depending on the economic strategically secure access to renewable energy to meet the
context, each of these may qualify as a potential barrier to entry
or expansion, 14 or potentially lead to an anticompetitive practice. 15
See responses to Questions 1 and 7 of DG COMP's Calls for
Contributions on Competition in Virtual Worlds and Generative AI (see
First, in relation to data, the necessity of very large, high-quality above fn. 3).
16
Anton Shilov, Nvidia's next-gen Blackwell AI Superchips could cost up to
datasets for pre-training, the preparatory phase of the AI
$70,000 — fully-equipped server racks reportedly range up to
foundation model, may represent a bottleneck. Accessing such $3,000,000 or more, Tom’s Hardware, 14 May 2024, available at
data appears very costly and availability seems limited. According https://www.tomshardware.com/pc-components/gpus/nvidias-next-gen-
to the consultation, several factors are contributing to this. One blackwell-ai-gpus-to-cost-up-to-dollar70000-fully-equipped-servers-
range-up-to-dollar3000000-report.
factor is that, despite successful pre-training experience on public 17
Anton Shilov, Nvidia's H100 AI GPU shortages ease as lead times drop
data, navigating the uncertainty on the application of copyright from up to four months to 8-12 weeks, Tom’s Hardware, 10 April 2024,
laws and the need for new, specialized datasets may pose available at https://www.tomshardware.com/pc-components/gpus/
nvidias-h100-ai-gpu-shortages-ease-as-lead-times-drop-from-up-to-
challenges in the future. Another factor is the costs of data
four-months-to-8-12-weeks; Anton Shilov, Wait times for Nvidia's AI
licensing agreements between large players and the right holders GPUs ease to three to four months, suggesting peak in near-term
of high-quality online content, such as publishers and news growth — the wait list for an H100 was previously eleven months,
outlets, which are at the moment high enough to pose a Tom’s Hardware, 16 February 2024, available at https://www.toms
hardware.com/tech-industry/artificial-intelligence/wait-times-for-
significant barrier to entry in the AI foundation model developing nvidias-ai-gpus-eases-to-three-to-four-months-suggesting-peak-in-
near-term-growth-the-wait-list-for-an-h100-was-previously-eleven-
months-ubs; Anton Shilov, Nvidia's H100 AI GPUs cost up to four times
13
See responses to Question 1 of DG COMP's Calls for Contributions on more than AMD's competing MI300X — AMD's chips cost $10 to $15K
Competition in Virtual Worlds and Generative AI (see above fn. 3). See apiece; Nvidia's H100 has peaked beyond $40,000: Report, Tom’s
also French Opinion 24-A-05, fn. 8, paras. 122 and ss.; the Portuguese Hardware, 2 February 2024, available at https://www.
Issues Paper, fn. 8, pp. 14 and ss.; and the UK CMA’s Update Report, tomshardware.com/tech-industry/artificial-intelligence/nvidias-h100-ai-
fn. 8, p. 6-7. gpus-cost-up-to-four-times-more-than-amds-competing-mi300x-
14
See responses to Questions 2 and 3 of DG COMP's Calls for amds-chips-cost-dollar10-to-dollar15k-apiece-nvidias-h100-has-
Contributions on Competition in Virtual Worlds and Generative AI (see peaked-beyond-dollar40000.
18
above fn. 3). See above fn. 9.

4
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

growing energy needs of data centres, which are needed for the applications to significantly benefit from feedback data loops,
operation of AI applications. This is the case, for instance, of the particularly in multimodal AI foundation models. Feedback data
Microsoft and Brookfield agreement for more than 10.5 loops can, for instance, occur in the deployment phase, when the
gigawatts (reported to be almost eight times larger than the model continues its learning from the interaction with the user
largest power purchase agreement ever signed by a company). 19 (e.g. correction or precision of the prompts, forms of feedback,
etc.) or from the data uploaded by the users for further
Third, a recurring issue is also the difficulty of finding highly elaboration (e.g. requests for translation of an uploaded pdf file).
skilled labour in generative AI. While startups may have the
ability to offer equity shares to attract talent, some respondents As data feedback loops may in turn reinforce network effects, the
noted the difficulties of competing with the salaries and latter may be significant because the models with the largest
conditions offered by large players, especially US-based. The user base are more likely to provide accurate answers and, in
consultation did not highlight a widespread use of non-compete turn, attract more users. 21 This may be the case, for example, for
clauses restricting the mobility of AI engineers to other jobs and AI applications with plug-ins which facilitate access to different
the market for the acquisition of AI talent seems to be dynamic services, because the higher the number of plug-ins, the more
at this stage. It highlighted, however, a general scarcity of highly attractive that specific AI interface will become. Such effects may
qualified experts in the AI field. The intense fight to procure this also appear or be boosted when generative AI applications are
precious input may thus lead to ”acqui-hires”, i.e. transactions integrated into larger digital ecosystems of other products and
whereby a player acquires all or almost all the key employees of services, thanks to data integration, seamless user experience
the target company, such as in the recent case of Microsoft and and fewer interoperability issues (see further on generative AI
Inflection where almost all of Inflection´s employees, in addition ecosystems below in the section on Ecosystems dynamics in the
to the chief executive of AI, were hired by Microsoft together with generative AI industry). 22 Important network effects may
a non-exclusive license. 20 It may also lead to attempts to therefore end up constituting barriers to entry for new rivals and
foreclose competitors by making it difficult to secure talent contribute to the creation or the strengthening of market power
needed to establish oneself or grow in the industry. in these markets.

Other potential barriers to entry and expansion Barriers to entry may also affect neighbouring markets, which
The generative AI sector also seems to present other barriers to may then indirectly also harm competition in generative AI
entry and expansion. Some of these are characteristic of digital markets. For instance, switching cloud service providers seems to
markets in general, such as economies of scale and scope, and remain challenging, due to pricing strategies, such as cloud
the presence of ecosystems of large players which may facilitate credits and committed spend discounts, the lack of
the penetration of neighbouring markets. interoperability and data portability, and various software
licensing practices, which may lead to customers’ lock-in. 23
Other characteristics are quite different and mark a substantial DG Competition has come across similar allegations in its on-
difference in the analysis of the anticompetitive risks potentially going investigations in the cloud sector.
arising in generative AI related markets. For instance, the fact
that there are, at least at present, very high marginal costs Competition enforcement in generative AI related
attached to the actual operation of the model (e.g. answering a
markets – A reference framework
prompt), alongside the high initial fixed ones for the pre-training
Competition authorities have an important role in preserving
of the foundation model which strongly favour large players.
competition in generative AI related markets, both at the
According to respondents and interviewees, some uncertainty upstream and the downstream level, and in preserving choice and
remains around how powerful data feedback loops and network innovation for consumers in a pivotal sector for the future of the
effects will be and whether they will constitute barriers to entry European economy.
and be associated with anticompetitive concerns as has been the
case in the past in relation to some digital markets.

21
While data cannot be continuously fed into a model after the pre- See also French Opinion 24-A-05, fn. 8, p. 63; and the Portuguese
training phase to perfect it, due to the costs and time required to Issues Paper, fn. 8, p. 40; and the UK CMA’s Update Report, fn. 8,
pp. 20-21.
pre-train a model, it still seems possible for AI models and 22
See also French Opinion 24-A-05, fn. 8, p. 58 and ss.; the Portuguese
Issues Paper, fn. 8, pp. 16 and 25; and the UK CMA’s Update Report,
19
Brookfield Press Release, Brookfield and Microsoft Collaborating to fn. 8, p. 20.
23
Deliver Over 10.5 GW of New Renewable Power Capacity Globally, See, for instance, French Competition Authority, Opinion 23-A-08 of
1 May 2024, available at https://bep.brookfield.com/press-releases/bep/ 29 June 2023 on competition in the cloud sector, available at
brookfield-and-microsoft-collaborating-deliver-over-105-gw-new- https://www.autoritedelaconcurrence.fr/en/opinion/competition-cloud-
renewable-power. sector, and Competition and Markets Authority, working papers in the
20
Tabby Kinder, Microsoft hires DeepMind co-founder Mustafa Suleyman cloud services market investigation, published on 23 May 2024,
to run new consumer AI unit, Financial Times, 19 March 2024, available at https://www.gov.uk/cma-cases/cloud-services-market-
https://www.ft.com/content/5feedf3a-ff7a-4c89-9b1d-f9b48834ff4c. investigation#working-papers.

5
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

The primary focus of competition authorities has been to ensure To the extent that such concentrations do not meet the turnover
that conduct and transactions by generative AI players remain thresholds under the EU Merger Regulation, the Commission will
compliant with competition law. DG Competition has been vigilant work together with Member States and the parties to assess
to catch early on any potential anticompetitive issues that could whether they will be reviewed under national merger control
emerge in these markets, both from the antitrust and, where regimes or referred to the European Commission in line with the
appropriate, the merger control viewpoint. legal requirements for such referrals as clarified in the recent
Illumina judgment of the European Court of Justice. 26
As regards antitrust enforcement in generative AI, and in light of
the market tendencies and emerging risks highlighted above, The following sections provide a non-exhaustive list of possible
DG Competition is looking into possible vertical or other concerns avenues that may be adopted, depending on the circumstances,
relating to the investments and partnerships between large for framing the analysis of potential anticompetitive issues that
digital players and generative AI developers. may arise in generative AI related markets (upstream and
downstream), within the current EU competition law framework. 27
It is also looking into whether agreements between Google and
OEMs (such as Samsung) for the pre-installation of Google’s Possible relevant market definitions and useful criteria for the
small AI model “Gemini Nano” on mobile devices may raise identification of market shares
anticompetitive concerns by making it more difficult for other Whilst defining relevant markets is an exercise that can only be
foundation models to be accessed or pre-installed on those conducted in the context of a full-fledged competition
devices. investigation, the information gathered by DG Competition so far
suggests that the following elements could be starting
DG Competition is also monitoring from the merger control hypotheses for the purpose of relevant market definition in
viewpoint investments and partnerships between large digital investigations in the AI sector.
players and generative AI developers, including the one involving
Microsoft and OpenAI. While it was preliminarily concluded that, It may be useful to distinguish between the upstream and the
even following the firing and re-hiring of the CEO of OpenAI and downstream levels. The relevant markets upstream may typically
the granting of an observatory seat for Microsoft on the OpenAI be those for the purchase by the AI developer of the necessary
board in November 2023, Microsoft did not acquire control of inputs. Examples would include the possible market for the
OpenAI on a lasting basis, DG Competition is keeping this and purchase of high-quality data for pre-training, the possible
other partnerships under close scrutiny. market for the purchase of cloud capacity from cloud services
providers, the possible market for the purchase of data centre
DG Competition also monitors whether transfers of highly skilled services, or the possible market for attracting AI engineering
employees between two undertakings (sometimes via acqui- talent. Depending on the specific case, and for each of these
hires), like the hiring by Microsoft of most of the employees of markets, there can of course be further segmentation and
Inflection, are subject to scrutiny under EU merger rules. While it product differentiation, for instance, in relation to data for
is for the parties to a transaction to assess whether their general purpose pre-training and specialised datasets for a
agreements need to be notified for merger control, they may specific domain in the form of model-as-a-service; in relation to
approach the Commission with a consultation in case of doubt. generalist cloud compute and specialised cloud services; in
relation to the type of skillset of the concerned employees; or in
In the case of Microsoft and Inflection, based on information relation to different types of AI accelerators (such as GPUs, TPUs,
provided to the Commission by these companies, the Commission or others).
considers that the transaction involves all assets necessary to
transfer Inflection’s position in the markets for generative AI The relevant markets downstream could typically be markets for
foundation models and for AI chatbots to Microsoft. 24 Taking the sale or supply of generative AI foundation model services.
further into account that Inflection itself announced on 19 March Further segmentation and product differentiation may include a
2024 that the “new Inflection” would shift its focus to its AI distinction, for instance, between consumer-facing and business
studio business, 25 the Commission regards the agreements models, cloud-based and on-device models, general purpose and
entered into between Microsoft and Inflection as a structural specific purpose models, or unimodal and multimodal models. At
change in the market that amounts to a concentration as defined this level, given the nascent state of the industry, where many AI
under Article 3 of the EUMR. foundation models are yet to be fully monetised 28 and therefore
revenue streams are not necessarily a proxy of market power,
24
See DG COMP’s Press Release of 18 September 2024, IP/24/4727,
Commission takes note of the withdrawal of referral requests by
26
Member States concerning the acquisition of certain assets of Joined cases C-611/22 P and C-625/22 P, Illumina v. Commission and
Inflection by Microsoft, available at https://ec.europa.eu/commission/ Grail v. Commission, 3 September 2024, EU:C:2024:677.
27
presscorner/detail/en/ip_24_4727. See responses to Question 11 of DG COMP's Calls for Contributions on
25
See Inflection’s Press Release, The new Inflection: an important change Competition in Virtual Worlds and Generative AI (see above fn. 3).
28
to how we’ll work, 19 March 2024, https://inflection.ai/the-new- See responses to Question 5 of DG COMP's Calls for Contributions on
inflection. Competition in Virtual Worlds and Generative AI (see above fn. 3).

6
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

alternative metrics for the assessment of market shares may example would be that of a large vertical buyer 32 or a
have to be considered. At present, these could include, for dominant player entering into an exclusive licensing
instance, the following: agreement to secure high-quality data content from one
specialised upstream source, should this prevent other
(1) The activity volume, such as the average number of players from pre-training or fine-tuning a model in that
users, the number of interactions or of tokens (units of specific domain. A further example would be a dominant
data) processed, which would suggest the frequency of integrated player active both in cloud and at the level of
use of a given AI foundation model; generative AI development refusing to give access to its
(2) The costs and power of the processing capacity used cloud infrastructure to competitors with the effect of
for inferencing 29 (as one of the possible proxies of foreclosing those competitors. Another example would be a
quality of performance of the models); dominant generative AI player with deep pockets engaging in
(3) The number of purchases or downloads of the model acqui-hires or predatory hiring of talent, with the effect of
(or downloads of the original model and number of foreclosing competitors by hiring their AI key staff. This
models derived from the original one for the open- practice may lead de facto to the dissolution of an active or
source ones); of a potential competitor by absorbing the expert talent in
(4) Other specific deployment parameters (such as the its entirety or near-entirety, thus leading to a loss of choice
number of calls to one’s APIs). or innovation or the reduction of competitive constraints
downstream. 33
Because of their importance in generative AI related markets, (2) The risk that large players offering generative AI foundation
market definition may also need to take into account network models may use their market power to limit choice or distort
effects and ecosystem dynamics. As noted in the European competition in downstream markets, when distributing and
Commission’s revised Market Definition Notice in relation to commercialising AI applications. This can take the form, for
digital ecosystems, 30 the presence of network effects, switching instance, of exclusivity clauses or leveraging behaviour,
costs and customers’ lock-in, and single- or multi-homing might including self-preferencing, refusal to supply, tying or
have to be part of the analysis (see further below the section on bundling, non-compete and lock-in strategies. 34 One example
Ecosystems dynamics in the generative AI industry). would be one where a large player would offer its own AI
foundation model on its marketplace alongside other models
Possible theories of harm
and would put in place conditions that lead to offer a
Based on the responses to the call for contributions and on-going
preferential display of its AI foundation model, while
market investigations, some potential anticompetitive concerns
demoting the display of the competitors’ ones. Another
that may materialise in generative AI related markets in the
example could involve the tying of a distinct digital product,
future were raised.
such as, for instance a search engine, to an AI foundation
DG Competition will therefore remain vigilant, particularly (but model owned by the same digital player, thus forcing
not exclusively) in relation to following five possible types of consumers to acquire or use both products. 35
competition risks in these markets: 31 (3) The risk that agreements between horizontal competitors
may reduce competitive constraints between such players or
(1) The risk that incumbent large digital players, which may enable unlawful exchange of commercially sensitive
currently enjoy preferential access to generative AI’s key information.
components, grant it to third parties on an exclusive basis, or (4) The risk that vertically integrated players may adopt pricing
prevent competitors from accessing it. This may affect policies for the purpose of margin squeezing other players.
access to any of the key inputs described above, including (5) The risk of killer or reverse killer acquisitions, that is
computing infrastructure (such as GPUs, supercomputing acquisitions aimed at eliminating nascent competition in
power, and cloud capacity), data or talent. One example generative AI related markets to protect the acquirer’s
would be that of a large digital player reserving exclusive
access to its AI computing infrastructure to a specific player 32
For an analysis of vertical foreclosure, see also Commission Regulation
or providing preferential access to a certain player and (EU) 2022/720 of 10 May 2022 on the application of Article 101(3) of
access at worse conditions to all other players. Another the Treaty on the Functioning of the European Union to categories of
vertical agreements and concerted practices (VBER) and the Guidelines
on vertical restraints, OJ C 248, 30 June 2022, p. 1–85.
33
Depending on the concrete assessment, this conduct may be relevant
29
Inferencing is typically the process of using a trained generative AI under merger control or antitrust scrutiny.
model to generate output, for instance, making predictions or 34
For companies offering online intermediation services, see VBER and
answering prompts based on the data it has been fed during the pre- Vertical Guidelines, fn. 32 above.
training phase. 35
Some of these theories of harm have already been applied in other
30
European Commission’s Notice on the definition of the relevant market, digital markets (for instance, self-preferencing in Google Shopping and
fn. 10, p. 40. exclusivity - in the form of anti-fragmentation - and tying in Google
31
See responses to Question 4 of DG COMP's Calls for Contributions on Android) and they are likely to be a suitable reference also in the
Competition in Virtual Worlds and Generative AI (see above fn. 3). generative AI domain.

7
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

position in those markets, which may harm choice and As mentioned above in the section on Possible relevant market
innovation in the longer term at that level of the supply definitions and useful criteria for the identification of market
chain. shares, if an ecosystem approach is considered appropriate in a
specific case, it may also have to be reflected in the way in which
Further, it may be noted that investments in small AI developers the relevant market is defined and market shares are accounted
by large companies are seen by the industry as important for for. This is because the assessment of market shares in a specific
developing and distributing AI systems, securing necessary relevant market, depending on the specific circumstances of the
capital, accessing intellectual property, and gaining technological case, may not be, alone, fully informative of the actual
insights. When not granting exclusivity rights, these partnerships competitive position of the analysed player.
can be pro-competitive. However, in some circumstances, they
may create the conditions for the concentration of key inputs in As recommended by the Market Definition Notice in relation to
the hands of few players, for foreclosure strategies or for other digital markets, it may also be important to consider supply and
distortions of competition, warranting monitoring by competition demand-side issues, like network effects or the capacity to multi-
authorities to maintain a level playing field. 36 This means that homing of users, to better understand market dynamics. 42
they may undergo merger control or antitrust scrutiny.
Furthermore, while not all generative AI players are structured as
Ecosystems dynamics in the generative AI industry a platform, the experience relating to multi-sided market
Given the connections between different products and services in definition may be of relevance in those cases where generative
the AI production and supply stack and the way in which AI players opt for a business model that charges only one side of
dominance in a generative AI related market may have an impact the market, for instance, by selling advertising, while offering
on several other markets, a holistic view of ecosystem dynamics some AI functionalities or services to final users for free (e.g.
in the generative AI industry may be particularly relevant as a chatbots or search engines).
way to better capture market realities.
Potential efficiency gains to be considered
The existence of dynamics within an ecosystem played an Since generative AI development and deployment are
important role, for example, in the recent DG Competition characterised by their technological innovation, due consideration
prohibition decision of the Booking/eTraveli merger. 37 The case should be given to concrete efficiency gains that may arise from
took into account, among other things, the importance of practices or agreements between undertakings in the generative
considering the specific features and effects of ecosystem AI sector, under the framework of both Article 101 and 102 TFEU.
dynamics, of network effects, as well as of behavioural biases, Specifically, regarding partnerships, examples of potential
such as defaulting, customer inertia or single-homing. 38 These efficiency gains may include:
features have also been analysed in other transactions, such as
Meta/Kustomer, 39 and in abuse of dominance cases like Google (1) combining complementary skills and assets of the involved
Android. 40 The attention to these issues is dictated by the nature players, which may result in the issuing of a better or new
of competitive dynamics in the technology space. product or technology that would not otherwise come to
light;
Some of these factors may be relevant (and likely to be applied) (2) disseminating technological expertise across the market,
also in the analysis of the competitive dynamics in generative AI which may lead to further innovation; and
related markets. 41 (3) reducing costs or dependencies when supply of a specific
input is limited, which may increase supply and strengthen
the internal market. 43
36
See also the French Opinion 24-A-05, fn. 8, paras. 290 ss. and the UK
CMA’s Initial Report on AI Foundation Models, available at Specific types of agreements may also fall under Article 101(3)
https://www.gov.uk/government/publications/ai-foundation-models- block exemptions. 44 Relevant efficiencies that benefit consumers
initial-report, reviewed by an Update Report, fn. 8, pp. 17-18 and
Figure 5.
37
European Commission, M.10615, Booking Holdings / eTraveli Group, 42
European Commission’s Notice on the definition of the relevant market,
Decision of 25 September 2023, pp. 199 and ss. fn. 10, p. 40.
38 43
Ibid., pp. 118 and ss. European Commission’s Guidelines on the applicability of Article 101 of
39
European Commission, M. 10262, Meta (formerly Facebook) / the Treaty on the Functioning of the European Union to horizontal co-
Kustomer, Decision of 27 January 2022, pp. 110 and ss. See also operation agreements, C(2023) 3445 final, p. 39.
44
European Commission’s Competition Merger Brief, Adobe/Figma: Much See, for R&D agreements, Commission Regulation (EU) 2023/1066 of
Ado(be) About Nothing?, 2/2024, forthcoming at https://competition- 1 June 2023 on the application of Article 101(3) of the Treaty on the
policy.ec.europa.eu/publications/competition-policy-briefs_en#merger- Functioning of the European Union to certain categories of research
brief. and development agreements; for specialisation agreements,
40
Case T-604/18, Google and Alphabet v Commission, 14 September Commission Regulation (EU) 2023/1067 of 1 June 2023 on the
2022, EU:T:2022:541, para. 116. application of Article 101(3) of the Treaty on the Functioning of the
41
See also See also French Opinion 24-A-05, fn. 8, para. 128; the European Union to certain categories of specialisation agreements; for
Portuguese Issues Paper, fn. 8, p. 33 and ss.; and the UK CMA’s Update technology transfer licensing, Commission Regulation (EU) No
Report, fn. 8, p. 16. 316/2014 of 21 March 2014 on the application of Article 101(3) of the

8
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

also need to be considered in the context of mergers in the AI 9 January 2024 two calls for contributions on competition and
sector. generative AI and held a workshop on 28 June 2024 to provide
an overview of the results. 46 It also sent a series of requests for
Generative AI related markets and the DMA information to several players in the generative AI sector to gain
DG Competition sees the role of competition law enforcement in a better understanding of the market dynamics.
generative AI related markets and the implementation of the
DMA as complementary, and equally important tools at its In cooperation with the US Department of Justice (DoJ),
disposal. The objective of the DMA is to ensure contestable and US Federal Trade Commission (FTC), and UK Competition and
fair markets in the digital sector where gatekeepers are present. Markets Authority (CMA), the European Commission also issued
It achieves these objectives by ordering designated gatekeepers on 23 July 2024 a joint statement on competition in generative
to comply with specific obligations in relation to a predefined list AI related markets, aimed at signalling the agencies’ common
of core platform services. readiness to prevent anticompetitive practices in this sector,
raising awareness among customers and consumers about
There are two main ways in which the DMA will be relevant for possible anticompetitive infringements they may witness, and
generative AI services. First, a generative AI player may offer a clarifying their expectations of market behaviour by relevant
core platform service and meet the gatekeeper requirements of players. 47 The European Commission also contributes its input in
the DMA. Second, generative AI powered functionalities may be discussions and meetings at the G7, promoting a level playing
integrated or embedded in existing designated core platform field for the development of AI markets.
services and therefore be covered by the DMA obligations. Those
obligations apply in principle to the entire core platform service Other factors promoting competition in generative AI related
as designated, including features that rely on generative AI. For markets
these purposes, DG Competition will continue to reassess the There are several factors that can be considered important to
situation as the services evolve, as services that do not fall under reduce potential barriers to entry or limit their effects, as well as
the DMA currently may well do so in the future because of the directly or indirectly promoting competition in generative AI
integration of AI powered services into core platform services. related markets. Some examples include:

The DMA will not however cover all competition concerns related (1) The presence of open-source models, which can promote
to generative AI services, in particular for non-designated choice and innovation in the sector and lower barriers to
companies and for conduct which goes beyond the scope of the entry. 48
DMA. (2) The availability of freely or easily accessible high-quality
databases, which can support the pre-training phase for the
Competition advocacy and other pro-competitive development of different AI foundation models and can also
initiatives in generative AI related markets promote wider offerings.
(3) The availability of freely accessible public supercomputers to
Competition authorities’ activities
researchers and stakeholders.
As the generative AI sector is nascent, efforts have been made by
(4) The availability and mobility of AI talent.
many competition authorities to better understand the
(5) The ability of customers and consumers to switch and multi-
competitive dynamics of generative AI related markets also
home across different cloud or AI foundation model
conducting market studies and issuing reports. 45 As noted in the
providers, potentially reducing the impact of network effects
Introduction, DG Competition published for this purpose on
and making it easier for rival players to penetrate the
market.
Treaty on the Functioning of the European Union to categories of (6) The presence and dissemination of differentiated AI
technology transfer agreements (TTBER); and for vertical agreements,
Commission Regulation (EU) 2022/720 of 10 May 2022 on the
foundation models, for instance specialised ones for specific
application of Article 101(3) of the Treaty on the Functioning of the sectors or functions. This may largely depend on access to
European Union to categories of vertical agreements and concerted specific datasets, and in turn on the regulatory frameworks
practices (VBER). applicable to the licensing of such data, as well as the
45
See, for instance, in the European Union, the French Opinion 24-A-05,
fn. 8; the Hungarian competition authority’s market analysis on the market price of such data.
impact of AI on market competition and consumer behaviour (launched (7) The presence of pro-competitive non-exclusive partnerships
on 4 January 2024) available at https://www.gvh.hu/en/press_room/ between generative AI developers and players with access to
press_releases/press-releases-2024/gvh-launches-market-analysis-on-
important inputs or access to consumers, enabling smaller or
the-impact-of-artificial-intelligence and the Portuguese Issues Paper,
fn. 8. See also, outside of the European Union, the UK CMA’s Initial
46
Report on AI Foundation Models, fn. 36 reviewed by an Update Report, The video recording of the workshop is available at the YouTube
fn. 8; and the US FTC’s market inquiry into the investments and channel of DG Competition.
47
partnerships being formed between AI developers and major cloud Joint Statement on Competition in Generative AI Foundation Models
service providers available at https://www.ftc.gov/news- and AI Products, 23 July 2024.
48
events/news/press-releases/2024/01/ftc-launches-inquiry-generative- See responses to Question 6 of DG COMP's Calls for Contributions on
ai-investments-partnerships. Competition in Virtual Worlds and Generative AI (see above fn. 3).

9
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

more innovative players to establish themselves on the Further, in the European Union, legislation has been adopted
market. enabling free access to text and data mining while protecting
(8) Some well-targeted interoperability standards across copyright holders. The provision contained in Article 3 of the EU’s
different AI foundation models and across different level of Copyright Directive contains an exception to EU copyright law for
the generative AI supply stack, which do not decrease text and data mining for scientific purposes and, while other
incentives to invest but that allow customers of one AI exceptions may be implemented for text and data mining,
foundation model, for instance, to interact upstream with Article 4(3) allows rightsholders to reserve the use of works
different cloud providers and downstream with different where appropriate. 52 Another example is the Data Act, which will
generative AI system deployers or users. 49 potentially facilitate switching, data portability and
interoperability between cloud providers and other data
Public policy activities relating to these factors may therefore processing services at no cost. 53
contribute to ensuring competitive generative AI related markets.
The AI Act itself, while not directly focusing on competition,
For instance, public authorities might engage in initiatives to ensures that all providers of general-purpose AI models operating
promote open-source models and, when appropriate and in the EU comply with copyright and related laws, in order ‘to
necessary, interoperability standards, and to strengthen, ensure a level playing field […] where no provider should be able
wherever possible, the factors that limit or reduce barriers to to gain a competitive advantage in the Union market by applying
entry to these markets. lower copyright standards than those provided in the Union.’ 54

Furthermore, the European Union, represented by the European Finally, as mentioned above, the DMA allows prohibiting ex ante
Commission, is already investing significant resources to make instances of potentially problematic conduct relating to the
public supercomputers accessible in several Member States. The integration of generative AI services into other digital products by
European High Performance Computing Joint Undertaking gatekeepers, as well as monitoring any potentially problematic
(EuroHPC) is a government-industry collaboration aimed at acquisitions on the basis of Article 14 of the DMA. 55
developing and adopting the most innovative and competitive
supercomputing systems in Europe, as well as expanding their Virtual Worlds
access to European users, including SMEs and startups. So far, it
successfully deployed eight supercomputers in different EU Market dynamics and emerging tendencies
Member States, whilst a ninth exascale supercomputer (Jupiter) is The virtual worlds industry features diverse players with distinct
currently under construction in Germany. 50 The accessibility of strategies, who invest in a variety of intertwined technologies and
such public supercomputers enables innovation, by making the services. Stakeholders confirm that from a conceptual point of
pre-training and the fine-tuning of foundation models accessible view, these can be grouped in three layers: enabling technologies,
to a wider set of stakeholders. 51 virtual worlds platforms and specific services. 56

As part of its AI innovation package of January 2024, the Enabling technologies power virtual worlds through a
European Commission launched several additional activities, combination of hardware and software technologies, providing
including financial support to generative AI projects through, for the infrastructure and the input and output devices that can
instance, Horizon Europe and the Digital Europe programme, as support immersive experiences and interactions. Large digital
well as support to education, training, skilling and re-skilling players, such as Meta, Apple, Sony or Microsoft have invested in
activities to strengthen the pool of generative AI talent European hardware technologies such as virtual reality headsets. High-
companies can draw from. Competition authorities may, for
instance, ensure that companies do not agree on illegal no- 52
See Articles 3 and 4(3) of Directive (EU) 2019/790 of the European
poaching or no-hire of AI engineers or other key employees or Parliament and of the Council of 17 April 2019 on copyright and
related rights in the Digital Single Market and amending Directives
promote Governments’ initiatives to limit excessively long or
96/9/EC and 2001/29/EC, OJ L 130, 17 May 2019, pp. 92–125.
unjustified non-compete agreements that may prevent them 53
Articles 23 and ss. of Regulation (EU) 2023/2854 of the European
from moving to other jobs. Policies aimed at fostering education Parliament and of the Council of 13 December 2023 on harmonised
and re-skilling of workers in the AI domain may also be helpful to rules on fair access to and use of data and amending Regulation (EU)
2017/2394 and Directive (EU) 2020/1828 (Data Act), OJ L, 2023/2854,
make AI talent as widely available to companies as possible. 22 December 2023.
54
See recital 106 of Regulation (EU) 2024/1689 of the European
Parliament and of the Council of 13 June 2024 laying down
harmonised rules on artificial intelligence and amending (Artificial
Intelligence Act) OJ L, 2024/1689, 12 July 2024, available at
49
See responses to Question 8 of DG COMP's Calls for Contributions on http://data.europa.eu/eli/reg/2024/1689/oj.
55
Competition in Virtual Worlds and Generative AI (see above fn. 3). See responses to Question 12 of DG COMP's Calls for Contributions on
50
See, for more information, https://digital-strategy.ec.europa.eu/en/ Competition in Virtual Worlds and Generative AI (see above fn. 3).
policies/high-performance-computing-joint-undertaking; https://eurohpc 56
See responses to Questions 2 and 3 of DG COMP's Calls for
-ju.europa.eu/index_en. Contributions on Competition in Virtual Worlds and Generative AI (see
51
See also the French Opinion 24-A-05, fn. 8, p. 52. above fn. 3).

10
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

speed networks, cloud computing, chips, relevant intellectual product will look like in the buyer’s home, as well as (v) education
property, AI and data are also key enabling factors, access to and training in virtual classrooms or “gamified” environments.
which can play a crucial role for vibrant competition in the virtual
worlds sector. If an undertaking is active across multiple layers of the value
chain, vertical integration may give rise to foreclosure concerns.57
A dynamic area within enabling technologies is the so-called Large digital players that control key technologies, such as virtual
“digital twins” technology, where several players, such as Bosch, reality headsets, cloud computing services, chips, or already
Dassault Systèmes, General Electric, IBM, Microsoft, SAP and existing platforms, could be in the position to dictate what the
others are active. Digital twins are virtual models of physical other elements of virtual worlds ecosystems will look like.
objects or systems, such as an engine, factory or even entire
supply chain. They can, for instance, be used for training purposes Examples of vertical integration raised by stakeholders include
but also to predict and prevent points of failure or bottlenecks on Meta, which through its vertical integration, has taken early
the factory floor or within the supply chain. They have reportedly prominence in the consumer-facing virtual worlds space. By
already led to improvements, e.g. in aircraft manufacturing and combining leading hardware (virtual reality headsets), app stores
maintenance and in healthcare for enhancing patient care and and virtual content with data created in its social media
medical training. With regard to digital twins and B2B platforms, Meta is in an advantageous position to drive user
applications in general, stakeholders see significant opportunities adoption and engagement. Apple is another prominent player
for the European industry, including SMEs, and highlight that an following a strategy of vertical expansion. Its virtual and
overall competitive environment and, in particular, access to live augmented reality headset utilises its own chip and operating
data and computing power, are essential to realise the full system, integrating its virtual worlds offering into its existing
potential of this technology. hardware and applications ecosystem. Nvidia is leading
specialised chip development for industrial virtual reality
Virtual worlds platforms on which virtual worlds are developed applications, building on its expertise in 3D gaming, and also
and accessed are currently taking shape and it is not yet possible offers its Omniverse solution to build digital twins. Similarly,
to define them clearly. Virtual worlds platforms can be Microsoft provides key underlying technologies and hardware for
distinguished between platforms in the B2C sector, such as virtual worlds, including headsets, Azure cloud computing services
gaming or virtual social spaces like Meta Horizon, and the B2B and digital twins, and gained ground in AI innovation.
sector, such as work environment platforms or industrial virtual
worlds like Nvidia’s Omniverse. There are also virtual worlds At the same time, vertical integration can also be pro-
maintained by public actors such as cities or public authorities. competitive, creating greater efficiencies, better quality, or lower
costs for consumers. That said, the strong positions of a number
Virtual worlds platforms have a number of specificities that set of large digital players in key enabling technologies and
them apart from other platforms. They stand out due to their platforms may warrant the attention of competition law
immersive and dynamic nature, compared to two-dimensional, enforcement so as to ensure that potential entrants are not
often static platforms, such as an online marketplace or a news excluded through either foreclosing access to key inputs for their
aggregator. They are persistent and continuous, allowing users to own virtual worlds or acquisitions.
enter and exit, similarly to a permanent location in the physical
world. They typically offer multiple ways of perception, sensing There are diverging views on whether virtual worlds should adopt
and manipulation, compared to touchscreens or displays. Finally, open standards and different platforms seem to explore different
virtual worlds platforms offer unique development tools and they options in that regard. Many stakeholders consider that without
cater to diverse applications. seamless interoperability, virtual identities (avatars), virtually
purchased goods or digital twin models cannot be moved
Some stakeholders consider that operating systems and app between competing platforms. However, proponents of
stores, some of which are already subject to DMA obligations, proprietary ecosystems emphasise that private standards and
should also be regarded as separate layers of the value chain closed ecosystems could also have benefits by offering better
(rather than part of the platform layer). security and more tailored experiences. Striking the right balance
is considered crucial, and the potential impact on innovation
On the third layer – services – companies specialise in offering should be carefully assessed.
specific and tailored services to businesses and customers on top
of virtual worlds platforms and within the virtual worlds
environment. The potential for such services is vast. Common
examples for such services currently include (i) virtual work in
immersive, remote workspaces, (ii) medical services, allowing
high precision, robot-assisted or remote surgeries, (iii) a variety of
57
industrial services based on digital twins, (iv) e-commerce of See responses to Questions 2, 3 and 8 of DG COMP's Calls for
Contributions on Competition in Virtual Worlds and Generative AI (see
virtual goods or the sale of physical goods, such as furniture and above fn. 3).
other design products, in virtual spaces that simulate what the

11
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

Potential barriers to entry and expansion Several contributors, however, point at a number of already
Stakeholders raise that strong scale effects can increase entry
58 existing risks.
barriers in virtual worlds markets. The costs of the necessary
Rising concentration and cross-ownership in the video games
investment and innovation are very high in the sector. Large
sector could lead to the risk of existing market power in video
digital players may have more financial means should they
game ecosystems potentially translating into market power in
decide to invest heavily in the underlying infrastructure,
virtual worlds. Similar risks could arise in relation to vertically
technology, content and talent. In contrast, European start-ups
integrated players, where market power on critical inputs – being
and SMEs often lack access to funding, which makes it difficult
technology and infrastructure such as headsets or cloud, but also
for them to scale up and match the offer of the established tech
platforms such as app stores or social media – could be
companies.
transferred into virtual worlds markets.
Network effects, especially of platform-based services that
Consolidation of market power by a few dominant platforms
connect a large number of users with service providers (e.g.
could lead to gatekeeping behaviours, which could restrict entry
multiplayer gaming or social media platforms that are
and competition. Dominant platforms could also engage in
transitioning into immersive virtual worlds), play a crucial role in
exclusionary practices, limiting consumer choice, thereby stifling
driving user adoption and engagement within virtual worlds and
innovation. The competitive advantage afforded by the use of
they can make it difficult for new entrants to gain foothold: as
data for personalised virtual experiences could further entrench
more users join a platform, its value increases, attracting even
the market position of established players, especially if they limit
more users. This draws service providers to the largest platforms,
data portability or their competitors’ access to such data sets. In
while small platforms may struggle to attract enough services,
this context, several contributors raise the need to extend the
especially if proprietary development toolkits and standards
scope of the DMA and reduce its thresholds.
make deployment for multiple platforms too costly. While this
phenomenon may be more prominent in B2C platforms, its
Contributors also flag the risk of aggressive acquisition
relevance should not be dismissed for B2B platforms either.
strategies. These can allow financially sound, leading tech
companies to absorb innovative startups, diminishing the
Customer lock-in can increase switching costs, if users have
vibrancy and diversity of the ecosystem and eliminating potential
already invested money and time into a platform. Potential
challengers or innovative technologies that could be disruptive to
restrictions on switching, e.g. because of limited interoperability,
their existing models. Beyond such “killer acquisitions”, some
cost of portability, exclusivity agreements or other factors that
contributors also point out a risk of “kill zones”: in certain
may result in customers’ losing assets, such as digital twin
segments, the presence of some ecosystems could discourage
models, can deter them from trying out new platforms.
other undertakings from investing or competing. In these zones,
Established digital platforms collect vast amounts of user data, startups and competitors are less likely to thrive due to the
which they use to improve services and target advertising. 59 perceived risks or lack of opportunities, which can stifle
Virtual worlds platforms will likely not be different, particularly investment and innovation and ultimately chill competition.
because the personalisation of the immersive experience will
Potential other anticompetitive practices, such as exclusivity
heavily rely on data relating to user behaviour and interactions.
agreements (e.g. between platforms and IP rightsholders), unfair
Without fair access to this data, new entrants could be
trading conditions (e.g. major platforms imposing their own
disadvantaged. Moreover, potential data portability restrictions
standards), self-preferencing, tying and bundling, refusal to
could contribute to increasing switching costs and lock-in effects,
supply interoperability information, predatory pricing and margin
strengthening existing positions of platforms and foreclosure of
squeeze may also arise in virtual worlds markets.
rivals by limiting their capacity to attract users. In that context,
stakeholders point out that user ownership of their own data is
critical. Conclusions and next steps
Generative AI and virtual worlds technologies will considerably
Potential competition concerns impact and shape numerous markets in the coming years. The
Some stakeholders argue in the call for contributions that it deployment of these technologies will undoubtedly bring
would be too early to make assumptions about possible significant positive changes, as well as innovation, new business
competition concerns in virtual worlds markets and that the EU models and new ways of doing things.
already has the right tools to address any issues that may arise.
They may also, however, give rise to competition concerns that
58
See responses to Questions 1 and 5 of DG COMP's Calls for may threaten the well-functioning of markets and negatively
Contributions on Competition in Virtual Worlds and Generative AI (see affect innovation, choice and quality. These concerns may include
above fn. 3).
59 exclusionary practices or other forms of foreclosure by dominant
See responses to Questions 1, 7 and 8 of DG COMP's Calls for
Contributions on Competition in Virtual Worlds and Generative AI (see players, such as exclusivity agreements, imposition of unfair
above fn. 3). trading conditions, self-preferencing, tying and bundling, refusal

12
 Competition in Generative AI and Virtual Worlds | Competition Policy Brief No 3/2024

to supply, margin squeeze or predatory pricing, reducing choice As regards virtual worlds, DG Competition is closely monitoring
and innovation for consumers. how these nascent markets evolve, as virtual worlds are expected
to transform traditional markets (such as manufacturing,
They may also consist of acquisition strategies aimed at education, software, gaming, media, etc.). This will also allow
eliminating nascent competitors (killer acquisitions) or absorbing DG Competition to apply existing knowledge of these markets to
most of their key employees (acqui-hires) and critical know-how. their virtual extensions, building also on its experience on the
Incentives to invest and innovate may be further chilled by the evolution of digital markets.
presence of ecosystems, network effects and tipping dynamics in
the so-called “kill zones” described above. 60 DG Competition currently has also preliminary investigations
ongoing into markets that are crucial for the development of
A heightened risk of horizontal collusion or exchange of generative AI and virtual worlds, like the markets for cloud or for
commercially sensitive information may also emerge, as well as different types of specialised chips (such as GPUs, TPUs and
wage-fixing and no-poach agreements between employers vis-à- other AI accelerators).
vis key employees.
As a follow up to the joint statement published with the US DoJ
Given the magnitude and extent of the impacts of these and FTC and the UK CMA, 61 and given the global nature of the
technologies and the risks they might entail, DG Competition is business operations of AI players, the European Commission is
actively monitoring the AI and virtual worlds sectors to ensure also committed to strengthening the coordination at the level of
that competition is not negatively affected. As mentioned above, the ECN, ICN and OECD in relation to enforcement in generative
DG Competition is already analysing the investments and AI related markets. It is also fully engaged in G7 discussions on
partnerships between large digital players and generative AI these issues, advancing the global dialogue on goals including
developers, as well as agreements for the pre-installation of the reduction of barriers to entry in AI development, sustaining
generative AI models on the devices to ensure they do not raise diversity of AI business models, and fostering innovation.
anticompetitive concerns.
DG Competition will use all tools at its disposal to address
DG Competition is also monitoring partnerships between large potential concerns in the generative AI and virtual worlds sectors,
digital players and AI developers from the merger control including antitrust, merger control and the DMA, to ensure that
viewpoint. these sectors remain competitive, contestable and fair.

61
60
See above the section on Virtual Worlds – Potential competition See above the section on Competition advocacy and other pro-
concerns. competitive initiatives in generative AI related markets and fn. 47.

13