Competition in artificial intelligence infrastructure

4. Potential competition policy responses in AI infrastructure

The economic and social impacts of AI are forecast to be widespread and significant, but they rely on AI infrastructure to also grow at pace. As highlighted in section 3, there are several features of the AI infrastructure value chain which make it susceptible to competition issues. It is therefore important that competition authorities consider the best responses to potential competition issues in these markets. This section discusses the potential application of competition tools to the sector, as well as the extent to which there has been activity so far.

4.1. Competition enforcement

Having effective competition enforcement is a key tool in ensuring that firm conduct is kept in check and competition is based on the merits of enterprise. The US Department of Justice (DOJ) Assistant Attorney General Gail Slater recently highlighted the importance of enforcement in markets such as AI infrastructure where technology is rapidly developing:

Antitrust enforcement focuses on creating a level playing field for businesses big and small and ensuring that market incumbents do not unfairly hinder newcomers and startups. This is always important, but it is crucial where the technology is still developing rapidly (US DOJ, 2025[127]).

Recently several authorities have launched investigations into the chip designers developing the latest AI chips.1 The supply chain at the chip design level has seen several significant global enforcement cases in the past prior to the development of AI. However, to date there has been no completed enforcement actions relating directly to the AI infrastructure supply chain.

This section highlights some of the potential areas that authorities may need to pursue potential antitrust cases, as well as highlighting some of the past cases as examples of risks which may arise again. We first discuss anti-competitive agreements before considering the risks of unilateral conduct.

4.1.1. Anticompetitive agreements

As highlighted in section 3, the AI infrastructure supply chain is characterised by an increasing number of agreements, often including investments in equity which fall below merger thresholds. These agreements have created an interrelated web of firms with crossholdings and relationships which could have the potential to weaken competition in the sector.

Vertical or conglomerate agreements are not per se illegal and in most jurisdictions they are subject to an economic effects analysis. However, given the importance of the sector, authorities should pay close attention to such agreements. So far, it appears that the agreements have received more attention through merger regimes (discussed under merger control below). Although, the US FTC has conducted an initial investigation into some of the arrangements between cloud providers and generative AI companies. A summary of the findings published in the US FTC staff report are included in Box 4.

Box 4. US FTC review of partnership agreements

The US Federal Trade Commission in January 2025 published staff findings on a study into partnerships involving generative AI companies and cloud providers. They sought to understand whether these relationships were being used to circumvent merger review or create anticompetitive advantages. The report highlighted three areas to watch regarding the potential implications of AI partnerships:

• The partnerships could affect access to certain inputs such as computing resources and engineering talent.

• The partnerships could increase contractual and technical switching costs for AI developer partners.

• The partnerships provide cloud service providers access to sensitive technical and business information unavailable to others.

Sources: US FTC (2025), Partnerships Between Cloud Service Providers and Ai Developers - FTC staff Report on AI Partnerships and Investments 6 (b) Study, https://www.ftc.gov/system/files/ftc_gov/pdf/p246201_aipartnerships6breport_redacted_0.pdf.

The agreements examined by the US FTC and other authorities primarily relate to agreements between generative AI providers and cloud providers, but such agreements are becoming increasingly prevalent at all levels of the AI infrastructure stack. Intervention may be warranted when agreements contain provisions which could lead to market foreclosure (such as exclusivity), for example preventing rivals from accessing key inputs (e.g. compute or reaching customers). This is especially critical in sectors like AI infrastructure, where access to advanced chips, can determine market viability.

Beyond vertical agreements, recently there have been arrangements between firms at the same level of the supply chain. Such horizontal arrangements and collaboration can raise the risk of information sharing or broader collusion (OECD, 2018[128]).

More generally the risk of collision may be relatively low at the leading edge of the AI infrastructure supply chain, given it is mostly operating with fast moving technology and high levels of innovation. At many levels of the supply chain has one firm with a leading market share. In markets with such asymmetric shares, there is typically lower risk of collusion (Motta, 2004[129]), especially when the market leaders can often make legal monopoly profits from their innovations. That said, there are other parts of the supply chain where technologies have matured, and the products may have become commoditised with oligopoly-like competition between a small number of players. In these areas the risk of collusion is likely to be higher (Asmat, 2019[130]). For example, Box 5 below highlights enforcement action from the 2000s relating to a computer memory chip cartel.

Box 5. The DRAM Cartel

Dynamic Random Access Memory (DRAM) is a type of semiconductor memory chip used in the primary memory systems of computers and other digital devices. As the DRAM market matured and commoditised, chips became interchangeable and industry standards were widely followed. This commoditisation in concentrated markets with high information sharing increased the risk of collusion.

The DRAM cartel was a significant international price-fixing cartel involving several manufacturers of DRAM chips between 1998 and 2002. The cartel included major DRAM producers including Samsung, Hynix, Infineon, Elpida, Micron Technology, NEC, Hitachi, Toshiba, Mitsubishi Electric, and Nanya Technology. These companies co-ordinated prices and restricted competition, artificially inflating DRAM prices sold to PC and server manufacturers.

A number of competition authorities took action and issued large sanctions, including:

• The US DOJ launched investigations in 2002 under the Sherman Antitrust Act. Between 2003 and 2006, several executives and companies pleaded guilty. For example: Infineon was fined USD 160 million in 2004. Hynix paid USD 185 million in 2005. Samsung paid USD 300 million in 2005. Executives at Samsung were also sentenced to prison. 

• In 2010, the European Commission fined nine companies a total of EUR 331 million for their role in the cartel. Micron Technology received immunity for whistleblowing.

Sources: European Union (2011[131]) Final report of the Hearing Officer - COMP/38.511 - DRAMS, https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:52011XX0621(02); United States District Court (2005[132]) United States of America v Samsung Electronics Company and Samsung Semicoductor Inc, https://www.justice.gov/atr/case-document/file/509296/dl; United States District Court (2005[133]) Unites States of America v. Infineon Technologies AG, https://www.justice.gov/atr/case-document/file/499551/dl; (2005[134]) United States of America v Hynic Semiconductor Inc, https://www.justice.gov/atr/case/us-v-hynix-semiconductor-inc; United States District Court (2006[135]) United States of America v Sun woo Lee, https://www.justice.gov/atr/case-document/file/501201/dl.

Lastly, the rapid growth in AI infrastructure may also have implications for the risk of collusion upstream in large-scale projects involving data centres, fibre networks, and energy-intensive compute facilities. These projects often require tenders for the largescale procurement of construction, including specialised engineering expertise. The high capital investment, and co-ordination across a limited number of capable firms, create conditions which increase the risk for bid rigging activity. The complexity and technical specificity of AI infrastructure tenders could obscure pricing benchmarks and make it easier for firms to align bids or rotate winners.2 Additionally, the urgency to deploy infrastructure may lead to compressed timelines and reduced scrutiny in procurement processes, weakening safeguards against cartel formation. As governments and tech giants increasingly rely on public-private partnerships to build AI infrastructure, competition authorities should be alert to the structural vulnerabilities in upstream markets and continue advocating for enhanced procurement design and detection to actively deter collusion such as bid rigging (OECD, 2025[136]).

4.1.2. Anticompetitive unilateral conduct

As highlighted in section 3 there are several highly concentrated parts of the AI infrastructure supply chain. Firms with very high market share and positions protected by barriers to entry will likely have a high degree of market power (OECD, 2022[137]). This raises the risk of firms potentially abusing their market power.3

The starting point for authorities in assessing whether conduct is an abuse will be to seek to prove dominance (or monopoly power). Dominance is often defined as the ability to act to a substantial degree independently of competitors, customers and suppliers (i.e. the ability to exercise market power) (OECD, 2021[138]). This can be expressed in metrics such as the ability to earn high profits on a sustained basis which are above costs (including a reasonable return on capital). Some jurisdictions will have market share-based thresholds which allow competition authorities to establish a firm’s dominance beyond a certain share of the market (OECD, 2024[139]). 4

In practical terms when assessing dominance, whether structural presumptions apply or not, the approach typically starts with defining the relevant market. In several layers of the AI infrastructure value chain, meeting the threshold of dominance would seem likely as several providers along the value chain appear to have near monopoly market shares on certain technologies and are generating significant profit margins (Narechania and Sitaraman, 2023[45]). One key aspect competition authorities may need to consider is the dynamic nature of the markets, for example in conducting market share analysis by focusing on the share of forward orders rather than the most recent years data an analysis.

One of the features of AI supply chains highlighted earlier in the paper is that there are high levels of both vertical integration and concentration (and likely market power) at different levels of the supply chain. Such conditions create the risk that firms will use their position in one market to distort competition in other areas in their favour. This potential for exclusionary behaviour could take different forms including outright refusing to supply, tying/bundling, use of rebates or other pricing strategies

There have been numerous examples in previous rounds of computing developments of tying and bundling being used in a potentially anticompetitive manner. These cases have not always been successful, demonstrating the challenges for authorities in bringing cases. For example, IBM was accused of tying memory components to its CPU in the 1970s (although the court sided with IBM on the basis that that the integration was technologically superior) (Rowles, 2000[140]). There have however been successful cases, for example, the European Commission fined Microsoft EUR 497 million for bundling Windows Media Player with Windows OS (European Union, 2007[141]), foreclosing competition in media players. Google was fined EUR 4.34 billion for tying its search engine and Chrome browser to the Android OS, leveraging dominance in mobile OS to gain advantage in search and browser markets (European Union, 2018[142]).

In relation to AI infrastructure such bundling or tying could emerge as a risk at different levels for example, most AI models operate off cloud computing resources where compute access is rented from third parties and increasingly specialised for AI modelling and training. This advanced cloud computing is primarily offered by three hyperscalers, Google, Microsoft and Amazon (OECD, 2025[5]). Competition concerns may emerge from vertical relationships where cloud service providers are also involved in developing and deploying AI models and applications. If a company has significant market power in the upstream cloud services sector, by bundling its AI model downstream together with this infrastructure, it could make it in the future hard for model developers to compete independently. Another risk is that AI chips are bundled with other components. Furthermore, risk can result for the AI infrastructure supply chain is in the supply of complementary hardware products. As firms begin to offer multiple complementary hardware products in the supply of the data centre technology, there is a risk that if a supplier has a dominant position in one where there are high barriers to entry, it may use that position to gain greater market share in other complementary hardware products.5

Authorities should also be alert to the risk of pricing strategies, such as the use of conditional rebates, which may seek to encourage exclusivity through the incentives delivered (OECD, 2016[143]). Such cases have occurred in previous rounds of chip development for example, the EU has previously acted against Intel regarding the structure of their rebates, although the finding was later annulled on appeal (Court of Justice of the European Union, 2024[144]). A summary of the case is discussed in Box 6 highlighting how both pricing behaviour in relation to chips can be viewed as potentially exclusionary, and also that such behaviour may create efficiencies, requiring thorough economic analysis.

Box 6. Intel rebates case

Prior to the latest wave of AI computing, CPUs were the core technology of modern computing. Intel was the dominant manufacturer with AMD a smaller rival.

In 2009, the European Commission fined Intel EUR 1.06 billion under Article 102 TFEU for abusing its dominant position in the x86 CPU market. The Commission alleged that Intel granted loyalty rebates to computer manufacturers (e.g. Dell, HP, Lenovo) to induce exclusivity and foreclose competition, particularly from AMD.

The case was subject to multiple appeals with the CJEU ultimately annulling the case in 2024. The decision highlighted the need for thorough economic analysis of any such conduct and in particular to respond to any economic arguments put forward by the defendant sufficiently.

The core theory relating to such rebate cases is that there is typically a smaller contestable share of the market that competitors can compete on, and the conditional rebates can mean that the effective price competitors would need to compensate customers with is so low as to make effective entry/growth impossible.

Notes: The European Commission also found against Intel for abusing its dominant position by paying providers for delaying products. These findings of naked restrictions were note overturned on appeal.

Sources: Judgement of the Court (fifth Chamber) 2024, Commission vs Intel Corporation, https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:62022CJ0240; OECD (2016[143]), Fidelity Rebates, https://doi.org/10.1787/fd9e16da-en.

As highlighted above, previous technological waves have seen complaints and investigations into exclusionary behaviour, including tying and rebates. While such cases are no doubt complex and have in the past faced challenges in court. This does not mean that authorities should not continue assessing the impacts of potentially exclusionary pricing schemes, but authorities must ensure they have clear theories of harm. The challenge of bringing cases in such rapidly developing markets is the time they take, and therefore enforcement must be complemented with effective advocacy where necessary.

4.2. Merger control

The AI infrastructure supply chain is experiencing rapid technological innovation and unprecedented growth. In this environment, mergers & acquisitions (M&A) and strategic partnerships have become important tools for industry participants to secure and effectively manage their supply chains. This can drive efficiency and has benefits for startups who can secure investment for growth and R&D. However, they may also raise some significant competition concerns (OECD, 2020[145]). Effective merger control is therefore important to ensure anticompetitive arrangements are kept in check.6

As highlighted in section 3, the sector has high concentration and increasing levels of vertical integration, investments and partnerships. This creates a need for competition authorities to carefully scrutinise deals even where firms are not directly competing. There have already been several such merger investigations in the supply chain. Box 7 highlights two recent merger investigations into acquisitions at different levels of the supply chain where authorities considered vertical and conglomerate theories of harm.

Box 7. Recent examples of non-horizontal merger investigations in AI infrastructure

Mellanox/Nvidia merger

The Nvidia–Mellanox merger, valued at USD 6.9 billion and completed in 2020, was scrutinised by several competition authorities, focusing on conglomerate theories of harm. The transaction involved Nvidia, the leading provider of GPUs used in datacentres and Mellanox, a leading provider of network interconnects in datacentres.

The European Commission cleared the deal unconditionally in December 2019, finding no horizontal overlaps and limited vertical concerns. However, it examined conglomerate theories of harm ultimately concluding that in practice even if Nvidia had the incentive and ability to leverage Mellanox’s products, it would only impact a small proportion of the GPU market and therefore not materially impact the competition from rivals.

The Chinese competition authority (SAMR) conducted a year-long Phase II review and ultimately granted conditional approval in April 2020. The SAMR identified conglomerate effects in neighbouring markets where the merged entity would hold dominant shares. To mitigate risks of tying, bundling, and foreclosure, SAMR imposed behavioural remedies, including commitments to supply on FRAND terms, maintain interoperability, and protect third-party confidential information. SAMR also required Nvidia to preserve organisational separation between GPU and interconnect teams and to maintain open‑source commitments. Recently the SAMR has also reportedly opened an antimonopoly case in relation to potential breaches of the commitments agreed in 2020.

ARM/Nvidia merger

The proposed ARM–Nvidia merger, announced in 2020, was a USD 40 billion deal in which Nvidia sought to acquire ARM from SoftBank. ARM is a supplier of chip architecture used widely across the tech industry, including by Nvidia’s competitors. The deal raised significant competition concerns globally, particularly around the risk that Nvidia could restrict access to ARM’s technology or distort competition in downstream markets like data centres.

Concerns centred on the potential harm to Nvidia's rivals through foreclosure strategies, such as limiting access to ARM's CPU intellectual property and hindering interoperability between related products. The theory was that this may ultimately benefit Nvidia's downstream activities and increase its profits. Of particular relevance to AI infrastructure, vertical and conglomerate effects were identified as a risk which, could restrict access to CPUs, network interface controllers and GPUs, impacting data transfer efficiency and server performance.

After extensive investigations and concerns raised from multiple competition authorities including the CMA, US FTC and the EU Commission, Nvidia and SoftBank abandoned the deal in February 2022.

Sources: European Commission (2019), Case M.9424 – NVIDIA / MELLANOX, https://ec.europa.eu/competition/mergers/cases/decisions/m9424_778_3.pdf; Garrod et al. (2020) Nvidia/Mellanox: China’s Close Scrutiny of Semiconductor Deals Continues, https://www.akingump.com/en/insights/alerts/nvidiamellanox-chinas-close-scrutiny-of-semiconductor-deals-continues; Perrone, H. (2025), “Chips in on a merger: The Arm-Nvidia case”, International Journal of Industrial Organization, Vol. 98, p. 103130, https://doi.org/10.1016/j.ijindorg.2024.103130.

As well as conglomerate and vertical concerns, acquisitions of nascent competitors and potential killer acquisitions could also present challenges for merger control in AI infrastructure. Killer acquisitions refer to the strategic purchase of nascent or potentially competitive firms with the intent to neutralise future threats, often by discontinuing their innovation or integrating them into a broader ecosystem. As highlighted in section 3, AI infrastructure markets rely on a high degree of innovation, with leading edge innovations often gaining dominant positions in a market. This can lead to incumbents seeking to acquire small startups as a defensive measure, which can undermine dynamic competition and risk incumbents entrenching their positions in key layers of the AI stack. On the other hand, given the potential need for large amounts of capital to function in the AI sector, and potential efficiencies from transactions, authorities need to carefully consider what the appropriate counterfactual would be (OECD, 2020[145]). Given the highly dynamic nature of any such assessment authorities may need to rely more on qualitative evidence, including assessing internal documents of the acquirer to understand the extent of any perceived threat. In addition, analysis on the valuation of the transaction may also provide additional information (OECD, 2020[145]).

While this report focuses on AI infrastructure markets which is primarily about physical hardware, such infrastructure markets also rely on the knowledge and skills of key staff. Recently there has also been an increase in moves to acquire key staff such as researchers, engineers and leadership in AI infrastructure markets.7 These arrangements where companies may seek to hire teams of key personnel can significantly reshape competitive dynamics. While usually hiring and firing is not the focus of merger frameworks, competition authorities have begun to assess such practices (Federle and de Amorin, 2024[146]), especially when the acquisition of staff leads to loss of competition or potential competition.

The CMA, and Bundeskartellmt for example, have previously examined the agreement between Microsoft, a cloud service provider, and Inflection, an AI company and model develop. The CMA found that hiring of key personnel and licensing IP constated a merger, however the CMA found that the transaction was unlikely to result in a realistic prospect that there would be a substantial lessening of competition (Competition and Markets Authority, 2025[147]). The Bundeskartellamt found that the merger did not meet the national thresholds for merger reviews (Bundeskartellamt, 2024[148]).

Such acquire-hires and other transactions involving innovative startups have exposed a potential gap in existing frameworks, where strategic investments, joint ventures, or exclusive supply arrangements may escape merger control scrutiny. Former European Commission Executive Vice President Margrethe Vestager noted that some competitively significant tech acquisitions fall below EU Merger Regulation thresholds yet still warrant scrutiny (European Commission, 2024[149]).8 Several authorities therefore have begun to introduce call in powers, giving them discretion to examine mergers even when not meeting the mandatory thresholds (Bary, 2025[150]). However, the use of such powers should be balanced carefully against the potential legal uncertainty they can create if not applied consistently or utilised excessively.

4.3. Market studies

Market studies are a flexible tool for competition authorities to analyse whether there are competition problems in a sector outside of enforcement investigations or merger control (OECD, 2018[151]). Using these tools to explore the AI infrastructure supply chain has the benefit of increasing authority staff knowledge. In addition, proactive efforts during the early stages of development can uncover issues which may be resolved before market dynamics become intractable, a common critique of competition authority responses to digital platforms in the past two decades.

There have been several market studies in relation to AI and the cloud infrastructure market including in France, Japan, Korea, the Netherlands, Denmark, the US and the UK. These market studies so far have focused on the cloud segment of the market and have found issues including (OECD, 2025[5]):

• switching barriers including the use of egress fees (exit charges)

• restrictive licenses which impose markups for using key software on rivals’ platforms

• issues with cloud credits potentially making prices so low smaller rivals can’t compete

• issues with bundling and tying.

The competition issues in the cloud were discussed in more detail in the recent OECD policy paper “Competition in the Provision of Cloud Computing Services” (OECD, 2025[5]) and is not repeated in detail again here.

So far, no competition authority appears to have launched a dedicated market study solely on semiconductors for AI or the supply chain for compute. Nonetheless, the topic is increasingly embedded in broader inquiries into cloud infrastructure, AI partnerships, and digital ecosystems.

France has however an ongoing study exploring access to energy by AI players and there have also been studies undertaken into generative AI which have considered the potential competition issues upstream. These studies included reports from France (Autorité de la concurrence, 2024[111]), Japan (Japan Fair Trade Commission, 2025[152]), Korea (Korean Fair Trade Commission, 2024[153]), Portugal (Autoridade de Concorrencia, 2023[154]), Canada (Competition Bureau Canada, 2025[155]) and the UK (Competition and Markets Authority, 2024[102]).

Some of the findings from these studies include:

• Promoting competition in Generative AI markets is intrinsically tied to promoting competition in these upstream markets (Autoridade de Concorrencia, 2023[154]).

• Potential concerns include the risk of abuse by IT component providers which have dominant positions, the potential for lock-in by cloud platforms and a lack of transparency in relation to investments and partnerships (Autorité de la concurrence, 2024[111]).

• A risk that if large technology companies control key AI development resources, this control could lead to exclusionary practices that limit competition (Competition Bureau Canada, 2025[155]).

• Access to compute remains key to developing and deploying AI modelling, especially for training most of the cutting-edge models. UK stakeholders reported a limited supply of AI accelerator chips and the potential importance of new, smaller AI models which may have implications for compute access needs (Competition and Markets Authority, 2024[102]).

• That there may be different competitive dynamics in AI chips for the training and inference phases of AI. In the training phase NVidias GPUs continue to hold a significant share globally due to factors like development environments and the CUDA software ecosystem. In the training phase Nvidia is likely to maintain its status as market leader, but there may be more competition longer-term at the inference phase. However, there are large switching costs from switching ecosystems and that this will may switching of GPUs even if there are comparable competitor products (Japan Fair Trade Commission, 2025[152]).

• That there are significant economies of scale and scope, potential network effects and first mover advantages in the provision of cloud services and GPU provision. These can pose significant barriers to entry Korea (Korean Fair Trade Commission, 2024[153]).

Given the importance and global significance of the supply chain, market studies could be a valuable tool to ensure competition agencies are well informed and identify potential anticompetitive conduct which might warrant investigation. Such studies can also provide the basis for recommendations to governments for ex ante interventions. While in many jurisdictions market studies only provide the ability to make recommendations, several jurisdictions also have market investigation powers which provide the ability to intervene in markets and address competition issues, without needing to find breaches of the law (OECD, 2024[4]). A lighter touch option could be to gather information by monitoring developments. This could take different forms, from simply devoting time to tracking developments, and engaging with market operators (OECD, 2024[4]). Monitoring and information gathering could be a relatively light-touch first step in understanding whether the sector requires further scrutiny. Monitoring could be complemented with access to expertise, which could be provided by specialist staff (OECD, 2024[4]).

4.4. Advocacy and co-operation

AI infrastructure is both highly dynamic and attracts a great deal government and business attention. Advocacy can play an important role in ensuring markets are shaped in as pro-competitive a manner as possible. Such advocacy may cover several areas such as engagement with market participants as well as with governments.

4.4.1. Advocacy with market participants

One disadvantage of traditional antitrust enforcement tools is that they can be slow to achieve results (von Thun and Hanley, 2024[156]). In fast-evolving AI infrastructure markets, competition advocacy could play an important role in bridging the gap between enforcement and market realities before any anti-competitive conduct risk crystalises.

Effective advocacy can help build trust and awareness among market participants, encouraging whistleblowers, complainants and leniency applicants to come forward (OECD, 2023[157]). In AI infrastructure, where abuse may be subtle or opaque to authorities due to the highly technical nature of the market, insiders will often be the best source of actionable intelligence. By engaging with developers, startups, and infrastructure providers, authorities can better understand emerging risks and competitive bottlenecks, such as discriminatory access to compute or exclusionary bundling.

Finally, advocacy can have a deterrent effect. Public statements, guidance documents and soft law instruments can signal enforcement priorities and shape firm behaviour. In AI compute, where firms may be tempted to leverage infrastructure control for exclusionary purposes, visible advocacy can discourage such conduct before it occurs. However, advocacy alone is insufficient and to ensure deterrence agencies will need to effectively enforce against anticompetitive conduct if it is occurring.

4.4.2. Advocating with other parts of government

As discussed in section 2, governments and authorities will potentially be looking to implement laws and regulations affecting AI infrastructure through several policy angles. For example, AI infrastructure’s environmental impacts through energy use and water consumption.

Competition authorities can play a role ensuring policy interventions are proportionate. Very rigid or prescriptive rules, especially those that favour large, vertically integrated firms can raise barriers for smaller players and new entrants in the AI compute ecosystem. Advocacy can help shape regulation that is risk-based, proportionate, and innovation-friendly, ensuring that other policy goals and competition are pursued in tandem. The OECD Competition Assessment Toolkit (OECD, 2019[158]) provides a framework competition agencies and governments can use to review proposed laws and regulations and promote more competition in their economies, leading to lower prices, greater choice and higher quality of goods and services.

As discussed in detail in section 2, various jurisdictions have passed laws to spur domestic manufacturing of chips or imposed export controls. Competition authorities also have a role to play in helping to ensure that state decisions on investments and trade policy are as pro-competitive as possible, seeking to integrate competitive neutrality in decisions where possible (OECD, 2024[159]). Authorities can also advocate to ensure governments making such large procurement and investment decisions are fully prepared for managing bid processes free from collusion (OECD, 2025[136]).

4.4.3. Co-operation

Almost all of the AI infrastructure chain is made up of global markets. The cross-border nature of the supply chain means that enforcement actions, merger reviews and policy interventions in one jurisdiction can have global implications. Accordingly, competition authorities around the world are recognising the need for closer international co‑operation. For example, the EU, UK CMA, US DOJ and US FTC recently issued a joint statement on AI models and foundations. It noted risks to competition relevant to AI infrastructure, including the concentration of key inputs such as specialised chips and substantial compute. It also included principles for protecting competition in the AI ecosystem including fair dealing, interoperability and choice (EU, CMA, US FTC & US DOJ, 2024[160]). Similarly, the G7 issued a Digital Competition Communiqué which also highlighted the potential for competitive bottlenecks in availability and access to key inputs including compute infrastructure and specialised chips (G7 Italia, 2024[161]). It highlighted principles such as fair access to key inputs like AI chips and called for vigorous antitrust enforcement and enhanced cooperation (G7 Italia, 2024[161]).

As AI infrastructure markets evolve, deeper co-operation will be essential to ensure that merger remedies, abuse investigations, and regulatory interventions are effective and consistent across jurisdictions. Divergent approaches can raise compliance costs for business, stifle innovation and exacerbate the risk of competition issues persisting in some regions.

In addition to working to minimise unnecessary divergence, the consistency of the markets and complexity of the technologies provides opportunities for authorities to share their knowledge, analysis and expertise (OECD, 2024[162]).

Given the interdependencies between AI infrastructure and other regulated sectors, such as energy, telecommunications, and data governance, competition authorities may also need to increasingly co‑ordinate with other domestic agencies. The development and operation of data centres, for example, relies heavily on access to stable and scalable energy supplies, while high-performance compute often depends on advanced connectivity and bandwidth provision regulated by telecommunication authorities. As AI infrastructure scales, issues such as grid access, spectrum allocation, and environmental impacts may intersect with competition concerns, particularly where infrastructure bottlenecks or preferential access arrangements arise. Co‑ordinated oversight can help ensure that enforcement and policy interventions are coherent, avoid regulatory gaps and support a level playing field across sectors.

4.5. Potential regulation

As highlighted in section 3, the AI infrastructure supply chain is complex and dynamic, which can make enforcement challenging. In particular, abuse of dominance cases can be complex and time‑consuming (OECD, 2021[163]). In fast-moving sectors like AI infrastructure which is evolving rapidly, such delays risk rendering enforcement ineffective, underscoring for some the potential need for more agile and anticipatory regulatory tools (Narechania and Sitaraman, 2023[45]).

In recent years several jurisdictions have moved to implement ex-ante regulatory regimes in digital markets.9 Each jurisdiction’s legislation is drafted differently but given these regimes were typically designed in response to digital platforms rather than physical infrastructure or hardware. Therefore, many of the existing provisions will not cover AI infrastructure (beyond the potential partial coverage of cloud platforms) (OECD, 2025[5]).

There have already been calls from some for similar ex ante regulatory interventions in AI infrastructure. It has been suggested, for example, that competition could be improved by introducing access obligations on dominant infrastructure providers (such as chip producers) (Narechania and Sitaraman, 2023[45]). The idea being that such provisions would prevent firms which are supplying key inputs from favouring some customers over others, promoting fair access and competition downstream. Such regulations have typically been seen in other more traditional infrastructure sectors such as telecoms where monopoly providers have often been subject to Fair Reasonable and Non-Discriminatory (FRAND) provisions or in the creation of standards in industries with patents (Silva, 2025[164]). It is unclear whether such traditional regulatory approaches would be appropriate for AI infrastructure given the rapidity of technical developments. In addition, while such provisions extend the circumstances in which suppliers must not discriminate, enforcing such provisions may have potential challenges. For example, defining terms such as ‘fair’ and ‘reasonable’ in an industry with such large outlays in research and development.

Standardisation and interoperability have also been proposed as a way to encourage effective competition in the AI stack (von Thun and Hanley, 2024[156]). This has primarily been in the context of promoting interoperability in cloud computing. However, there are also issues of interoperability in AI infrastructure if there is a need to switch between different hardware components (e.g. switching chipsets). Moves towards standards and interoperable systems can have benefits for competition in relation to reduced switching costs, but have potential risks. For example, standards may simply further embed the systems supported by dominant providers (Perez, 2017[165]). Given these challenges and the continuing rapid market developments, authorities should consider exercising caution in pushing for standards early in the market’s development. An alternative approach could be to consider advocating for government support for open-source technologies to be developed which may provide alternatives to proprietary technology, serving as the potential base for future standards.

More broadly, while regulations can potentially spur competition and innovation, they also impose costs on market participants and may have unanticipated consequences. Some commentators suggest that early intervention may be unnecessary. Hagiu and Wright (2025[166]) for example argue that each large player with a strong position in one layer has a strong incentive to commoditise the other layers, especially the ones where their market position is weaker. Hagiu and Wright (2025[166]) suggest that there may end up being at least seven major players operating at multiple (possibly, all) levels of the AI stack which would be a significant improvement compared to other digital markets.

Before considering an intervention, authorities may need stronger evidence that the risk of firms preventing access or distorting competition is sufficiently high and that traditional enforcement tools are not sufficiently timely to resolve the concerns (Meyers and Bourreau, 2025[126]).

Notes

← 1.  For example, the French Authority conducted dawn raids in an investigation in the GPU sector. Its report on Generative AI indicated that the sector is being closely scrutinised by their Investigation Services department (Autorité de la concurrence, 2024[111]). The Chinese competition authority has also opened an investigation into Nvidia regarding potential breaches of Chinese Anti-monopoly law (Reuters, 2025[177]).

← 2.  For example, the EU Commission has recently carried out unannounced antitrust inspections in the data centre construction sector. The Commission stated it had concerns firms may have violated EU antitrust rules that prohibit cartels and restrictive business practices (European Commission, 2024[173]).

← 3.  Depending on the jurisdiction, the exploitation of single firm’s market power is commonly referred to as ‘abuse of dominance’ or ‘monopolisation’ (OECD, 1996[167]).

← 4.  For example, EU case law presumes that firms are dominant with market share over 50% (OECD, 2006[183])

← 5.  There have been investigations recently reported into Nvidia the leading provider of GPUs, for example:

• The US Department of Justice has reportedly launched an investigation into Nvidia after complaints from competitors that it may have abused its market dominance in selling chips that power artificial intelligence. US DOJ investigators are reportedly looking at whether Nvidia is giving preferential supply and pricing to customers who buy its complete systems (King and Nylen, 2024[169]; TheGuardian, 2024[170]; United States Securities and Exchange Commission, 2025[181]).

• A statement from the State Administration for Market Regulation (SAMR) announcing a probe into Nvidia for potential violating China’s anti-monopoly laws. The SAMR did not elaborate on how Nvidia might have violated China's anti-monopoly laws. It also said that the US chipmaker is suspected of violating commitments it made during its acquisition of Mellanox Technologies under terms outlined in conditional approval of that deal. As part of the merger clearance of Nvidia, Mellanox Nvidia was required for six years from the date of the decision, to supply its GPUs and Mellanox’s high-speed network interconnection devices, relevant software and accessories to mainland China on fair, reasonable and non-discriminatory terms (Kharpal, 2025[179]).

← 6.  In 2025, the OECD adopted a new recommendation on merger review (OECD, 2025[174]) which: calls for a clear legal framework for merger review to be effective, efficient, and timely; provide clear principles applicable to merger notifications and review procedures; ensure that merger assessment is effective and transparent; provide clear guidance for the design, assessment, and adoption of remedies.

← 7.  For example, Nvidia recently spent over USD 900 million to hire key staff at Enfabrica and license the AI startup’s technology (CNBC, 2025[180]).

← 8.  The recent Illumina–Grail case, decided by the Court of Justice of the European Union (CJEU) in September 2024, was highly significant for EU merger control, particularly regarding jurisdiction over below-threshold mergers and the use of Article 22 of the EU Merger Regulation (EUMR). The CJEU ruled that EU Member States cannot refer mergers to the European Commission under Article 22 unless they themselves have jurisdiction to review the transaction under national law (European Union, 2025[175]).

← 9.  The EU’s Digital Markets Act (DMA) is one of the most well-known examples, imposing pre-emptive obligations on designated “gatekeepers” to ensure contestability and fair conduct in core platform services. Similar regimes have also been brought in in the UK, Korea and Japan, with several other jurisdictions also considering introducing such frameworks.