2025云计算与人工智能领域竞争政策研究报告

Issue Paper ACompetitionPolicyforCloud and AI ZachMeyersMarc Bourreau June2025 As provided for in CERRE's bylaws and procedural rules from its “Transparency & IndependencePolicy”, all CERREresearch projects and reports are completed in accordance with the strictestacademic independence. The project, within the framework of which this report has been prepared, received the supportand/or input of the following CERRE member organisations:Amazon,Arcep, BIPT,andMicrosoft.However, they bear no responsibility for the contents of this report. The views expressed in this CERREreport are attributable only to the authors in a personal capacity and not to any institution with whichthey are associated. In addition, they do notnecessarily correspond either to those of CERRE, or ofany sponsor or of members of CERRE. © Copyright 2025, Centre on Regulation in Europe (CERRE) ExecutiveSummary As European leaders focus on boosting Europe’s competitiveness, theymustensure businessesincrease their productivity. Better deployment of new technology will be an important way to achievethat. Artificial intelligence (AI) could eventually boost competition across the economy–both bydisrupting incumbents in the tech sector and by helping firms in many sectors become moreproductive, particularly services industries which have traditionally struggled to use technology toboost their output and efficiency. To maximise the economic benefits of AI foundation models, competition authorities must ensurethat competition thrives throughout the value chain–so that AI remains as cheap, high-quality andwidely available as possible and incentives to innovate are maximised. This issue paper provides an overview of how effective competition between providers of AIfoundation models is functioningtoday–and how the sector could develop. We consider upstreaminputs, in particular the provision of computing power and data for AI foundation models. This reportdoes not consider in detail downstream uses of AI–such as competition between applications anddevices that deploy AI, or future markets that might developusing AI, such asmarketsfor AI agents. Despite competition authorities’ initial worries,several potential barriers to entry for AI developershave proven less significant than feared.Currently, there is a thriving ecosystem of diverse AIfoundation models. However, there is uncertainty about the future trajectory of competition in thesector, thanks to the significant role of a few large firms across the AI value chain, the potentiallygrowingdependence of smaller models on larger ones, a shift towards lower up-front costs and higheroperational costs, and the growing importance of open AI models. At the moment, these shifts suggestthe possibility of sustainable and intense competition. However, there remain some potentialchokepoints–such as access to certain datasets, including a user’s own usage history of a service–where targeted regulatory interventions may prove necessary in order to help ensure the sectorremains contestable. The European Commission and national competition authorities are particularly focused on AIdevelopers’ relationship with the largest cloud computing providers(called ‘hyperscalers’). Therelationship between these giants and AI developers is multifaceted given the vertical integration ofthe hyperscalers across the AI value chain. Hyperscalers may: •compete to provide ‘accelerated compute’ (the specialised computing power needed to trainAI foundation models and to allow those models to produce a respond to user requests)to AIdevelopers;•provide an important channel to market for AI foundation models;•invest in many AI developers’ firms;•provide their own AI foundation models in competition with AI developers; and•be significant users of AI, which is often integrated into their other digital services. We suggest that: •Given the diversity of AI models today, the significant growth in the number of models, andthe ability of many diverse types of AI developers to attract investment, the AI sector lookscompetitive even ontraditionalstaticmetrics of market power. •However, these metrics may understate the level of competition.A dynamic competitionanalysis would point tothe frequent radical innovations achieved by AI firms, largeinvestments by large and small AI firms and private equity and venture capital investors, andthe diversity of business models in the sector as companies experiment.Barriers to entry in AIcontinue todrop, with many AI developers now using large, open-source (or open-weights)foundation models as the basis for developing more specialisedservices–avoiding the hugecosts of building an entirely new model. Similarly, the shift away from large-scale trainingtowards more fine-tuning and inference might provide more scope for smaller cloudcomputing companies, which do not have the large data centres of the hyper-scalers, to serveAI developers.While overall promising, these developmentsalsocome with some risks