人工智能计算主权：跨地区、云服务商和加速器的基础设施控制

Zoe Jay Hawkins, Tech Policy Design Institute & Oxford Internet Institute, Universityof OxfordVili Lehdonvirta, Department of ComputerScience, Aalto University&Oxford Internet Institute, Universityof OxfordBoxi Wu, Oxford Internet Institute, Universityof Oxford Abstract:The concept of‘compute sovereignty’has become a focal point in government and industry discussionson artificial intelligence (AI) governance. What is it and who has it?Based on previous literature, weproposetobreak these questions down tothree levels: (1) how much AI compute a country has on its territory, (2) what is thenationality of the companies who own the AI compute data centres, and (3) what is the nationality of the acceleratorvendors whose chips power the AI compute data centres? We examine these questions empirically through the lensof cloud computing infrastructure, focusing on nine leading public cloud providers that represent approximately 70percent of the global market.The data is collected using a methodology previously published in Lehdonvirta, Wu,and Hawkins (2024).The findings suggest thatthepossession of “compute sovereignty” varies between countriesdepending on the level of analysis. Determining the most relevant level depends on governments’ policy aims andnational contexts,and involves policy trade-offs. Policies aimed at attracting data centres to a country’s territory canenhance supply security of critical computational resources while also introducing increased consumptionof energy,water and land use resources,with corresponding localised socioeconomicand environmental impacts. Regional andsupply chain approaches involve different trade-offs. 1.Introduction In a time ofincreasing geopolitical competition and uncertainty, the location, ownership and control of artificialintelligence (AI) technologies and their underlying material infrastructures have become strategic considerations forgovernments(Wang and Chen, 2018; Miller, 2022).AI compute–the specialised computational resourcesthatarerequired to train and run AI models–has emerged as asought-aftercommodity inaglobal race for AI leadership.Theamount ofcompute required to train frontier AI modelshasdoubledapproximately every six months (Sevilla etal, 2022), driven by continued pursuit of computational scalingof AI performance(Pilz et al., 2025). For example,progress in advanced reasoning capabilities in models such as OpenAI’s o3 or DeepSeek’s R1 have relied oncontinued computational scaling during inference time.NVIDIACEO Jenson Huangclaimedthat “the amount ofcomputation we need as a result of agentic AI, as a result of reasoning, is easily 100 times more than we thought weneeded this time last year” (Nellis & Cherney, 2025).The perceived need to secure sufficient access to AI computefor domestic firms, public sector organizations, and/orresearchershas come to be referred to using such terms as“sovereign AI compute” or “compute sovereignty” (Ghahramani, 2023;Sastry et al., 2024; Lehdonvirta et al., 2024;Government of Canada, 2025).As policymakers seek to assert regulatory authority, ensure supply chain resilience,and protect strategic autonomy, ‘compute sovereignty’ is becoming a topic of interest in debates about AIgovernance (Sastry et al., 2024).Sovereignty–understood broadly as a government’s supreme authority within a given territory–has a complex and contested relationship with digital technologies (Pohle & Thiel, 2020). Policies pursuing digitalsovereignty haveevolved significantly over the last three decades and particularly since the early Internet debates ofthe 1990s. Different stakeholders hold differing views on digital sovereignty; from autocratic arguments regardingthe primacy of the nation state in Internet governance (as opposed to multistakeholder approaches), through tonations’ increasing pursuit of autonomy from foreign government interference (Musiani, 2024). In westerndemocracies, digital sovereignty has recently been conceptualised through thelens of national or regional strategicautonomy in the digital realm(Blancato, 2024).Amid rising competition to develop advanced AI systems and the corresponding scarcity of computational resources, notions of ‘digital sovereignty’ are increasing entwined with debates about the governance of AI and theemerging paradigm of compute governance (Roberts et al., 2023; Sastry et al., 2024; Lehdonvirta et al., 2024). Since2022, almost all western democratic countries have developed policies and initiatives to increase strategic autonomyover computational resources, including a turn to industrial policies that aim to onshore the compute supply chain.Indoing so, governments are assessing their dependence on corporate and state actors who own or control key nodes inthe vast, global supply chain that characterises AI compute; from the owners of AI data and models, to the producers of valuable AI accelerators and the developers and operators of the data centres that house accelerators. Throughthese assessments–su