2026-02-13-sovereign-ai-strategies-sandoval-et-al

Research Paper How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies Francisco Javier Varela Sandoval, Isabella Wilkinson, Alex Krasodomski and Rowan Wilkinson Digital Society Programme February 2026 -- 1 of 48 -- Chatham House, the Royal Institute of International Affairs, is a world-leading policy institute based in London. Our mission is to help governments and societies to build a secure, sustainable, prosperous and just world. -- 2 of 48 -- 1 Chatham House Contents Summary 2 01 Introduction 4 02 Why build sovereign AI? 7 03 Strategy meets reality 18 04 The road ahead 32 05 Conclusion 41 About the authors 43 Acknowledgments 44 -- 3 of 48 -- 2 Chatham House Summary — Middle powers that fail to secure influence over the development, deployment and governance of artificial intelligence (AI) will likely forfeit control over their economies, societies, political systems and positions in the global economy. ‘Sovereign AI’ capabilities – that is, the extent to which countries can preserve autonomy in their AI policy choices, ensure the growth of AI meets national interests, and avoid creating unhealthy dependencies on external actors – are thus becoming as fundamental to national power as military strength or economic policy. — Today, middle powers face a critical choice: where to compete, where to partner, and where to accept dependency. While these countries cannot match the scale of the US or China in terms of investment and advanced technology, middle powers still have leverage to shape the next decade of AI. Failure to make deliberate strategic choices now could result in the loss of autonomy in an intensifying global AI race. — Middle powers can pursue one of four pragmatic approaches to AI sovereignty. Each involves trade-offs. These paths will shift as the technology evolves: — Specialize – carve out limited niches in the global AI supply chain that provide leverage and partial autonomy. — Align – side fully with an AI superpower, accepting deep dependencies in exchange for access and protection. — Share – pool sovereignty with like-minded countries through blocs and partnerships to amplify collective influence. — Hedge – cherry-pick capabilities from a diverse range of foreign suppliers, while building national capacity in targeted areas. — Countries must assess their position across eight critical building blocks – data, compute (computational power), advanced models, energy, industry, talent, infrastructure and trust – and determine where they can realistically build capability versus where they must depend on others. In terms of AI, no middle power should expect to operate independently. The critical question is not whether to depend on other nations for AI, but on which nations, to what extent and on what terms. -- 4 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 3 Chatham House — Current national motivations in pursuing AI sovereignty differ. Some countries prioritize national security, while others focus on economic growth, public service capacity, geopolitical leverage, values alignment or frontier AI leadership. But in this sphere all middle powers share two common challenges: 1) building autonomy and competing for opportunities in the global AI value chain despite limited resources, and 2) strategizing amid uncertainty. — Given the unpredictability of AI trajectories, this paper calls for pragmatic strategy over maximalism. The goal is not independence (as full autonomy remains unrealistic) but strategic flexibility: the ability to switch providers, adapt to disruptions and avoid coercion. — Middle powers that act strategically can secure durable influence over the technologies that underpin their political systems, economies and societies, not to mention their international standing. AI sovereignty will inevitably remain partial for these countries, but efforts in this direction will not be in vain. Those that delay or pursue unfocused strategies risk being left with dependencies they did not choose and ambitions they cannot realize. -- 5 of 48 -- 4 Chatham House 01 Introduction States are taking steps to improve national AI capabilities and technological control. But with the US and China dominating AI, complete independence remains unreachable. Middle powers need pragmatic pathways to seek meaningful AI sovereignty. The drive for sovereignty Strengthening the capacity of states to exercise power over digital technology is not a new challenge. For decades, governments have grappled with the ways that digital technology has reshaped their economies, societies and cultures, and the limits of their own national capacity to influence that transformation.1 Artificial intelligence (AI) has reinvigorated debates over what constitutes sovereignty in relation to technology. This is especially true among middle powers, which, in the context of AI, are understood as countries that cannot compete with the US and China in the AI race, but nonetheless aim to influence AI’s development, deployment and governance – both on a national and international level. At the AI Action Summit in Paris in February 2025, French president Emmanuel Macron described the future of AI as ‘presenting a political challenge, a question of sovereignty and strategic self-determination’,2 later remarking that ‘there is no such thing as happy vassalage’.3 Various EU policy initiatives, such as the International Digital Strategy, have called for the bloc to commit to sovereign digital infrastructure as a strategic imperative for becoming more technologically 1 Edler, J., Blind, K., Kroll, H. and Schubert, T. (2023), ‘Technology sovereignty as an emerging frame for innovation policy. Defining rationales, ends and means’, Research Policy, 52(6), https://doi.org/10.1016/j.respol.2023.104765. 2 Macron, E. (2025), ‘L’intelligence artificielle, dont le développement ne cesse de s’accélérer, bouleverse nos vies par ses immenses pouvoirs’ [Artificial intelligence, whose development continues to accelerate, is revolutionizing our lives with its immense power.], LinkedIn Pulse, 9 February 2025, https://www.linkedin.com/pulse/ lintelligence-artificielle-dont-le-d%C3%A9veloppement-ne-cesse-macron-arpue. 3 Hoez, J. (2025), ‘Macron’s ‘Choose Europe for Science’ speech’, The French Dispatch, 5 May 2025, translation, https://www.frenchdispatch.eu/p/macrons-choose-europe-for-science-speech-sorbonne. -- 6 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 5 Chatham House independent and to strengthen security across the region.4 The International Digital Strategy was published soon after the new EU AI Continent Action Plan, which outlined steps for the EU to become a global leader in AI.5 Dozens of other countries have now set out their own national strategies for ‘sovereign AI’ – the ability of a country to influence, develop and deploy AI in line with national interests. The UK’s AI Opportunities Action plan promises major sovereign AI infrastructure investment, international partnerships and incentives to bring AI companies onshore; prompting commitments for more than £14 billion in fresh inward investment across the country.6 Further afield, Saudi Arabia has set out its ambition to become an AI superpower by 2030 through capital investments, headlined by the launch of a state-owned company and major US deals on AI and defence.7 No two countries’ sovereign AI strategies are completely alike. But strategic ambitions aside, no middle power can feasibly compete with the scale and expenditure of US and Chinese sovereign AI efforts.8 How states – regardless of size and influence – might fit into an AI world order is still unclear. Some voices – notably major US AI companies – argue that the world is on the cusp of an AI revolution that will result in the greatest transformation of political and economic power in human history. Sceptics are less sure, both on the timeline and the scale of the projected change.9 For governments, their sovereign AI strategies represent attempts to gain leverage over a technology expected to shape economic competitiveness, national security and societal resilience. This paper’s core objective is a pragmatic assessment of the gap between the current AI superpowers – the US and China – and middle powers, and to highlight realistic strategies for middle powers to build and exercise AI sovereignty in that context. Full AI self-sufficiency – even for AI superpowers – is not possible. AI’s global supply chains are too interconnected, and interdependencies run deep. Middle powers are severely constrained on nearly every critical input into building sovereign AI, from talent and data to raw materials and semiconductor chips. We conclude that complete technological independence is unachievable for most countries, but certain levels of control and leverage are possible. 4 European Commission (2025), ‘The EU sets out its International Digital Strategy’, press release, 5 June 2025, https://ec.europa.eu/commission/presscorner/detail/en/ip_25_1370. 5 European Commission (2025), ‘The AI Continent Action Plan’, https://digital-strategy.ec.europa.eu/en/library/ ai-continent-action-plan. 6 UK Department for Science, Innovation and Technology (DSIT) (2025), ‘AI Opportunities Action Plan’, https://www.gov.uk/government/publications/ai-opportunities-action-plan/ai-opportunities-action-plan; Adams, L. (2025), ‘UK unveils £14b AI infrastructure plan’, Tech EU, 13 January 2025, https://tech.eu/2025/ 01/13/uk-unveils-14b-ai-infrastructure-plan. 7 England, A. and Al Omran, A. (2025), ‘Donald Trump lauds Saudi Arabia as he unveils AI and defence deals’, Financial Times, 13 May 2025, https://www.ft.com/content/5302d5d2-d375-4327-905c-7b1ad5173105; Harshan, A. (2025), ‘Saudi Vision 2030: The rise of humain & Kingdom’s strategic AI ambitions’, Global Business Outlook, 19 May 2025, https://globalbusinessoutlook.com/technology/saudi-vision-rise-humain-kingdoms- strategic-ai-ambitions. 8 Marshall, C. and E&E News (2025), ‘Here’s What’s in ‘Stargate,’ the $500-Billion Trump-Endorsed Plan to Power U.S. AI’, Scientific American, 22 January 2025, https://www.scientificamerican.com/article/heres-whats-in-stargate- the-usd500-billion-trump-endorsed-plan-to-power-u-s; Chan, K. et al. (2025), ‘Full Stack: China’s Evolving Industrial Policy for AI’, RAND, 26 June 2025, https://www.rand.org/pubs/perspectives/PEA4012-1.html; Ma, S. and Tan, G. (2025), ‘China sets up 60b yuan national AI fund to accelerate tech innovation’, China Daily, 4 April 2025, https://www.chinadaily.com.cn/a/202504/18/WS6802358ea3104d9fd38204b5.html. 9 Bengio, Y. (eds) (2025), ‘International AI Safety Report 2025’, UK Department of Science, Innovation and Technology, 29 January 2025, https://www.gov.uk/government/publications/international-ai-safety-report-2025. -- 7 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 6 Chatham House Chapter 2 maps the rationales for middle power investments in sovereign AI, ranging from national security to economic competitiveness. Chapter 3 proposes that the abilities of middle powers to exercise sovereignty over primary (the technical foundations) and secondary (the enabling environment) building blocks for AI are constrained by entrenched dependencies on the US and China. Chapter 4 argues for an urgent reset in strategic thinking, sketching out pragmatic, informed pathways for different forms of sovereign AI. The recommendations have relevance to a diversity of national contexts excluding the US and China. Methodology For this paper, the authors conducted research on policy and strategic documentation associated with national approaches to AI, contextualizing this information with secondary sources (news releases, scientific reports, company updates, analysis and commentary from media, research institutes and universities). Countries were selected on the basis of diversity in their AI implementation, innovation and investment (gauged by the Global AI Index 2024) and the prospects for regional cooperation (particularly in Europe and Southeast Asia). An important part of these criteria is the presence of a national strategic plan on, or related to, AI.10 EU-wide AI initiatives are included alongside national plans. The authors supplemented and validated this research with anonymized, informal, semi-structured interviews with middle power sovereign AI builders from several of the case study countries and workshops, including representatives from the EU (France and Belgium), ASEAN (Thailand, Singapore and Vietnam), and the UK. The paper’s approach also benefits from insights gathered during the Paris AI Action Summit and the International Association for Safe and Ethical AI Conference, both held in February 2025. 11 All countries have a stake in the development of AI. This paper’s approach and recommendations are designed to encourage strategic thinking among a diverse array of policymakers. The work should be of particular interest to strategic decision-makers in middle powers. Exercising power over AI is a new and unfamiliar undertaking, with few options and many constraints. But there are also emerging pathways for states to carve out autonomy in the sector, including new multilateral forums for decision-making, the proliferation of big data and increasingly globalized supply chains. The nature of AI as a networked technology with multiple inputs – not to mention its entanglement in geopolitical dynamics – means that any ambition to deploy AI in the domestic interest within national borders has an inescapably international strategic dimension. 10 The authors note that the 18 countries (and EU) surveyed is a small snapshot of global sovereign AI initiatives and hope to conduct a broader survey in future work. 11 Noting approaches to building sovereign AI capabilities around the world are in flux and rapidly developing, this paper is a depiction of the global state of sovereign AI as of January 2026. -- 8 of 48 -- 7 Chatham House 02 Why build sovereign AI? Motivated by the fear of falling behind in the global AI race, middle powers have introduced sovereign AI strategies with a focus on national security, economic advantage, public capacity and geopolitical positioning. Increasingly wary about being left behind in the AI race, middle powers have rushed to publish sovereign AI strategies. These strategies call for national investment and policy change in order to ensure that AI is a boon to national security, the economy, modern public service provision, national values and culture, as well as a bargaining chip in geopolitical competition or cooperation. Among the countries (and the EU) surveyed for this paper, there is no common agenda or set of shared principles underlying different sovereign AI strategies. But there are a recurring set of interlinked rationales, presented in Table 1, which provide a foundation for comparative analysis.12 Each rationale category spotlights combinations of countries to provide a snapshot of how diverse sovereign AI ambitions share characteristics. 12 Research draws from primary (i.e. state and industry policies, investment and action plans) and secondary (e.g. news and independent commentary and analysis from research institutes) sources. National investment and policy change can ensure that AI is a boon to national security, the economy, modern public service provision, national values and culture, as well as a bargaining chip in geopolitical competition or cooperation. -- 9 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 8 Chatham House Table 1. Rationales underlying national sovereign AI strategies and global rankings Global AI Index 2024 edition National security Economic growth Public service capacity- building Leverage geopolitical competition Values alignment Domination of the frontier Middle powers Canada 8th • • France 5th • • Germany 7th • • Japan 11th • • • India 10th • • Indonesia 49th • • • Israel 9th • • • • Saudi Arabia 14th • • • • Singapore 3rd • • • South Korea 6th • Spain 18th • • Sweden 25th • • • Thailand 43rd • • • United Arab Emirates 20th • • • • United Kingdom 4th • • • Vietnam 58th • • • European Union – • • • • AI superpowers United States 1st • • • • • • China 2nd • • • • • • Source: Compiled by the authors, based on Tortoise (2024), ‘The Global AI Index’, https://www.tortoisemedia.com/data/global-ai#rankings. Note: The Global AI Index ranks 83 countries based on their international AI capacity. They are scored based on 83 indicators, falling within three umbrella areas: implementation, innovation and investment. -- 10 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 9 Chatham House National security For most countries, national security is a pressing reason for building and investing in domestic AI capabilities. The US frames technological leadership as essential to its national security – seeking to onshore critical infrastructure, export AI only to trusted allies and restrict the supply of advanced chips to ‘countries of concern’. 13 The Stargate Project, which was announced in 2025, embodies this strategy, combining massive domestic investment ($500 billion) in AI infrastructure with efforts to curb China’s competitiveness. 14 Likewise, China’s approach is security-driven, focused on regime stability, public safety and control over data and AI model use.15 China’s interim measures for governing AI acknowledge the importance of ‘placing equal emphasis on development and security’. 16 Focusing on middle powers, the UK’s approach to sovereign AI is supported by a different national security rationale. The country’s Strategic Defence Review 2025 commits to ‘innovat[ing] at a wartime pace’, denoting advanced AI systems as essential for next-generation defence.17 Meanwhile, the AI action plan identifies three compute (computational power) tiers: publicly owned or allocated, privately owned but UK-based, and international via foreign partners.18 While fully homegrown AI capabilities might be preferential from a national security perspective, for the UK, such developments are not realistic in economic or practical terms.19 Israel also harbours ‘AI superpower’ ambitions, notably captured by its AI National Program. 20 Its sovereign AI strategy is highly securitized. The country has long prioritized building technology and cybersecurity hubs for international trade and to bolster its national security. Its investments in sovereign AI capabilities – for example, in military robotics and autonomous warfare21 – fulfil both aims. For other middle powers, a growing national security driver of sovereign AI is the legal and jurisdictional risk of relying on foreign cloud and AI providers, especially those in the US. Mistrust of US data governance and technology infrastructure 13 The White House (2025), ‘FACT SHEET: Ensuring U.S. Security and Economic Strength in the Age of Artificial Intelligence’, 13 January 2025, https://bidenwhitehouse.archives.gov/briefing-room/statements-releases/ 2025/01/13/fact-sheet-ensuring-u-s-security-and-economic-strength-in-the-age-of-artificial-intelligence. 14 da Silva, J., Sherman, S. and Rahman-Jones, I. (2025), ‘Tech giants are putting $500bn into ‘Stargate’ to build up AI in US’, BBC News, 22 January 2025, https://www.bbc.co.uk/news/articles/cy4m84d2xz2o. 15 For an explainer, see: Citak, E. (2023), ‘China ChatGPT ban: Implications for the chatbot industry and beyond’, Tech Briefly, 24 February 2023, https://techbriefly.com/2023/02/24/china-chatgpt-ban-implications. 16 China Law Translate (2023), ‘Interim Measures for the Management of Generative Artificial Intelligence Services’, 13 July 2023, https://www.chinalawtranslate.com/en/generative-ai-interim; Murphy, B. (2021), ‘Outline of the People’s Republic of China 14th Five-Year Plan for National Economic and Social Development and Long-Range Objectives for 2035’, translated 12 May 2021 by Etcetera Language Group, Inc., Centre for Security and Emerging Technology, https://cset.georgetown.edu/wp-content/uploads/t0284_14th_Five_Year_Plan_EN.pdf; Creemers, R. et al. (2022), ‘14th Five-Year Plan for National Informatization’, translation published 24 January 2022, Stanford Digi China, https://digichina.stanford.edu/wp-content/uploads/2022/01/DigiChina-14th-Five-Year-Plan-for-National- Informatization.pdf. 17 UK Ministry of Defence (2025), ‘Strategic Defence Review: Making Britain Safer: secure at home, strong abroad’, https://assets.publishing.service.gov.uk/media/683d89f181deb72cce2680a5/The_Strategic_Defence_Review_ 2025_-_Making_Britain_Safer_-_secure_at_home__strong_abroad.pdf. 18 UK DSIT (2025), ‘AI Opportunities Action Plan’. 19 Aspen Digital and Chatham House (2025), ‘Strategic Reorientation on A.I. Competition with China’, 6 February 2025, https://www.aspendigital.org/report/strategic-reorientation-on-ai-competition-with-china. 20 Israel National AI Program (undated), ‘Shaping Israel’s AI future: Strategy, Infrastructure, Operating Environment’, https://aiisrael.org.il. 21 Reuters (2023), ‘Israel aims to be ‘AI superpower’, advance autonomous warfare’, 22 May 2023, https://www.reuters.com/world/middle-east/israel-aims-be-ai-superpower-advance-autonomous-warfare- 2023-05-22. -- 11 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 10 Chatham House shapes middle powers’ national AI and cloud procurement strategies. This anxiety is pushing countries towards sovereign cloud solutions, domestic compute capacity and open-source models that reduce exposure to foreign jurisdictions and regulations.22 Concerns about the US CLOUD (Clarifying Lawful Overseas Use of Data) Act 2018 – which allows American authorities to compel US-headquartered companies to provide access to data stored both in the US and abroad – are growing, even among allies. In November 2025, a Canadian government white paper warned that the country’s data sovereignty is at risk if legal control over data stored in Canada is claimed and acted upon by another jurisdiction.23 In the same month, the European Commission launched market investigations into cloud computing services from Amazon and Microsoft to determine if their practices limit competitiveness in the EU cloud computing sector, in accordance with the Digital Markets Act’s provisions. 24 China’s data governance regime – including the mandatory requirement for companies storing data within its jurisdiction to enable state access25 – has long sparked global concern. Anxieties about over-dependence on China-supplied technology – and its implications, such as surveillance and strategic risks – have shaped technology policies in London,26 Brussels27 and Washington.28 22 US Department of Justice (undated), ‘CLOUD Act Resources’, Criminal Division, https://www.justice.gov/ criminal/cloud-act-resources. 23 Karadeglija, A. (2025), ‘Data stored in Canada can be subject to foreign courts, government paper warns’, The Canadian Press, 3 November 2025, https://www.thecanadianpressnews.ca/politics/data-stored-in-canada-can- be-subject-to-foreign-courts-government-paper-warns/article_70b5ea4f-b431-5d93-9dbd-1a17c3a39398.html. 24 European Commission (2025), ‘Commission launches market investigations on cloud computing services under the Digital Markets Act’, press release, 18 November 2025, https://ec.europa.eu/commission/presscorner/ detail/en/ip_25_2717. 25 Creemers, R., Webster, G. and Triolo, P. (2018), ‘Translation: Cybersecurity: Law of the People’s Republic of China (Effective June 1, 2017)’, Stanford University DigiChina, https://digichina.stanford.edu/work/ translation-cybersecurity-law-of-the-peoples-republic-of-china-effective-june-1-2017. 26 Bowler, T. (2020), ‘Huawei: Why is it being banned from the UK’s 5G network?’, BBC News, 14 July 2020, https://www.bbc.co.uk/news/newsbeat-47041341. 27 Tagliapietra, S., Trasi, C. and Sebastian, G. (2025), ‘A smart European strategy for electric vehicle investment from China’, Bruegel, 16 July 2025, https://www.bruegel.org/policy-brief/smart-european-strategy-electric- vehicle-investment-china. 28 Wilkinson, I. and Wilkinson, R. (2025), ‘The TikTok transfer raises worrying questions for allies like the UK’, Chatham House Expert Comment, 3 October 2025, https://www.chathamhouse.org/2025/10/tiktok-transfer- raises-worrying-questions-allies-uk. Concerns about the US CLOUD Act 2018 – which allows American authorities to compel US-headquartered companies to provide access to data stored both in the US and abroad – are growing, even among allies. -- 12 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 11 Chatham House Economic growth Building domestic capabilities for national economic growth is another prevalent rationale in the AI sovereignty strategies reviewed for this paper. Vietnam’s 2021 strategy on AI outlines the country’s aspirations to be a ‘center for [AI] innovation’, and to become one of ASEAN’s top four nations on AI R&D and applications by 2030.29 The country’s strategy commits to developing a competitive AI ecosystem – including attracting investment and mobilizing capital – and promoting the application of AI in businesses.30 In December 2024, the Vietnamese government and major semiconductor designer and producer NVIDIA agreed to open a new AI R&D centre, focusing on software development to accelerate AI adoption.31 AI automation in manufacturing and agriculture, two of Vietnam’s major industries, could be highly beneficial for the country.32 Japan’s approach to building sovereign AI capabilities also centres on economic advantages and productivity. Recent deals with NVIDIA commit to public, private and academic partnerships geared towards accelerating adoption and ‘homegrown’ innovation.33 A common theme of these new initiatives is a commitment to protect Japan’s data sovereignty with tailored solutions. 34 The country’s approach to AI benefits from the population’s ‘high degree of affinity with generative AI’ due to existing digitalization, the robotics industry, technology R&D and is partly motivated by ‘a sharp decline in the working population’. 35 This state of play highlights twin drivers of sovereign AI investments: a sober assessment of future productivity challenges and existing economic advantages. 36 Further afield, Canada’s Sovereign AI Compute Strategy directly responds to the constraints faced by middle powers: high costs and limited domestic capacity in compute. The country’s strategy outlines major commitments to public supercomputing infrastructure and an AI Compute Access fund as drivers of economic growth.37 Moving to Europe, Germany was one of the first countries to publish an AI strategy, launched in 2018. The strategy calls for Made in Germany AI (or Made in Europe AI) to leverage existing economic strengths to create both value and societal 29 Vietnam government (2021), ‘National Strategy on Research and Development and Application of AI (2021- 2030)’, 10 December 2021, accessed via OECD, https://wp.oecd.ai/app/uploads/2021/12/Vietnam_National_ Strategy_on_RD_and_Application_of_AI_2021-2030.pdf. 30 Ibid. 31 NVIDIA (2024), ‘NVIDIA to Open Vietnam R&D Center to Bolster AI Development’, press release, 5 December 2024, https://nvidianews.nvidia.com/news/nvidia-to-open-vietnam-r-d-center-to-bolster-ai-development. 32 AI for Vietnam Foundation (2025), ‘What is AI, and why does it matter for Vietnam’s future?’, 10 February 2025, https://www.aiforvietnam.org/what-is-ai-and-why-does-it-matter-for-vietnam-future. 33 Osaki, M. (2024), ‘Japan tech Leaders Supercharge Sovereign AI With NVIDIA AI Enterprise and Omniverse’, NVIDIA Blogs, 12 November 2024, https://blogs.nvidia.com/blog/japan-tech-leaders-sovereign-ai-omniverse. 34 Ibid. 35 Japan AI Strategy Council (2023), ‘Tentative Summary of AI Issues’, May 2023, https://www8.cao.go.jp/cstp/ ai/ai_senryaku/2kai/ronten_youshi_eiyaku.pdf. 36 For further reading, see: Ichikawa, T. (2025), ‘Norms in New Technological Domains: Japan’s AI Governance Strategy’, CSIS, 17 June 2025, https://www.csis.org/analysis/norms-new-technological-domains-japans-ai- governance-strategy; Japan Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry (2024), ‘AI Guidelines for Business Ver1.0’, 19 April 2024, https://www.meti.go.jp/shingikai/mono_ info_service/ai_shakai_jisso/pdf/20240419_9.pdf. 37 ISED Canada (2025), ‘Canadian Sovereign AI Compute Strategy’, updated 6 May 2025, https://ised-isde.canada. ca/site/ised/en/canadian-sovereign-ai-compute-strategy. -- 13 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 12 Chatham House benefit.38 This strategy links with efforts in other areas designed to enhance German and European AI leadership, such as upskilling and training strategies, as part of a €5 billion federal package on AI.39 As the country seeks to build domestic AI capabilities and improve its competitiveness, Germany’s sovereign AI rationale depends on coordination with other middle powers in the EU bloc. Public service capacity-building Many countries view the development of sovereign capabilities in AI as an enabler of other public policy goals. These efforts are often consistent with whole-of- government digitalization strategies. This is certainly the case in Singapore, India, Indonesia and Thailand, four South and Southeast Asian countries with markedly different political systems and levels of digitalization, but similar public capacity ambitions for AI. Singapore’s Smart Nation initiative aims to advance the digitalization and therefore efficiency of public services with an ‘AI for public good’ approach, from healthcare and housing to security and immigration control. 40 India is also a hub of growing digital public infrastructure (DPI) solutions – secure, open and interoperable digital systems – across the public sector, such as the country’s digital ID system, Aadhaar.41 India’s strategic approach to building sovereign AI capabilities is similarly rooted in public capacity ambitions, particularly in relation to AI for social welfare in healthcare, agriculture and linguistic diversity. 42 Despite their economic and infrastructural differences, Singapore and India’s strategic approaches clearly promote the use of AI to strengthen governing capacity. For Indonesia, Southeast Asia’s largest digital market, investments in sovereign AI are viewed as essential for catering to the country’s different cultures and languages, building public capacity and opening the door to AI applications. The multilingual Sahabat-AI models are a case in point: a collection of open-source large language models (LLMs) ‘built by Indonesians for Indonesians’, which are the result of a public–private partnership involving NVIDIA.43 The country’s latest strategy emphasizes the need to avoid foreign dependencies and identifies five priority areas – 38 German Federal Ministry on Research, Technology and Space (undated), ‘Artificial Intelligence’, https://www.bmbf.de/EN/Research/EmergingTechnologies/ArtificialIntelligence/artificialintelligence_node.html. 39 European Commission (undated), ‘Germany AI Strategy Report’, https://ai-watch.ec.europa.eu/countries/ germany/germany-ai-strategy-report_en. 40 Smart Nation Singapore (2025), ‘The Smart Nation vision: Discover the foundation of Singapore’s vision to be a Smart Nation’, updated 10 July 2025, https://www.smartnation.gov.sg/about/our-vision/smart-nation- vision; Smart Nation Singapore and Ministry of Digital Development and Information (2024), ‘Smart Nation 2.0: A Thriving Digital Future for All’, https://file.go.gov.sg/smartnation2-report.pdf. 41 Naidu, S. et al. (2020), ‘Mapping Digital Identity Systems: India’, Digital Identities: Design and Uses’, 13 October 2020, https://digitalid.design/research-maps/india.html. 42 Mohanty, A. and Sahu, S. (2024), ‘India’s AI Strategy: Balancing Risk and Opportunity’, Carnegie India, 22 February 2024, https://carnegieendowment.org/posts/2024/02/indias-ai-strategy-balancing-risk-and- opportunity?lang=en. 43 Gomes, J. (2024), ‘Indonesia Tech Leaders Team With NVIDIA and Partners to Launch Nation’s AI’, NVIDIA Blogs, 13 November 2024, https://blogs.nvidia.com/blog/indonesia-tech-leaders-sovereign-ai. -- 14 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 13 Chatham House all in the public sector – including bureaucratic reform and smart cities (using digital technology to improve urban living). Indonesia’s forthcoming AI roadmap is expected to provide further clarity.44 Finally, Thailand’s strategic approach to developing sovereign AI links the building of public capacity with overall regional competitiveness. Thailand’s ambition is, like Indonesia and Singapore, to become a leading AI hub. Strengthening governing capacity is critical to achieving this goal. Annual reporting by the Thai government recognizes the importance of ‘AI readiness’ in policymaking,45 while the country’s strategy develops in line with the number of users (in the public and private sectors) adopting AI for innovation.46 Front and centre of the national strategy is the potential for AI-powered public services to improve well-being, quality of life and governing efficiency.47 LLMs developed for the Thai language may accelerate these efforts.48 Leverage geopolitical competition Building domestic AI capabilities can also further geopolitical and strategic aspirations. While the US–China AI race is a major geopolitical consideration in developing these technologies, it is by no means the only factor. Seeking to maximize their AI sovereignty – without necessarily striving for global AI dominance – middle powers must navigate competing offerings from leading AI providers. Some middle powers, such as the United Arab Emirates (UAE), have picked a side. Others are more agile in navigating the geopolitical crosswinds of US–China AI competition, aiming to build a third way. The US and China have long appreciated the geopolitical value of the UAE, due to its access to cheap electricity and significant public wealth – including a commitment to invest $200 billion in AI infrastructure.49 In the last year, however, the UAE has moved closer to the US,50 leveraging its position for an unprecedented deal: the country will host OpenAI’s first international deployment of Stargate. This deal is part of a series of major investments into data centres – with the 1 gigawatt (GW) Stargate UAE cluster in Abu Dhabi – and strategic partnerships with US and UAE AI companies, including the Acceleration Partnership deal with the US government.51 For the UAE, the reasoning is clear: geopolitical competition can be leveraged to secure investments and partnerships for the country’s sovereign AI. 44 Sagena, U. (undated), ‘Priorities and Challenges of Indonesia’s Artificial Intelligence National Strategy (Stranas KA)’, SAFEnet, https://safenet.or.id/2022/05/priorities-and-challenges-of-indonesias-artificial-intelligence-national- strategy-stranas-ka. 45 AI Thailand (2024), ‘Annual Report 2024’, https://www.ai.in.th/wp-content/uploads/2025/06/NAIS-Annual_ 2024_ENG_Web.pdf. 46 AI Thailand (undated), ‘Thailand national AI strategy and action plan (2022-2027)’, https://www.ai.in.th/en/ about-ai-thailand. 47 AI Thailand (2022), ‘Thailand National AI Strategy and Action Plan (2022-2027)’, https://www.ai.in.th/ wp-content/uploads/2022/12/2022-NAIS-Presentation-eng.pdf. 48 For example, see SCB (2024), ‘Typhoon’, https://www.scb.co.th/en/about-us/news/jan-2024/scb-10x-typhoon. 49 Bhat, D. (2025), ‘Can the Gulf buy its way to AI supremacy?’, Rest of World, 24 June 2025, https://restofworld.org/ 2025/gulf-ai-investment-us-china-race; Business Times (2025), ‘Huawei seeks AI chip customers in Middle East, South-east Asia’, Business Times Singapore, 11 July 2025, https://www.businesstimes.com.sg/international/ global/huawei-seeks-ai-chip-customers-middle-east-south-east-asia. 50 Cornwell, A. (2024), ‘UAE seeks closer AI, tech ties in Biden talks as China interest stirs US concern’, Reuters, 23 September 2024, https://www.reuters.com/technology/uae-seeks-closer-ai-tech-ties-biden-talks-china- interest-stirs-us-concern-2024-09-23. 51 OpenAI (2025), ‘Introducing Stargate UAE’, 22 May 2025, https://openai.com/index/introducing-stargate-uae. -- 15 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 14 Chatham House Saudi Arabia’s newly launched sovereign AI initiative – HUMAIN, a Public Investment Fund-owned company – faces a similar geopolitical landscape but adopts a slightly different rationale. The country’s national strategy outlines Saudi Arabia’s aspirations to be a global leader in AI by 2030, with the aim of attracting at least $20 billion in investment.52 HUMAIN emphasizes Saudi independence (from US and Chinese AI offerings) with major investments in local, Arabic LLMs.53 At the same time, Chinese technology companies – particularly Huawei Cloud – are proactive in Saudi Arabia, supporting government cloud usage and brokering strategic partnerships across the telecommunications sector. Analysis suggests this two-pronged approach means Chinese AI has a central role in public–private partnerships while strictly adhering to local data laws.54 This presence is a sovereign AI puzzle for Saudi Arabia, building sovereign capabilities but reckoning with intense dependencies on China-supplied infrastructure. South Korea’s approach focuses on more investment. Gaining geopolitical autonomy is a priority issue for the country, having announced a $75 billion investment in sovereign AI in June 2025, with a new AI policy unit and a dedicated presidential secretary on AI.55 Crucially, the lead strategist has called for these investments in an effort to achieve ‘full-stack’ sovereignty – that is, end-to-end control of the AI system so no critical part depends on a foreign company or government – and freedom from the ‘neo-imperialism’ of the US–China technology race.56 Alignment with national values National strategies on technology and sovereignty are heavily influenced by societal, ethical and political values, alongside discrete notions of public good. These can significantly contrast between nations, which has implications for cooperation. EU efforts to develop sovereign AI illustrate the core considerations of values alignment for democracies. The EU’s 2025 AI Continent Action Plan champions democratic values, cultural diversity and trustworthy and human-centric AI.57 Across the EU, efforts have focused on the development of open, multilingual models that cater to European languages to enable broader societal input, which can better reflect local values.58 For example, the Open Euro LLM, which is a series of models that promote transparency in AI, seeks to build and improve access to multilingual foundation models that can then be fine-tuned for local applications.59 Similarly, AI 52 Saudi Data and AI Authority (undated), ‘National Strategy for Data & AI’, https://sdaia.gov.sa/en/SDAIA/ SdaiaStrategies/Pages/NationalStrategyForDataAndAI.aspx. 53 Public Investment Fund Saudi Arabia (2025), ‘HRH Crown Prince launches HUMAIN as global AI powerhouse’, 12 May 2025, https://www.pif.gov.sa/en/news-and-insights/press-releases/2025/hrh-crown-prince-launches- humain-as-global-ai-powerhouse. 54 Benito, A. (2025), ‘How China is gaining ground in the Middle East cloud computing race’, Rest of World, 5 May 2025, https://restofworld.org/2025/china-cloud-middle-east. 55 Garam, L. (2025), ‘“100조 쏟아부어 AI강국 만들겠다”…이재명 시대 ‘AI 로드맵’ 살펴보니 [‘“I will pour a hundred trillion won to create an AI powerhouse” .. Looking at the ‘AI Roadmap’ in the era of Lee Jae-myung]’, MK, 4 June 2025, https://www.mk.co.kr/news/business/11334520. 56 Kumar, R. and Cho, Y. (2025), ‘South Korea’s Sovereign AI Gambit: A High-Stakes Experiment in Autonomy’, The Diplomat, 4 July 2025, https://thediplomat.com/2025/07/south-koreas-sovereign-ai-gambit-a-high-stakes- experiment-in-autonomy. 57 European Commission (2025), ‘The AI Continent Action Plan’. 58 AI Sweden (2024), ‘AI Sweden and Fraunhofer IAIS to develop language models for all of Europe’, 16 May 2024, https://www.ai.se/en/news/ai-sweden-and-fraunhofer-iais-develop-language-models-all-europe. 59 OpenEuroLLM (undated), ‘A series of foundation models for transparent AI in Europe’, https://openeurollm.eu. -- 16 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 15 Chatham House Singapore’s Southeast Asian Languages in One Network (SEA-LION) is a multilingual AI model that aims to provide LLMs with a better understanding of ‘diverse contexts, languages and cultures’.60 Such cases underline how regional or international blocs can collectively build sovereign AI. At the country level, there are also efforts to uphold cultural and linguistic diversity in sovereign AI initiatives, from Sweden’s GPT-SW361 to Spain’s Alia62 (available in Catalan, Valencian, Galician and Basque, in addition to other European languages). Switzerland’s Apertus initiative63 provides a fully open, public model designed to maximize auditability and multilingual support while avoiding reliance on foreign commercial providers. These efforts correlate with rising concerns about the deployment of models from the US and China with ‘non-European’ values. China’s DeepSeek, for example, is facing bans in German app stores due to data protection concerns.64 Both NVIDIA and OpenAI promote international partnerships based on values alignment. NVIDIA touted localization and respect for national values as the basis of its latest series of agreements and partnerships in Asia. Likewise, as part of Stargate, the OpenAI for Countries initiative promotes ‘build[ing] on democratic AI rails’.65 That said, its first official partner, the UAE, is a federation of monarchies. Non-democracies are motivated by value alignment, too, with prevailing concerns about content control over the outputs of foreign-developed AI models and the subsequent implications for regime stability and public order. Consequently, ChatGPT does not operate in China. Sovereign AI capabilities can be repurposed for digital authoritarian ends, giving governments the tools to automate repression and surveillance. Indonesia’s sovereign AI model (Sahabat-AI) caters to local languages in an effort to align AI with national values and facilitate local applications. The strategies reviewed for this paper capture attempts to align the development of sovereign AI capabilities with national values. These values range from interpretations of public good and societal benefit to ideas about political organization, rights and freedoms. 60 Sea-lion AI (undated), ‘Catalysing AI Innovation for Southeast Asia’, https://sea-lion.ai. 61 AI Sweden (undated), ‘GPT-SW3’, https://www.ai.se/en/project/gpt-sw3. 62 Barcelona Supercomputing Center (2025), ‘ALIA, Europe’s first public, open and multilingual AI infrastructure’, BSC News, 21 January 2025, https://www.bsc.es/news/bsc-news/alia-europes-first-public-open-and-multilingual- ai-infrastructure. 63 Apertus (undated), ‘About Apertus’, https://www.swiss-ai.org/apertus. 64 Ersen, H. and Murray, M. (2025), ‘DeepSeek faces ban from Apple, Google app stores in Germany’, Reuters, 27 June 2025, https://www.reuters.com/sustainability/boards-policy-regulation/deepseek-faces-expulsion- app-stores-germany-2025-06-27. 65 OpenAI (2025), ‘Introducing OpenAI for Countries’, https://openai.com/global-affairs/openai-for-countries. Country-level efforts correlate with rising concerns about the deployment of models from the US and China with ‘non-European’ values. -- 17 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 16 Chatham House Domination of the frontier Only the US and China seek to dominate the frontier, or most advanced, AI tools. For these two countries, building domestic AI capabilities is about readiness for – and protection from – future technological disruption, triggered by exceptionally advanced AI capabilities or artificial general intelligence (AGI). The term AGI, regardless of its likelihood or achievability, refers to machine cognition that matches human-level reasoning. It is argued that such a development would fundamentally restructure global power dynamics, including economies and military capabilities. Research and public documentation suggest China has been working towards AGI since 2017,66 although the country’s policy priority appears to be accelerating the diffusion of less advanced AI for everyday uses. Leading US AI companies also profess their dreams to attain AGI capabilities and, in their words, ‘transform the world for the better’.67 For leading AI powers, the pursuit of frontier capabilities and AGI functions as a tool for shaping global norms, securing first-mover advantages and influencing international standards. Investments in frontier AI are not just about technological advancements, these projects carry strategic intent, underpin economic competitiveness and create leverage in scientific collaboration and talent attraction. Are middle powers motivated to attain frontier or AGI capabilities? It is certainly in their strategic interest to build readiness for – or even stewardship over – future technological disruptions. However, joining the global race towards AGI might not be. French AI company Mistral, for example, criticises the ‘religious’ obsession with AGI,68 while simultaneously promoting itself as a frontier AI company. France’s strategic approach to sovereign AI capabilities prioritizes making the country an AI ‘powerhouse’ and R&D for pioneering innovation in a limited number of scientific and technological fields with direct societal benefits, like improved healthcare.69 Other middle powers engage with frontier AI selectively, often concentrating on specific subcategories with clear societal, economic or industrial value. This sectoral approach allows middle powers to maintain awareness of cutting-edge 66 Hannas, W., Chang, H-M., Riesenhuber, M. and Chou, D. (2023), China’s Cognitive AI Research: Emulating Human Cognition on the Way to General Purpose AI, policy brief, Washington, DC: Center for Security and Emerging Technology, https://cset.georgetown.edu/publication/chinas-cognitive-ai-research. 67 Amodei, D. (2024), ‘Machines of Loving Grace: How AI Could Transform the World for the Better’, Dario Amodei Blog, October 2024, https://www.darioamodei.com/essay/machines-of-loving-grace. 68 Al-Sibai, N. (2024), ‘Mistral CEO Says AI Companies Are Trying to Build God’, https://futurism.com/the-byte/ mistral-ceo-agi-god. 69 President of the Republic of France (2025), ‘Make France an AI Powerhouse’, February 2025, https://www.elysee.fr/admin/upload/default/0001/17/d9c1462e7337d353f918aac7d654b896b77c5349.pdf. For leading AI powers, the pursuit of frontier capabilities and AGI functions as a tool for shaping global norms, securing first-mover advantages and influencing international standards. -- 18 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 17 Chatham House developments, contribute to scientific knowledge and support domestic innovation ecosystems, without the resource burden of competing directly with superpowers in the AGI race. Engagement with frontier AI also provides middle powers with strategic flexibility. By cultivating technical expertise and research capacity in select areas, these states can better understand the capabilities and risks of emerging AI technologies, anticipate potential disruptions and align investments with national priorities. Such selective engagement reflects a broader pattern in the global AI landscape, where countries balance ambition, capacity and practical benefits. While superpowers pursue comprehensive AGI and frontier strategies for global influence, middle powers – given their constraints – are more likely to focus on readiness, selective leadership in niche domains and ensuring that emerging capabilities can be harnessed for societal, economic or industrial benefit. This distinction highlights the diversity of motivations driving sovereign AI development and the different pathways that nations take to participate in the global AI ecosystem. -- 19 of 48 -- 18 Chatham House 03 Strategy meets reality The US and China lead the global AI race, leaving middle powers far behind. Despite the ambitions of national sovereign AI strategies, global dependencies on US and Chinese technology will continue to run deep. The US and China are leading the global AI race by a significant margin. However, it is the US that has achieved overwhelming computational dominance. By March 2025, the US accounted for 75 per cent of global AI supercomputer performance versus China’s 15 per cent.70 Traditional technology powers like Germany, Japan and the UK have become increasingly relegated to a subsidiary role, in a dramatic reshuffling of technological influence. Middle powers cannot match US or Chinese national AI investment. But, at the same time, middle powers intensely fear being excluded from emerging AI developments. Both the US and China have championed major bilateral partnerships, promising to supply their allies and customers with the infrastructures and ecosystems required for competitive AI. Yet, facing the growing margin between AI ‘leaders’ and AI ‘followers’ and the risks of a global split between US and Chinese offerings, middle powers and their international blocs are proactively seeking a more sovereign, secure path forward. Should AI become the foundation of economic growth, national security and public governance, dependence on either US or Chinese companies carries strategic risk. If middle powers get this negotiation wrong, they risk being locked into systems that they neither own nor control, ceding influence over the foundational technological layers of their politics, economies and societies. If middle powers get it right, they can build selective autonomy: leveraging partnerships, shared infrastructure, data and 70 Pilz, K. F. et al. (2025), ‘The US hosts the majority of GPU cluster performance, followed by China’, Epoch AI, 5 June 2025, https://epoch.ai/data-insights/ai-supercomputers-performance-share-by-country. -- 20 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 19 Chatham House governance leadership to secure a measure of agency in an interdependent AI order. The task is not to match the superpowers’ scale, but to safeguard the capacity to choose, adapt and steer the application of AI. Dependencies Most countries’ domestic AI industries are dependent on models developed outside their borders. While many have released smaller models or built on larger models to reflect local and national languages, the most powerful, scalable and readily accessible ‘foundation’ models (i.e. general-purpose, advanced models suitable for multiple uses and adaptation) are still provided by US and Chinese companies. The US dominates access to cutting-edge proprietary models and chatbots (such as Google’s Gemini, Microsoft’s Copilot or OpenAI’s ChatGPT-5), while China is emerging as a global leader in advanced open-source alternatives. The global proliferation of China’s open reasoning models71 like DeepSeek’s R3, Alibaba’s Qwen and Zhipu AI’s GLM-4.5 might democratize access to powerful capabilities and lower technical barriers, but they do not alter the fact that middle powers are dependent on others. Whether proprietary or open, frontier AI models remain concentrated in the hands of the two global AI superpowers, leaving others structurally dependent on external technology. Middle powers are also dependent on foreign-provided infrastructure for computing power.72 As AI models become more advanced, their demands for computing power skyrocket. US-based hyperscalers73 – such as NVIDIA, Oracle and Microsoft – remain the dominant forces in enabling Western AI adoption. Globally, US and Chinese providers dominate the cloud market, with US players (Microsoft, Google and Amazon Web Services) operating two-thirds of the market.74 71 A recent wave of LLMs have so-called reasoning capabilities. This means that models are prompted to solve problems step-by-step, generating a ‘chain of thought’ before delivering a response. For further information about the development of DeepSeek R1’s reasoning capabilities, see: Huggingface (undated), ‘DeepSeek-R1’, https://huggingface.co/deepseek-ai/DeepSeek-R1. 72 Also known as ‘compute’, which refers to how powerful and efficient AI’s physical and technical infrastructure is, such as data centres. 73 Companies that provide major data centres and cloud infrastructure for others to use, capable of rapidly scaling up to meet growing demand for compute. 74 Statista (2025), ‘Revenue of the public cloud market in the United States from 2020 to 2029’, Statista, 14 May 2025, https://www.statista.com/forecasts/1309968/us-revenue-from-public-cloud; Richter, F. (2025), ‘AWS Stays Ahead as Cloud Market Accelerates’, Statista, 4 November 2025, https://www.statista.com/chart/18819/ worldwide-market-share-of-leading-cloud-infrastructure-service-providers. The US dominates access to cutting-edge proprietary models and chatbots (such as Google’s Gemini, Microsoft’s Copilot or OpenAI’s ChatGPT-5), while China is emerging as a global leader in advanced open-source alternatives. -- 21 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 20 Chatham House Dependence also extends to data, talent and intellectual property. While governments have access to valuable public datasets, these are far too limited in scale and diversity to train competitive general-purpose models. The richest sources of training data – such as data generated through reinforcement learning from human feedback (RLHF)75 – remain concentrated in private, US and Chinese firms. Meanwhile, advanced AI talent is scarce and costly, and core intellectual property is controlled by major US and Chinese companies, in part through non-disclosure agreements with staff to prevent corporate espionage – though this has silenced critics and brought about self-censorship.76 Together, these gaps constrain the ability of middle powers to build, adapt and commercialize sovereign AI systems. This chapter highlights components essential for building AI sovereignty and tracks the constraints and availability of each in relation to dependencies on the US and China. It distinguishes between primary building blocks that provide a technical foundation and secondary blocks that constitute the broader enabling ecosystem. Table 2. Primary and secondary ‘building blocks’ for sovereign AI Inputs Description Primary Data The volume, quality, diversity and accessibility of datasets required to train, fine-tune and operate AI systems effectively. Compute Access to high-performance computing hardware, particularly advanced GPUs and TPUs,77 necessary for training and running AI models. Frontier models The availability of state-of-the-art AI models, either through domestic development, partnerships or open-source access. Secondary Energy Reliable, abundant and preferably low-cost energy supply to power AI infrastructure and data centres. Infrastructure The physical and digital infrastructure enabling AI deployment, including data centres, networks and distribution systems. Industry The domestic ecosystem of companies, universities and institutions capable of developing, deploying and maintaining AI systems. Talent The human capital and research ecosystems capable of designing, developing, deploying and governing AI systems, including researchers, engineers and policy experts. Trust The domestic capacity and willingness to adopt and effectively utilize AI systems across government, business and society. 75 Reinforcement learning from human feedback (RLHF) is a pre-deployment training technique in which humans assess and label model behaviours or outputs, and these evaluations are then used to guide further optimization of the model. 76 Shivakumar, N. and Bhattacharjee, S. S. (2025), ‘How NDAs Became the AI Industry’s Tool for Surveillance and Silence’, Tech Policy Press, 20 June 2025, https://www.techpolicy.press/how-ndas-became-the-ai-industrys-tool- for-surveillance-and-silence. 77 GPUs (graphics processing units) and TPUs (tensor processing units) are essential for the training of neural networks and the functioning of AI models, as they can handle complexity and parallel processing. -- 22 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 21 Chatham House Primary building blocks Data Middle powers cannot match the scale at which the US or China are able to gather data to train AI models. However, curated, high-quality government and local data can become strategic assets in sovereign AI development. Data are the foundational input of AI capability. Models are only as good as the data they are trained on. Access to good datasets directly determines model performance and applicability. Sovereign technology strategies often prioritize data collection, organization, availability and security.78 The US and China retain an enormous advantage through global user data – and closed datasets – generated by commercial providers’ digital services (for example, ByteDance and Google). China has long identified data collection as essential for global technology leadership, with a large domestic user base generating valuable big data through public services delivered by digital technology and interoperable e-commerce platforms, in addition to data-hungry online services outside its borders.79 Middle powers’ relative lack of homegrown digital services, the reduced capability for public data collection and generally limited progress in training models on government data have together created a capability gap. Building AI-ready government data is a critical step in developing sovereign AI capabilities – for example, through the UK’s National Data Library80 and Singapore’s data initiative.81 Large volumes of high-quality sovereign data will be necessary for some domestic AI applications. How those data are collected is a valuable part of the global AI supply chain. However, middle powers cannot catch up with the US and China in terms of quantity and quality of AI-ready training data, nor can these smaller countries be wholly self-sufficient in terms of AI. 78 UK DSIT (2025), ‘AI Opportunities Action Plan’; von der Leyen, U. (2024), Europe’s Choice: Political Guidelines for the Next European Commission 2024-2029, European Commission, 18 July 2024, https://commission.europa. eu/document/download/e6cd4328-673c-4e7a-8683-f63ffb2cf648_en?filename=Political%20Guidelines%20 2024-2029_EN.pdf; Canada ISED (2024), ‘What We Heard Report: Consultations on AI Compute’, 22 November 2024, https://ised-isde.canada.ca/site/ised/en/what-we-heard-report-consultations-ai-compute. 79 Dawson, J. and Wheeler, T. (2022), ‘How to tackle the data collection behind China’s AI ambitions’, Brookings, 29 April 2022, https://www.brookings.edu/articles/how-to-tackle-the-data-collection-behind-chinas-ai-ambitions. 80 See: Worth, S., Gurumoorthy, A. and Simperl, E. (2025), ‘Developing the UK National Data Library for public benefit: 10 key reflections’, Kings College London, https://www.kcl.ac.uk/developing-the-uk-national-data- library-for-public-benefit-10-key-reflections. 81 See: Singapore Government Developer Portal (undated), ‘Data.gov.sg’, https://v2.developer.tech.gov.sg/ products/categories/data-and-apis/data-gov-sg/overview. Middle powers’ relative lack of homegrown digital services, the reduced capability for public data collection and generally limited progress in training models on government data have together created a capability gap. -- 23 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 22 Chatham House Compute Compute scarcity is the central bottleneck for AI sovereignty. To secure the necessary compute or alternatively develop more resource-efficient AI models, middle powers must specialize in supply-chain niches or form partnerships. Physical computational infrastructure constrains AI sovereignty. Compute determines model size, training speed and the scale of deployment. Without sufficient compute, even excellent data and talent cannot produce advanced AI capabilities. Computing infrastructure is expensive, which constrains the progress of middle powers. The OECD estimates that just 13 member countries (advanced economies) have enough public cloud compute – understood as ‘on-demand [online] services from commercial providers available to the general public’ – for both developing and deploying advanced AI systems. 82 Meanwhile, the US and China have clear advantages, ‘creating a new geopolitical stratification where only [they] enjoy full-stack supply chain control’. 83 Investments like the US Stargate project ($500 billion) and Chinese providers’ projected $70 billion AI investment in 202684 – not including continued Chinese state investments – dwarf middle power efforts, such as those recently announced by South Korea ($75 billion),85 the UK ($19 billion)86 and Saudi Arabia ($20 billion).87 However, some middle powers have carved out space in the AI supply chain through excellence in narrow but vital inputs. Dutch leadership, through ASML, in the production of machinery essential to the manufacture of high-performance semiconductors is a case in point. 88 Furthermore, Canada’s sovereign AI compute strategy prioritizes building domestic compute capacity and measures like giving domestic AI institutes and firms access to limited compute.89 Following this approach, middle powers may develop sufficiently specialized physical capabilities as part of the AI supply chain, which may enable limited investments in domestic compute. Nevertheless, the US – and US companies – will race ahead. Their dominance as the global gatekeepers for access to hardware will limit wider efforts to build AI sovereignty. 82 Lehdonvirta, V. et al. (2025), ‘The geography of AI compute: Mapping what is available and where’, OECD AI Policy Observatory, 29 October 2025, https://oecd.ai/en/wonk/the-geopgraphy-of-ai-compute-mapping- what-is-available-and-where. 83 Rondeaux, C. (2025), ‘Compute or be computed’, New America, 20 May 2025, https://www.newamerica.org/ planetary-politics/blog/compute-or-be-computed. 84 Goldman Sachs (2025), ‘China’s AI Providers Expected to Invest $70 Billion in Data Centers Amid Overseas Expansion’, 3 November 2025, https://www.goldmansachs.com/insights/articles/chinas-ai-providers-expected- to-invest-70-billion-dollars-in-data-centers-amid-overseas-expansion. 85 Garam (2025), ‘“100조 쏟아부어 AI강국 만들겠다”…이재명 시대 ‘AI 로드맵’ 살펴보니 [‘“I will pour a hundred trillion won to create an AI powerhouse” .. Looking at the ‘AI Roadmap’ in the era of Lee Jae-myung]’. 86 UK DSIT (2025), ‘AI Opportunities Action Plan’; Adams (2025), ‘UK unveils £14b AI infrastructure plan’. 87 Saudi Data and AI Authority (undated), ‘National Strategy for Data & AI’. 88 Tarasov, K. (2022), ‘ASML is the only company making the $200 million machines needed to print every advanced microchip. Here’s an inside look’, CNBC, 23 March 2022, https://www.cnbc.com/2022/03/23/ inside-asml-the-company-advanced-chipmakers-use-for-euv-lithography.html. 89 Canada ISED (2024), ‘Sovereign Compute Strategy’, https://ised-isde.canada.ca/site/ised/en/canadian- sovereign-ai-compute-strategy. -- 24 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 23 Chatham House Frontier models Access to powerful models – proprietary or open-source – remains concentrated in the US and China. Middle powers can only mitigate, not eliminate, this dependence through strategic collaboration. States need access to frontier models to develop downstream applications and services. Based on current trajectories, these advanced AI models will cost an average of $1 billion to train by 2027.90 This is contingent on compute – for example, the procurement of powerful GPUs and the construction of data centres – and operational costs associated with model training itself. For most countries, the investment required will continue to be a barrier to training powerful models. Most middle power national AI policies emphasize the importance of strategic partnerships for guaranteeing access to frontier AI models. Nonetheless, these governments continue to magnify or even exaggerate the role of their country. For example, as part of its aim to turn the UK into ‘an AI maker not an AI taker’, the government has called for UK ‘national champions’ at critical layers of the global AI supply chain,91 and President Emmanuel Macron promotes Mistral as the French frontier challenger to US and Chinese AI models.92 DeepSeek’s release in January 2025 was a game-changer, showing middle powers – with limited primary inputs (such as data and compute) – that less costly open-source alternatives to major proprietary models might be a feasible pathway to sovereignty. International efforts to build AI models cooperatively – such as the ‘Airbus for AI’ initiative93 – are promising, as are alternative technology paradigms backed by countries like India, which could lead to the deployment and availability of powerful AI models with digital public infrastructure.94 Domestic models specific to domains and industries – like Saudi Aramco’s METABRAIN, which is designed for use in and around the oil industry – may in the future emerge as the best opportunity for middle power sovereign AI development.95 There is a growing recognition of the risks of total dependence on US and Chinese AI providers by middle powers and enthusiastic exploration of sovereign alternatives. 90 Cottier, B. et al. (2024), ‘How Much Does It Cost to Train Frontier AI Models?’, Epoch AI, 3 June 2024, updated 13 January 2025, https://epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models. 91 UK DSIT (2025), ‘AI Opportunities Action Plan’. 92 Dillet, R. (2025), ‘Macron gets down to business’, Tech Crunch, 10 February 2025, https://techcrunch.com/ 2025/02/10/mistral-gets-down-to-business. 93 See: Tan, J., Jackson, B., Berjon, R. and Coyle, D. (2025), Airbus for AI: A global strategy for public value creation, Bennett School for Public Policy, University of Cambridge, https://publicai.co/airbus-for-ai.pdf; the airbus for AI initiative gains its inspiration from Airbus, which started as a multilateral industrial consortium for European countries and businesses to better compete with Boeing. 94 Wilkinson, R., Krasodomski, A. and Wilkinson, I. (2025), The case for expanding digital public infrastructure: How open, scalable technology can serve citizens, preserve sovereignty and save money, Research Paper, London: Royal Institute of International Affairs, https://doi.org/10.55317/9781784136604. 95 Khowaiter, A. O. (2024), ‘Remarks by Mr. Ahmad O. Khowaiter at the Global AI Summit 2024’, Saudi Aramco, 10 September 2024, https://www.aramco.com/en/news-media/speeches/2024/remark-by-mr-ahmad- khowaiter-at-the-global-ai-summit-2024. Based on current trajectories, advanced AI models will cost an average of $1 billion to train by 2027. -- 25 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 24 Chatham House The global distribution of access to narrow AI systems may also be highly unstable and subject to political and regulatory changes. Faced with the enormous energy, compute and talent costs of frontier and general-purpose AI development, in the short term, it is likely most states will be forced to opt for existing model providers. In the longer term, there are indications that a third way is possible, either through significant national or regional investment, through sufficiently well-resourced coalitions, or by focusing on less powerful models to meet local use cases. Box 1. Open-source AI Open-source and open-science approaches are increasingly important pathways to AI autonomy. They illustrate how public or consortium-based models can support national languages and values, comply with domestic regulatory frameworks and avoid vendor lock-in.96 These efforts show that sovereign AI does not rely solely on cultivating domestic technology, open-source ecosystems can provide flexible, adaptable and values-aligned alternatives for governments seeking greater control over AI. But an open-source approach is not a complete solution. The European Open-Source AI Index reveals significant differences in how ‘open’ various AI models actually are – measured by factors like availability, access and documentation quality. For instance, HuggingFace’s Smo1LM scores highly across all openness metrics, while Cohere AI’s Command A scores poorly in areas like hardware architecture documentation and base model data availability.97 Furthermore, even when companies and governments use open models, they typically fine-tune them for specific purposes while remaining dependent on larger technology companies for the underlying software infrastructure. Open-source AI is also a significant component of US and Chinese global dominance of AI. China’s Global AI Governance Action Plan, for example, is predicated on the global diffusion of Chinese open-source models. Developing and diffusing open-source AI globally is also a core part of the US AI strategy.98 For middle powers seeking to build AI in the public interest, reliance on foreign open-source models is significantly better than reliance on foreign proprietary models. Many national-interest AI efforts are the product of cross-pollination and mutual learning, made possible by increased openness and reporting around, for instance, model weights – the learned parameters of a trained model that determine the model’s behaviour – and training data. But dependence on open-source models, like DeepSeek’s reasoning models and Meta’s Llama, may still present a dependence issue for middle powers. 96 Vendor lock-in takes place when an organization becomes structurally dependent on a single AI provider to the extent that switching away becomes prohibitively expensive or technically impossible, even if better options exist. 97 European Open-Source Index (2026), ‘EU-based community-driven research on open-source generative AI systems’, https://osai-index.eu. 98 Heikkilä, M. (2025), ‘China leapfrogs US in global market for ‘open’ AI models’, Financial Times, 26 November 2025, https://www.ft.com/content/931c8218-a9d7-4cbd-8b08-27516637ff41. -- 26 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 25 Chatham House Secondary building blocks Energy Cheap, reliable energy is emerging as a critical enabler of AI sovereignty, creating new advantages for energy-rich middle powers. Operating AI is extremely energy intensive, with the power required to train frontier AI models doubling annually.99 As such, the cost of energy can make or break commercial AI deployment. As in other fields, energy security affects strategic autonomy. Both the US and China have made enormous commitments to expand energy production. In renewables, China has seen unprecedented growth in its energy sector: the country reported 887 GW of total installed solar power in 2024, up from 609 GW in the previous year.100 While the US lags in solar energy production, despite growing demand, 101 the country has resorted to the use of fossil fuels to meet its energy demands under President Donald Trump.102 Middle powers are similarly divided in their approaches to meeting AI’s energy requirements. France is ramping up its energy production, building Europe’s largest AI campus with Abu Dhabi’s state-owned technology investment fund (MGX) and US chipmaker NVIDIA.103 Others, like Japan with its ‘Watt-Bit Collaboration’ framework,104 are more cautious, opting for maximizing efficiencies in existing infrastructure or reprioritizing energy use, focusing on reforms to planning and regulation to incentivize the construction of sustainable data centres. Other middle powers can use existing access to energy or their climates as a bargaining chip. The competitive edge of the Gulf States in particular is founded on access to cheaper fossil fuel energy, while other countries, such as Sweden, Canada and Finland, have cooler climates that can reduce the energy demand of data centres. A report from the national energy regulator notes that Canada’s cool climate, clean energy and relatively low electricity costs make it an attractive destination for data centres.105 99 Emberson, L. and Rahman, R. (2024), ‘The power required to train frontier AI models is doubling annually’, Epoch AI, 19 September 2024, https://epoch.ai/data-insights/power-usage-trend. 100 Ener data (2025), ‘China installs record capacity for solar (+45%) and wind (+18%) in 2024’, 22 January 2025, https://www.enerdata.net/publications/daily-energy-news/china-installs-record-capacity-solar-45- and-wind-18-2024.html. 101 Goldman Sachs Research (2024), ‘GS SUSTAIN: Generational Growth: AI, data centers and the coming US power demand surge’, 29 April 2024, https://www.goldmansachs.com/insights/goldman-sachs-research/ generational-growth-ai-data-centers-and-the-coming-us-power-demand-surge. 102 Reuters (2025), ‘Trump’s budget bill boosts fossil fuels, hits renewable energy’, 2 July 2025, https://www.reuters.com/sustainability/climate-energy/trumps-budget-bill-boosts-fossil-fuels-hits-renewable- energy-2025-07-02. 103 Semafor Gulf (2025), ‘Abu Dhabi’s MGX, Nvidia to build Europe’s largest AI campus’, Semafor, 21 May 2025, https://www.semafor.com/article/05/21/2025/abu-dhabis-mgx-nvidia-to-build-europes-largest-ai-campus-in-paris. 104 This Japanese policy aims to jointly plan energy systems and digital infrastructure, focusing on regulatory reform, energy efficiency and grid coordination to support sustainable growth in data centres and AI compute. 105 Canada Energy Regulator (2024), ‘Market Snapshot: Energy demand from data centers is steadily increasing, and AI development is a significant factor’, 2 October 2024, https://www.cer-rec.gc.ca/en/data-analysis/ energy-markets/market-snapshots/2024/market-snapshot-energy-demand-from-data-centers-is-steadily- increasing-and-ai-development-is-a-significant-factor.html. -- 27 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 26 Chatham House By 2030, it is plausible that energy will be the primary bottleneck for AI development and deployment, limiting even AI superpowers like the US. Of all the ingredients required for a thriving sovereign AI ecosystem, energy may give middle powers the highest potential leverage in developing sovereign AI systems. Infrastructure Domestic data centre and connectivity investments offer middle powers partial sovereignty gains, even if built with overseas components or by foreign firms. AI infrastructure determines how effectively the technology can be developed, deployed and accessed across society. Without sufficient core hardware and software as well as networking infrastructure, the potential economic benefits of developing AI capabilities will fail to materialize. Strong infrastructure also matters for a state’s strategic autonomy over how these benefits are captured and maximized. The US has a huge advantage in global cloud infrastructure, with its cloud computing market set to generate $467 billion in annual revenue by 2028 and key players – Microsoft, Google and Amazon Web Services (AWS) – accounting for close to two-thirds of that number.106 US investment in data centres is accelerating, with flagship projects like Stargate. China’s Informatization Plan – the world’s first ‘industrial policy for the digital age’ – prioritizes the development of an integrated national digital ecosystem supported by advances in digital infrastructure.107 While China leads on the number of data centres (over 400) in the Asia-Pacific region, the country trails far behind the number of data centres in the US (with well over 5,000). 108 Middle powers do not have the resources to match these investments. Most nations’ cloud computing in particular – despite efforts to build alternatives like the troubled GaiaX in Europe – remain almost entirely dependent on US providers. Even the UAE is dependent on strategic international partnerships with the US, NVIDIA 106 Statista (2026), ‘Revenue of the public cloud market in the United States from 2018 to 2030’, Statista, 20 January 2026, https://www.statista.com/forecasts/1309968/us-revenue-from-public-cloud; Richter (2025), ‘AWS Stays Ahead as Cloud Market Accelerates’. 107 Creemers, R. and Triolo, P. (2022), ‘Analyzing China’s 2021–2025 Informatization Plan: A DigiChina Forum’, Stanford Digi China, 24 January 2022, https://digichina.stanford.edu/work/analyzing-chinas-2021- 2025-informatization-plan-a-digichina-forum. 108 Taylor, P. (2025), ‘Number of data centers worldwide as of November 2025, by country or territory’, Statista, 19 November 2025, https://www.statista.com/statistics/1228433/data-centers-worldwide-by-country. The US has a huge advantage in global cloud infrastructure, with its cloud computing market set to generate $467 billion in annual revenue by 2028 and key players – Microsoft, Google and Amazon Web Services – accounting for close to two-thirds of that number. -- 28 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 27 Chatham House and OpenAI to build Stargate UAE, soon to be the world’s largest data centre. The EU’s €20 billion investment in computing infrastructure through a network of ‘AI Factories’ 109 – approximately 5 per cent of the Stargate initiative’s $500 billion – will be supported by preferential procurement, though current dependencies on US cloud providers are set to continue.110 Over the medium term, the gap between supply and demand is set to grow. Middle powers will struggle to meet infrastructural demands without leveraging network benefits like resource-sharing and regional strategic partnerships. Industry Cultivating both a resilient domestic AI industry and an ecosystem to support small and medium-sized enterprises could help middle powers use foreign technologies within national value chains. Robust domestic industrial capacity provides the institutional foundations and route to market for sustained AI growth, capability, economic benefits and strategic autonomy. These critical industry features may secure national AI infrastructure and even support market and technological development. However, the effectiveness of a sovereign AI strategy depends on the ability to deploy industrial systems at scale. China and the US boast clear advantages when it comes to their domestic AI industries. But their industrial bases differ significantly. While China’s AI industry is crowded with start-ups and established technology companies branching out into AI models and services, the US AI industry is globally unparalleled in terms of finance. US private investment into AI was $109.1 billion in 2024, majorly outpacing China (at $9.3 billion) and the UK ($4.5 billion).111 This is the result of a thriving venture capital ecosystem that feeds into start-ups and major technology giants and their rapidly developing AI capabilities. Both the US and China boast strong links between industry research labs and universities, both domestic and foreign. For example, US-based OpenAI offers access to its models and ChatGPT Edu – designed specifically for university settings – to leading UK universities.112 Meanwhile, local governments in China, such as the Shenzhen authorities, offer foreign and domestic AI professionals free housing, rent reduction and other subsidies to incentivize new tech hubs.113 In universities, Chinese students are offered extra credits, and faculties extra compensation, for contributing to the open-source AI ecosystem.114 109 The European Commission defines AI factories as ‘dynamic ecosystems’ that bring together primary building blocks – compute and data – with secondary blocks like talent to train cutting-edge models and drive collaboration across Europe, particularly in connecting its supercomputing centres to businesses, industry and universities. European Commission (undated), ‘AI Factories’, https://digital-strategy.ec.europa.eu/en/policies/ai-factories. 110 European Commission (2025), ‘The AI Continent Action Plan’. 111 Stanford University HAI (2025), ‘2025 AI Index’, https://hai.stanford.edu/ai-index/2025-ai-index-report/economy. 112 University of Oxford (2025), ‘Oxford and OpenAI launch collaboration to advance research and education’, press release, 4 March 2025, https://www.ox.ac.uk/news/2025-03-04-oxford-and-openai-launch-collaboration- advance-research-and-education. 113 Vaughan, S. (2025), ‘From Vouchers to Vidas: China’s Innovative Plan for AI Dominance’, Foreign Policy research Institute, 10 September 2025, https://www.fpri.org/article/2025/09/from-vouchers-to-visas-chinas- innovative-plan-for-ai-dominance. 114 Schaefer, K. and Nunlist, T. (2025), ‘The AI Plus initiative – China’s blueprint for AI diffusion’, Trivium, 4 September 2025, https://triviumchina.com/research/the-ai-plus-initiative-chinas-blueprint-for-ai-diffusion. -- 29 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 28 Chatham House Middle powers have long recognized the strategic importance of developing a robust industrial base. This now extends to the emerging AI industry. As early as 2018, the French Institute for Research in Computer Science and Automation was tasked with accelerating academic spin-offs – companies formed to commercialize technological developments by university researchers – and developing cooperation programmes with industry.115 At the EU level, the Continent Action Plan commits to large-scale computing infrastructure to support start-ups and industry to become more competitive with frontier AI models and applications.116 But middle powers must also navigate dependencies on non-homegrown providers of data and compute. Singapore, for instance, focuses on AI talent development and specialized applications in finance and urban planning, with a view to maintaining sovereignty despite limited resources. The country partners with AWS to help scale AI across business operations, for example on ‘cloud credits’.117 Middle powers, particularly those in Europe, are frequently identified as either deterring or failing to mobilize untapped capital resources. To address this shortcoming, these states should seek to unlock greater investment risk tolerance, particularly in relation to homegrown talent and AI developers, alongside an enhanced regulatory environment for venture capital and fundraising for start-ups.118 This potential cultural shift away from the AI superpowers might be seen as an essential hedge against potential risks, for example, in the event of an AI bubble bursting. It will take years to gauge whether the EU’s investment in an AI industrial base will boost independence. Regardless of the outcome, multinational-level industrial policy is a step in the right direction towards sovereignty. Industrial partnerships are a viable pathway for capability development in universities and the start-up ecosystem: particularly those involving US-based companies, although China-based companies may soon provide a competing offer. Talent Talent mobility and retention are critical considerations for AI sovereign strategies. International partnerships, visas and education are important policy tools to address these issues. AI development is enabled by human capital. Even with sufficiently skilled personnel at the AI development stage, countries often struggle to benefit from new capabilities due to a lack of talent available to deploy and govern AI. Research ecosystems – 115 European Commission (2021), ‘France AI Strategy Report’, European Commission, 1 September 2021, https://ai-watch.ec.europa.eu/countries/france/france-ai-strategy-report_en. 116 European Commission (2025), ‘The AI Continent Action Plan’. 117 Cloud credits are pre-paid allowances for the use of cloud services. Singapore Economic Development Board (EDB) (2025), ‘AWS and DISG Launch AI Springboard in celebration of SG60 to support 300 Singapore-based enterprises build AI Centers of Excellence’, Singapore Economic Development Board, 25 June 2025, https://www.edb.gov.sg/en/about-edb/media-releases-publications/aws-disg-launch-ai-springboard-to- support-300-firms.html. 118 Marcus, J. S. (2024), ‘Europe can produce its own tech giants – here’s how’, CEPS, 8 July 2024, https://www.ceps.eu/europe-can-produce-its-own-tech-giants-heres-how. -- 30 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 29 Chatham House ranging from investments in academic labs to technology transfer arrangements (such as licensing) – are equally important for producing and scaling AI outputs from human capital. The US is a hub for science, technology, engineering and maths, with leading universities capable of attracting and curating talent. Its AI industry boasts world-leading financial remuneration, strong pathways to market (such as through technology transfer arrangements) and industry–academic collaboration (like joint research, training and data or compute access). That said, sector layoffs and tightening visa policies are currently endangering the US’s foreign AI talent base.119 China, on the other hand, is taking steps in the opposite direction, with state investments in AI innovation centres and labs, and financial incentivizes for talented Chinese citizens returning from having studied or worked abroad. 120 Chinese companies are also making major strides towards training – and retaining – the next generation of AI professionals.121 Middle powers once again lag behind the AI superpowers in terms of human capital. China’s access to homegrown talent and US levels of financial rewards are unrealistic ambitions for smaller countries. While some middle powers can attract global talent through world-class universities, competitive degree programmes, visa schemes and academic links with industry, much of this potential is currently squandered by failures in talent retention. As a case in point, the UK’s research ecosystem incubated and nurtured the AI company Deepmind, allowing it to grow, but the country could not keep it. The company was acquired by Google for around $500 million in 2014.122 This precipitated a steady stream of talented AI professionals moving to the US and, within the UK, away from academia, due in significant part to earning potential.123 In seeking to make domestic research ecosystems more profitable – and thus retain leading human capital – middle powers are increasingly proactive. An over-emphasis on research – which can negatively impact effective deployment – may be a short-term trap for AI development. Instead, working towards the creation of preferred national and commercial AI standards might be a way of boosting domestic AI development.124 In addition, the UK has committed to work across the talent pipeline: for instance, by launching a scholarship (with investment from industry), investing in elite headhunting and exploring immigration policy changes. Just one UK-founded AI frontier lab could make all the difference to the country’s prospects of sovereign AI.125 Strategic partnerships with powerful AI companies are for the most part 119 Knight, W. (2025), ‘Trump’s Crackdown on Foreign Student Visas Could Derail Critical AI Research’, WIRED, 29 May 2025, https://www.wired.com/story/trump-administration-foreign-student-visa-brain-drain. 120 The Economist (2024), ‘China has become a scientific superpower’, 12 June 2024, https://www.economist.com/ science-and-technology/2024/06/12/china-has-become-a-scientific-superpower. 121 Qi, X. (2025), ‘Chinese tech giants ramp up AI talent recruitment amid surging demand’, Global Times, 15 June 2025, https://www.globaltimes.cn/page/202506/1336180.shtml. 122 Shu, C. (2014), ‘Google Acquires Artificial Intelligence Startup DeepMind For More Than $500M’, Tech Crunch, 26 January 2014, https://techcrunch.com/2014/01/26/google-deepmind. 123 Westgarth, T., Chen, W., Hay, G. and Heath, R. (2022), Understanding UK Artificial Intelligence R&D commercialisation and the role of standards, Oxford Insights, May 2022, https://oxfordinsights.com/wp-content/ uploads/2023/10/DCMS_and_OAI_-_Understanding_UK_Artificial_Intelligence_R_D_commercialisation__ accessible-1.pdf. 124 Westgarth, Chen, Hay and Heath (2022), Understanding UK Artificial Intelligence R&D commercialisation. 125 UK DSIT (2025), ‘UK AI Opportunities Action Plan’. -- 31 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 30 Chatham House unavoidable, as demonstrated by the UK’s Alan Turing Institute launching additional open-source training with Meta’s support.126 Pooling talent resources is another emerging AI sovereignty pathway: influential EU policy agendas also push for cross-EU training (such as joint doctoral schools) and improved retention (for example, through academic benefits, access to data and compute, and support for societal impact research).127 The EU AI Office – responsible for streamlining the EU bloc’s governance of AI and preparing guidance for companies, regulators and other actors – might play an important role here. By sharing research and training resources – and investing in talent niches – middle powers have the potential to develop globally competitive human capital and research ecosystems for AI, but retaining these assets is difficult. Trust Without public trust and widespread adoption, efforts to build sovereign AI capabilities will likely fail. Consequently, legitimacy and usability are central to sustainable sovereignty. AI capabilities only translate into national advantages when widely and effectively adopted. User trust and capacity – dependent on use, literacy and availability – determine strategic returns across society, government and the private sector. The US and China excel when it comes to AI users. Both countries recognize the strategic value of investments in digital literacy and deployment. But both have mixed track records of success, with major inequalities in domestic adoption and utilization across demographic groups, partially as a result of gaps in literacy and broadband.128 Both countries also have differing political models – one a constitutional republic, the other a one-party authoritarian state – which impact 126 UK DSIT (2025), ‘UK’s best AI engineers can apply now to build tech for public services in $1 million fellowship’, press release, 11 July 2025, https://www.gov.uk/government/news/uks-best-ai-engineers-can-apply-now-to- build-tech-for-public-services-in-1-million-fellowship. 127 Scientific Advice Mechanism to the European Commission (2024), ‘Successful and timely uptake of artificial intelligence in science in the EU’, April 2024, https://scientificadvice.eu/advice/artificial-intelligence-in-science; Twinch, E. (2024), ‘EU advised to create institute to promote use of AI in research’, Research Professional News, 15 April 2024, https://www.researchprofessionalnews.com/rr-news-europe-politics-2024-4-eu-advised-to- create-institute-to-promote-use-of-ai-in-research. 128 Pick, J., Ren, F. and Sarkar, A. (2024), ‘Digital Inequalities in China in 2020: Spatial and Multivariate Analysis’, Applied Science, 14(13), https://doi.org/10.3390/app14135385; Gallardo, R. and Whitacre, B. (2024), ‘An unexpected digital divide? A look at internet speeds and socioeconomic groups’, Telecommunications Policy, 48(6), https://doi.org/10.1016/j.telpol.2024.102777. Both the US and China recognize the strategic value of investments in digital literacy and deployment. But both have mixed track records of success, with major inequalities in domestic adoption and utilization across demographic groups, partially as a result of gaps in literacy and broadband. -- 32 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 31 Chatham House how their citizens adopt and engage with technology. For example, China’s tighter control over its technology ecosystem enables more efficient scaling and diffusion than its US counterpart, whose approach favours frontier efforts over integration. The US technology ecosystem has long been lauded as more innovative than China’s, though there are signals this gap is closing quickly. China’s strategic planning emphasizes narrowing skill gaps and empowering ‘netizens’ through access to digital infrastructure, in order to enable broad testing and scaling.129 The US, on the other hand, benefits from extensive enterprise adoption of AI and venture capital investments that have supercharged deployment, with private investment in AI at $109.1 billion in 2024, dwarfing China’s $9.3 billion over the same period of time.130 Some applications – like autonomous vehicles – boast strong adoption but persistent public mistrust still exists. In the AI Index Annual Report for 2025, 80 per cent of Chinese respondents reported excitement about AI products and services, compared to 34 per cent in the US.131 Many middle powers lack the huge domestic markets and user base that the AI superpowers possess. Large populations alone are insufficient for the rapid scaling and adoption of AI in general, let alone homegrown AI. Indonesia is a case in point: with over 280 million people and a big digital economy, its adoption and utilization capacity is limited by skills and literacy gaps, exacerbated by middling (but rising) internet penetration.132 India’s $1.25 billion sovereign strategy prioritizes public–private partnerships and homegrown AI models, with a view to democratizing access outside major cities.133 Middle powers recognize how important user trust is to developing AI, particularly in the public sector. The pursuit of a national foundation model has been championed across many sovereign AI agendas. This could take the form of a national AI model that could potentially substitute powerful foundation models like OpenAI’s ChatGPT, localizing control over model development and deployment. National models could build trust and scalability, but barriers remain to feasibility – such as investment, regulatory concerns and data access134 – and future user adoption. Dependencies on foreign powers for AI products and services will do little to build public trust: and it is unlikely middle powers can develop national AI models without external input. But middle powers seeking to test and scale sovereign capabilities are well-placed to use trust as a lever to grow domestic AI capacity: not only through investments in users (i.e. skills and access), but also by pursuing domain-specific applications of general-purpose models.135 129 State Council of the People’s Republic of China (2025), ‘China unveils 2025 plan to boost skills’, State Council of the People’s Republic of China, updated 27 April 2025, https://english.www.gov.cn/news/202504/27/ content_WS680de811c6d0868f4e8f21c2.html; Creemers et al. (2022), ‘14th Five-Year Plan for National Informatization’. 130 AI Index Steering Committee, Institute for Human-Centered AI, Stanford University (2025), Artificial Intelligence Index Report 2025, https://hai.stanford.edu/ai-index/2025-ai-index-report. 131 Ibid., p. 402. 132 UNESCO Global AI Ethics and Governance Observatory (2023), ‘Indonesia’, https://www.unesco.org/ ethics-ai/en/indonesia. 133 Ghosh, S. (2024), ‘India’s US$1.25 billion push to power AI’, Nature India, 17 March 2024, https://www.nature.com/articles/d44151-024-00035-5. 134 This reflects input from a Chatham House roundtable, ‘Middle Power Pathways to Sovereign AI’, March 2025, London: Chatham House. The event was convened under the Chatham House Rule. 135 For a further discussion on bargaining, see: Blancato, F. G. (2025), Embraced Dependence: A Study on the UK’s Approach to Cloud Computing, SSRN, 7 May 2025, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5229150. -- 33 of 48 -- 32 Chatham House 04 The road ahead Middle powers face constraints, but viable pathways to AI sovereignty exist. Smart sovereignty strategies can allow countries to take opportunities for autonomy while managing necessary dependencies. Middle powers cannot win the global AI race. The dependencies outlined in previous chapters are too deep to unwind quickly or completely. Full-stack national ownership over AI’s various building blocks is not a realistic near-term goal for any middle power. But constraint does not mean paralysis. Middle powers retain meaningful strategic choices about how to manage dependence, where to build national capability and who to partner with. The question is not whether to be sovereign, but what sovereignty looks like when autonomy is necessarily partial and complex. This chapter presents four pragmatic pathways – specialize, align, share or hedge – that middle powers can pursue, alone or together, to secure influence over their AI futures. — Specialize: Middle powers focus resources on a specific segment of the AI supply chain and leverage that position for geopolitical influence. — Align: Middle powers deliberately choose deep technological dependence on either the US or China in exchange for guaranteed access, reduced costs, security and preferential treatment. — Share: Middle powers pool resources, coordinate policy and negotiate as a regional or interest-based bloc to exercise collective influence. — Hedge: Middle powers assemble a hybrid AI stack by deliberately choosing building blocks from multiple foreign providers while targeting domestic investments – such as in industry, material resources and workforce talent – to strengthen the national AI ecosystem. -- 34 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 33 Chatham House Box 2. Timelines and the future of AI Assessing sovereign AI strategies necessitates reckoning with the theoretical likelihood of an ‘AI supremacy’ moment. After which the drive for sovereignty does not cease, but might become moot, as at that point either the US or China will wholly dominate the development and deployment of AI – to the likely detriment of less powerful players.136 This concentration of power could render traditional notions of AI sovereignty obsolete, as the world fragments between those with access to advanced AI, AGI or even artificial super intelligence (ASI), and those without (including adversaries and those worried of falling behind).137 The technical feasibility of AGI aside, fears of a global split into powerful AI haves and have-nots is a motivational factor behind the development of sovereign capabilities for both middle powers and superpowers. Policymakers and industry can reference AGI as a justification for ramping up investments in primary and secondary building blocks, framed as an effort to safeguard national resilience irrespective of the likely timeframe for achieving AGI.138 Figure 1. How prospects for sovereign AI may relate to timelines to AGI Source: Compiled by the authors. 136 Artificial general intelligence (AGI) is contested terminology but generally refers to AI with capabilities matching or surpassing human capabilities. Artificial super intelligence (ASI) refers to capabilities that far surpass human capabilities; Bengio, Y. (2024), ‘Implications of Artificial General Intelligence on National and International Security’, YoshuaBengio, 30 October 2024, https://yoshuabengio.org/2024/10/30/implications- of-artificial-general-intelligence-on-national-and-international-security. 137 Pavel, B. et al. (2023), ‘AI and Geopolitics’, RAND, 3 November 2023, https://www.rand.org/pubs/ perspectives/PEA3034-1.html. 138 However, as observed in other areas of technological progress across history – in robotics and manufacturing, for instance – technologies, upon global diffusion, become tiered. Technological dominance by one superpower may not dampen the quest for sovereignty: if anything, it might supercharge it. 0 15 10 Time to AGI (years) 5 50+ Prospects for sovereign AI Low High Align Share Hedge Specialize -- 35 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 34 Chatham House Experts are divided on the near-term likelihood of AGI. Estimates range from two years to decades.139 Building a more resilient AI supply chain and a national AI industry, as well as strengthening state capacity to steward AI development are resource-intensive endeavours. For some countries, states may have already run out of time to achieve these ambitions. Many factors affect the relationship between AGI and the potential for middle powers to achieve sovereign AI: longer timelines to AGI would allow middle powers more autonomy to achieve their ambitions, whereas shorter timelines favour existing AI superpowers. Figure 1 demonstrates this moving technological target for policymakers. There are three scenarios: — Short-term (AGI in 2–5 years): Superpower dominance quickly forces middle powers towards dependence or niche specialization. — Medium-term (AGI in 5–20 years): Available time for sustained political will to pursue moderately ambitious sovereign AI strategies. With a window for strategic partnerships, targeted investments and building competitive advantages in specific domains. — Long-term (AGI in 20+ years): Extended timeline enables full but slower development of domestic AI ecosystems, talent pipelines and independent technological capabilities. Experts disagree on timelines to AGI. The expectations for different middle power sovereign AI strategies (included in the rest of this chapter) depend on this timeline. With the prospect of AGI, policymakers must strategize amid intense uncertainty. How middle powers interpret ‘time to AGI’ might influence their AI ambitions: ranging from full capitulation to a superpower to aspirations for full-stack independence. The rest of this chapter walks through four pragmatic sovereign AI pathways for middle powers. This is relevant to middle powers whose AI ambitions fall short of global hegemony but are resistant to pure capitulation. While a conclusive stance on which AGI scenario is the most likely to occur is beyond the scope of this paper, the below recommendations fall into the intermediate-to-long-term category. 139 See: Hastings-Woodhouse, S. (2025), ‘Why do people disagree when powerful AI will arrive?’, Blue Dot Impact, 2 June 2025, https://bluedot.org/blog/agi-timelines. Leading voices on AGI in the short term include: Aschenbrenner, L. (2024), ‘Situational Awareness: The Decade Ahead’, June 2024, https://situational-awareness.ai; Kokotajlo, D. et al. (2025), ‘AI 2027’, 3 April 2025, https://ai-2027.com; Kokotajlo, D. et al. (2025), ‘AI Futures Timelines and Takeoff Model: Dec 2025 Update’, 31 December 2025, Less Wrong, https://www.lesswrong.com/ posts/YABG5JmztGGPwNFq2/ai-futures-timelines-and-takeoff-model-dec-2025-update. -- 36 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 35 Chatham House Specialize For middle powers with existing industrial strengths, specialization offers a pathway to strategic influence: by achieving dominance in a specific segment of the AI supply chain to gain leverage. Rather than competing across the full stack, states concentrate resources on becoming indispensable in one area. These might be advanced hardware components, specialized chip fabrication, energy infrastructure for data centres, or AI applications in specific domains. The Netherlands’ ASML is the global leader in photolithography, producing the machinery necessary for semiconductor manufacturers to make more powerful chips. ASML’s dominance as currently the world’s only provider of extreme ultraviolet lithography systems for chip printing is unmatched, which provides the Netherlands with a geopolitical bargaining chip while igniting fears about bottlenecks, black market or secondary component acquisitions140 and changeable global markets, such as in trade restrictions and export controls. 141 Taiwan’s TSMC – operating since 1987 – is the world’s largest semiconductor ‘foundry’, which means it offers manufacturing and design services to clients like NVIDIA. TSMC’s critical position in the global semiconductor supply chain is of strategic value to Taiwan. However, TSMC must contend with geopolitical realities in relation to the potential for instability between China and Taiwan.142 Readiness to carve out niche specialization in global AI supply chains demands both intelligence (about the likely trajectory of frontier capabilities) and large investment. Which companies may become the TSMCs and ASMLs of the 2030s is an open question. They will likely be dependent on the advancement of AI technology itself as well as market dynamics. With the well-resourced US and China crowding the frontier, middle powers must lean on sovereign secondary building blocks for AI, such as investments in human capital and research ecosystems, accompanied by strategic partnerships with and investment from powerful AI companies. At the same time, middle powers must be strategic forward-planners, wary about maintaining control over specialized and promising capabilities. The acquisition of UK-grown Deepmind by US-based Google in 2014 is a cautionary tale. At the time, the deal – valued at over $500 million – was viewed as a vote of confidence in the UK’s AI ecosystem. Today, the UK’s failure to keep Deepmind onshore should be considered a major strategic loss. The case underlines the importance of a long-term, risk-based approach to developing and retaining sovereign capabilities. Specialized capabilities must be insulated against geostrategic impacts, such as volatile US politics and governmental entanglement with major technology companies. 140 Olcott, E. (2025), ‘China boosts AI chip output by upgrading older ASML machines’, Financial Times, 19 December 2025, https://www.ft.com/content/d10398db-b8b4-40f3-8c6d-b340470f5f3c. 141 Investing.com (2025), ‘ASML slightly lower after Samsung delays US chip plant completion’, https://uk.investing.com/news/stock-market-news/asml-slightly-lower-after-samsung-delays-us-chip-plant- completion-93CH-4157531. 142 Shepardson, D. (2024), ‘US official says Chinese seizure of TSMC in Taiwan would be ‘absolutely devastating’’, Reuters, 9 May 2024, https://www.reuters.com/world/us/us-official-says-chinese-seizure-tsmc-taiwan-would- be-absolutely-devastating-2024-05-08. -- 37 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 36 Chatham House This approach to sovereignty prioritizes national control over a specialized primary capability (such as niche hardware essential for the future training and operational needs of advanced AI) that can be bolstered by diverse foreign investment. To achieve this, countries should: — Conduct supply chain audits to identify specialist niches where existing national strengths can translate into global leverage, then concentrate investment in 1–2 areas rather than allocating resources too broadly. — Establish foreign investment review mechanisms with the authority to block acquisitions of AI-relevant capabilities that constitute strategic assets. — Negotiate formalized partnerships with frontier AI companies that exchange specialized capabilities for guaranteed long-term access to advanced models or compute infrastructure. Align For many middle powers, the most strategically coherent pathway to sovereign AI is deliberate alignment with either the US or China. Rather than fragmenting limited resources across multiple partnerships or attempting to build independent capabilities, aligned states commit fully to one superpower’s ecosystem in exchange for guaranteed access, preferential treatment, security and reduced costs. In this scenario, middle powers choose not to bear the costs associated with building niche specialization in the global AI supply chain. Nor do they seek interest-based alliances with other middle powers to pool resources and gain collective leverage vis-à-vis the US and China. As an example, middle powers could explore partnerships with frontier labs, such as Google Deepmind or Deepseek, that are advancing the development of more efficient, smaller models, referred to as lightweight AI. Greater research into strategies like model distillation – where smaller, more efficient models are trained to replicate the behaviour of larger, more complex models – could provide an AI approach for middle powers that is more accessible and reduces the computational costs of both training and running AI models. This is not a surrender from weakness but a strategic choice. The UAE exemplifies alignment executed from a position of strength. As the first international partner in the US Stargate initiative, the emirates have traded aspirations of full technological independence for privileged access to American AI capabilities and massive infrastructure investment. But Abu Dhabi negotiated this alignment carefully, leveraging its capital, energy resources and geopolitical position to extract substantial concessions, notably by anchoring itself into the US data centre ecosystem and gaining access to advanced models. The UAE is dependent on US technology, but it is a valued partner, not a vassal state. The UAE’s investments allow the US to reinforce its global dominance. Alignment offers stability and speed. Middle powers avoid the costs and risks of developing capabilities domestically or managing complex multi-vendor architectures. They gain access to the most advanced technologies without -- 38 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 37 Chatham House shouldering the full burden of innovation. For states with acute security threats, limited technical capacity, or strong existing geopolitical relationships with a superpower, alignment can be the most rational option. But alignment carries clear costs. Junior partners have limited room for manoeuvre if the superpower’s strategic priorities shift. Those that align face vulnerability to disruption if geopolitical relationships sour. They may be forced to adopt the AI superpower’s standards, regulations and values even when these conflict with domestic preferences. And countries that choose to align are excluded from technologies, markets and partnerships controlled by rival blocs. Alignment is not a hedge: it is a bet that one superpower will remain both willing and able to provide access over the long term. The key for middle powers pursuing this pathway is informed dependence: formalizing the terms of alignment, understanding what is being exchanged, monitoring risks and maintaining options for an exit if circumstances change radically. Alignment should be deliberate, not passive. To achieve this, countries should: — Where possible, formalize terms through bilateral agreements specifying mutual obligations, access guarantees, data sovereignty provisions and conditions under which arrangements may be suspended. — Maintain domestic technical capacity to integrate, monitor and evaluate foreign AI systems to preserve options should alignment terms deteriorate. — Leverage strategic assets – geographic position, market access, energy resources – to secure infrastructure investment, preferential access or technology transfer at the point of alignment. Share Middle powers that view total alignment with a superpower as unacceptable and lack the capacity for individual specialization can pool resources and negotiate collectively. Regional or interest-based blocs allow states to exercise influence that they cannot achieve alone, through joint investments in infrastructure, harmonized regulation, shared R&D programmes and collective bargaining with AI providers and superpowers. This frees up middle powers to pursue sovereign AI strategies without necessarily having to match superpower scale in compute or data.143 Proposals for a jointly governed, cross-border model or infrastructure development project – in the same way that Airbus was formed by regional European collaboration to compete with dominant aircraft manufacturers – reflect aspirational but growing interest in shared sovereignty arrangements among trusted partners.144 These models often complement national strategies: some states curate national champions while simultaneously contributing to open-source or consortium-based efforts that 143 Abecassis, A. et al. (2025), ‘A Blueprint for Multinational Advanced AI Development’, Oxford Martin School, University of Oxford, 24 November 2025, https://www.oxfordmartin.ox.ac.uk/publications/a-blueprint-for- multinational-advanced-ai-development. 144 See: Tan, Jackson, Berjon and Coyle (2025), Airbus for AI: A global strategy for public value creation. -- 39 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 38 Chatham House distribute cost and reduce strategic dependence. France’s Mistral and Canada’s Cohere – two leading labs with substantial backing from their home governments – are potential case studies for this dual approach of backing national AI champions while benefiting from collective and cross-border action. Embedding such hybrid approaches within multilateral or cross-country AI initiatives would expand the set of viable alternatives to countries seeking to diversify away from dominant foreign ecosystems. The EU offers the most developed example of shared sovereignty in AI. Through initiatives like the European High-Performance Computing Joint Undertaking, Horizon Europe research programmes, and the AI Act’s harmonized regulatory framework, Brussels is attempting to build collective capabilities that no single member state could afford. The EU will not become a frontier AI leader, but by pooling sovereignty it can secure a degree of strategic autonomy and collective leverage vis-à-vis the US and China. Shared sovereignty works best when states face common threats, trust each other sufficiently to coordinate, and can enforce collective discipline. Smaller regional groupings such as Gulf states might leverage energy abundance to build compute infrastructure. A small group of Nordic countries might make use of their cooler climates to coordinate investments in green data centres. Smaller groupings like these may achieve coordination more easily than large, diverse blocs like the EU whose member states (despite the rhetoric of European sovereignty) remain overwhelmingly dependent on US hyperscalers for cloud services and frontier models. Shared sovereignty amplifies a nations voice, but it does not automatically deliver autonomy, nor will it match superpowers in terms of scale and scope. The strategic appeal of the share pathway lies in its defensive value: collective action reduces the risk that any single member state can be isolated or coerced. Countries should: — Establish multilateral infrastructure funds with regional or interest-based partners to jointly procure compute, build data centres or develop shared models in targeted domains. — Harmonize procurement standards and regulatory frameworks across partner states to create economies of scale and collective bargaining power with technology providers. — Create joint research consortiums that share technical talent and compute while structuring intellectual property rights to enable both collective advancement and national commercialization The EU will not become a frontier AI leader, but by pooling sovereignty it can secure a degree of strategic autonomy and collective leverage vis-à-vis the US and China. -- 40 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 39 Chatham House Hedge For middle powers with strong technical capacity, diversified economies and a preference for strategic flexibility, hedging offers an alternative to alignment or collective action. Rather than committing to one superpower’s ecosystem or pooling resources with a regional bloc, states that hedge deliberately assemble a hybrid AI stack – cherry-picking capabilities from multiple foreign providers while building targeted national strengths in areas where independence is essential. Japan’s recent AI strategy illustrates this agile approach. Tokyo relies on NVIDIA and US cloud providers for compute and access to frontier models, partners with domestic firms like SoftBank for telecommunications infrastructure, and invests in Japanese-language AI models for government services. The country’s latest AI Promotion Act seeks to balance the needs of providers with the national requirement for better interoperability – aiming to guide technical innovation through long-term thinking on AI rule-making.145 Open-source AI makes hedging more viable. National builders can adapt multiple open-weight models146 to local contexts and languages without depending on a single proprietary provider. Governments can encourage competition between open-source or public AI models to avoid lock-in while accessing near-frontier capabilities. For middle powers with under-resourced languages or unique regulatory requirements, the ability to customize models locally creates genuine opportunity. But hedging is technically demanding. It necessitates sophisticated procurement frameworks that mandate interoperability, penalize vendor lock-in and require compatibility with multiple competing systems. It demands expertise to integrate disparate technologies – US cloud infrastructure, European privacy tools, Chinese hardware, domestic domain models – into coherent national systems. And it requires continuous vigilance: the global AI market’s extreme consolidation means vendor lock-in can happen through stealth, as proprietary application programming interfaces (APIs), data formats, isolated talent and training dependencies accumulate over time. Hedging also carries geopolitical costs. Diversifying across US, Chinese and European providers may trigger suspicion from all three, particularly in sensitive domains like defence or critical infrastructure. And managing a multi-vendor environment increases complexity and integration costs compared to committing fully to one ecosystem. These costs underscore the erosion of middle power flexibility in an ever-polarized world order. This pathway works best for middle powers, such as Singapore and Japan, that can afford the coordination costs, have deep technical expertise to manage hybrid systems and face no immediate existential security threats that would push them towards alignment. The goal is not independence – full autonomy 145 Habuka, H. (2025), Japan’s Agile AI Governance in Action: Fostering a Global Nexus Through Pluralistic Interoperability, report, Washington, DC: Center for Strategic and International Studies, 9 October 2025, https://www.csis.org/analysis/japans-agile-ai-governance-action-fostering-global-nexus-through-pluralistic. 146 Open-weight models are those that have their trained parameters (‘weights’) available and open to the public. -- 41 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 40 Chatham House remains unrealistic – but rather strategic flexibility: the ability to switch providers, adapt to disruptions and avoid becoming captive to any single superpower’s technology ecosystem. Countries aiming to hedge should: — Mandate multi-vendor interoperability standards in government AI procurement to prevent proprietary lock-in through closed APIs, incompatible data formats or opaque integration requirements. — Invest in sovereign capacity selectively – build genuine independence in critical domains (defence, essential infrastructure, sensitive government services) while accepting pragmatic dependence elsewhere. — Establish contingency mechanisms, including reserved funding and pre-negotiated alternative partnerships, to enable rapid provider substitution if primary suppliers restrict access due to geopolitical tensions. -- 42 of 48 -- 41 Chatham House 05 Conclusion Middle powers can make deliberate choices about where to specialize, where to depend on other countries and when to cooperate. In doing so, it is possible to secure durable influence over the global trajectory of AI. This paper highlights the fact that, for middle powers, achieving AI sovereignty is highly complex and multi-layered. There is no simple solution for mitigating dependence on foreign players. Inevitably, middle power ambitions will be constrained by the realities of superpower dominance over investment, data, frontier models and compute, alongside enabling factors, such as trust, infrastructure, energy, industry and talent. Recognition of these constraints is a pragmatic and strategic necessity for sovereign AI. For middle powers, it is important that strategic thinking is more informed in relation to the building blocks of AI (notably, where these are concentrated in the global AI supply chain) and realistic about the feasibility of AI ambitions and their longevity. Without the comprehensive capacities of the US or China, middle powers can retain some agency in determining how to build domestic AI capabilities and when to rely on or cooperate with others. Traditional measures of a country’s international standing – such as economy size, military capacity and diplomatic reach – still matter greatly for AI. But distinctive capabilities such as semiconductor manufacturing, ready access to affordable energy, or even regulatory and governance leadership can boost the global status of middle powers. This more fluid understanding of national strength demonstrates how power can be exercised not only through technical dominance, but also through the capacity to shape standards, govern access and manage dependencies. Against this backdrop, middle powers face a strategic landscape defined not by a binary choice of success or failure, but by a range of viable, pragmatic pathways – specialize, align, share or hedge – as detailed in the previous chapter. Each of these strategies carries risks, trade-offs and opportunities. None guarantees full sovereignty, but all offer ways to secure meaningful agency to develop and deploy AI in the national interest – in a system where complete autonomy is unattainable. -- 43 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 42 Chatham House The real challenge for middle powers is not to replicate the superpowers’ full-stack strategies, but to craft resilient and sustainable positions within an interdependent global AI order. By recognizing AI sovereignty as multi-layered, middle powers can move beyond the false dichotomy of ‘winning’ or ‘losing’ and instead focus on positioning themselves strategically. Through deliberate choices about where to specialize, where to depend and where to cooperate, middle powers can secure durable influence over the global trajectory of AI – with the aim of ensuring that sovereignty in the age of artificial intelligence remains not the privilege of a few, but a negotiated resource shared by many. -- 44 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 43 Chatham House About the authors Francisco Javier Varela Sandoval was the International Strategy Forum Academy fellow with the Digital Society Programme at Chatham House. His research focused on sovereign AI development pathways. Before joining Chatham House, Francisco was a fellow at France’s National Institute for Public Service, focusing on comparative administration. Over a 15-year career, he has worked as a policy practitioner, researcher and lecturer in Mexico. Isabella Wilkinson is a research fellow in the Digital Society Programme at Chatham House. Her work focuses on the geopolitics of global technology governance (specifically AI), new centres of decision-making power on technology and disinformation. Isabella is also a DPhil Public Policy student at the Blavatnik School of Government, University of Oxford, where her research explores the political economy of collective action among companies in global AI governance. Alex Krasodomski leads the Digital Society Programme. His work focuses on AI, emerging technology and centres of tech power. He leads work aimed at strengthening state capacity and cooperation, identifying feasible paths towards global technology governance, and on routes to market for public technology. He was previously research director at Demos, director of the Centre for the Analysis of Social Media and a co-founder of the AI start-up CASM Technology. He led the Good Web Project and is a fellow at the Institute for Strategic Dialogue (ISD). Rowan Wilkinson is a research associate in the Digital Society Programme, where she supports research on digital public infrastructure, the information space, tech sovereignty and AI governance. Her expertise lies at the intersection of technology, humanitarianism and international development. -- 45 of 48 -- How middle powers can weather US and Chinese AI dominance The case for ‘sovereign AI’ strategies 44 Chatham House Acknowledgments This paper is part of a long-term research project on credible AI governance. The expertise and insights offered by anonymous peer reviewers was essential to the finalization of this paper. We have immense gratitude for the time and energy they dedicated to reviewing this paper. This paper draws on interviews and workshops throughout 2025, including fieldwork in Southeast Asia, conducted by Francisco Javier Varela Sandoval as part of his Academy Fellowship at the Chatham House. The authors are grateful to all those who participated, giving their time and valuable comments. Brainstorming and early drafts would not have been possible without the valuable input of Joshua Entsminger. The team thanks him sincerely for his ideas and suggestions for the paper’s final form. Many thanks are due to Chatham House’s publications team and the staff of the Global Governance and Security Centre for their support, particularly Bianka Venkataramani. -- 46 of 48 -- All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical including photocopying, recording or any information storage or retrieval system, without the prior written permission of the copyright holder. Please direct all enquiries to the publishers. Chatham House does not express opinions of its own. The opinions expressed in this publication are the responsibility of the author(s). Copyright © The Royal Institute of International Affairs, 2026 Cover image: Conference participants look at a model of the Stargate initiative data centre currently under construction in Abu Dhabi, 3 November 2025. Photo credit: Copyright © Giuseppe Cacace/AFP/Getty Images ISBN 978 1 78413 671 0 DOI 10.55317/9781784136710 Cite this paper: Sandoval, F. J. V., Wilkinson, I., Krasodomski, A. and Wilkinson, R. (2026), How middle powers can weather US and Chinese AI dominance: The case for ‘sovereign AI’ strategies, Research Paper, London: Royal Institute of International Affairs, https://doi.org/10.55317/9781784136710. This publication is printed on FSC-certified paper. designbysoapbox.com -- 47 of 48 -- The Royal Institute of International Affairs Chatham House 10 St James’s Square, London SW1Y 4LE T +44 (0)20 7957 5700 contact@chathamhouse.org | chathamhouse.org Charity Registration Number: 208223 Independent thinking since 1920 -- 48 of 48 --