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.
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. 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. While the US lags in solar energy production, despite growing demand, the country has resorted to the use of fossil fuels to meet its energy demands under President Donald Trump.
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. Others, like Japan with its ‘Watt-Bit Collaboration’ framework, 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.
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 – accounting for close to two-thirds of that number.
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. 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. 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).
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 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’ – 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.
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). 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. 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. In universities, Chinese students are offered extra credits, and faculties extra compensation, for contributing to the open-source AI ecosystem.
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. 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.
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’.
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. 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 – 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. 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. Chinese companies are also making major strides towards training – and retaining – the next generation of AI professionals.
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. 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.
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.
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. Strategic partnerships with powerful AI companies are for the most part unavoidable, as demonstrated by the UK’s Alan Turing Institute launching additional open-source training with Meta’s support.
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). 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.
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.
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. Both countries also have differing political models – one a constitutional republic, the other a one-party authoritarian state – which impact 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. 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. 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.
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. India’s $1.25 billion sovereign strategy prioritizes public–private partnerships and homegrown AI models, with a view to democratizing access outside major cities.
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 access – 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.
