Sovereign AI
Sovereign AI is a loosely defined term used in policy and industry debates for national or regional efforts to increase control over artificial intelligence (AI) capabilities and reduce critical dependence on foreign providers.[1][2] Projects described with the term can involve computing infrastructure, cloud services, models and data, technical skills, public procurement and regulation.[3][4][5] The term has no single agreed scope.[1][6][7] Some initiatives rely on imported processors, foreign cloud technology or international partnerships while retaining local control over data, deployment or access.[7][2]
The expression became more prominent in the mid-2020s as governments announced national and regional computing programmes, locally adapted models, investment funds and skills initiatives. Supporters associate these policies with national security, control over public-sector systems, and representation of local languages and cultures. Critics question their cost and performance, the environmental burden of large data centres, continued dependence on global supply chains, and the role of technology vendors in defining and supplying sovereignty projects.[1][8][9]
Concept and scope
[edit]Sovereign AI overlaps with the broader concepts of technological sovereignty, digital sovereignty and data sovereignty. In the literature on technological sovereignty, the central objective is often reliable access to technologies considered critical without an uncontrollable structural dependence on another country. This does not necessarily require domestic production of every component: international division of labour, redundant suppliers and trusted partners can form part of a sovereignty strategy.[6] German digital minister Karsten Wildberger similarly described European digital sovereignty as the ability to choose where data is stored and who operates infrastructure, while continuing to work with United States companies.[10]
In practice, the label has been applied to projects at different levels of the AI supply chain. Some focus on local data centres or cloud services; others support open or local-language foundation models, domestic AI companies, public-sector adoption, specialist training or the ability of regulators and public purchasers to set conditions for use.[1][3][5] Regional programmes in the European Union show that sovereignty can also be pursued collectively rather than only by individual states.[11]
The breadth of the term can obscure important distinctions. Capgemini chief executive Aiman Ezzat divided digital autonomy into control over data, operations, regulation and technology, and argued that no country controls the entire technology value chain. Capgemini itself marketed services as sovereign while operating them on infrastructure supplied by United States cloud companies.[7] In this usage, sovereignty refers to governance and operational choice rather than complete ownership of hardware and software.
Development
[edit]The wider debate over technological sovereignty grew as governments reconsidered dependence on foreign components and platforms. Edler described the policy challenge as balancing access to critical technologies with the benefits of international openness and cooperation.[6] Nvidia chief executive Jensen Huang began promoting the specific expression sovereign AI to governments in 2023, arguing that countries should develop AI systems reflecting their languages, knowledge and culture. Reuters reported in 2025 that the message had gained support among European leaders while Nvidia sought to supply processors for many of the resulting projects.[2] Huang described sovereignty as autonomy rather than isolation; the Associated Press also reported criticism that Nvidia's dominance could concentrate power.[9]
By 2025 and 2026, programmes described in sovereignty terms had appeared in Europe, Asia and the Middle East, ranging from national computing centres to multilingual models and public investment funds.[1][11][12][13] The European Union's broader 2026 technology-sovereignty package covered chips, cloud computing and AI, with the stated aim of avoiding risky dependence on a single supplier or country.[11]
Policy approaches
[edit]Compute and cloud infrastructure
[edit]Access to computing capacity is a recurring element of sovereign-AI programmes. In 2025 the European Union outlined a €20 billion plan for three to five AI "gigafactories", each intended to contain more than 100,000 advanced processors. The plan combined public support with private investment and also sought to develop European-made AI semiconductors.[8] South Korea announced that it would obtain 10,000 high-performance graphics processing units (GPUs) through public–private cooperation for a national AI computing centre. Reuters noted that Nvidia then controlled about 80 per cent of the global GPU market.[14]
Taiwan opened a 15-megawatt cloud-computing centre in Tainan in December 2025 as part of its sovereign-AI programme. The centre hosts the Nano 4 supercomputer and combines Taiwan's semiconductor-manufacturing and systems-integration capacity with processors designed by the United States company Nvidia.[15]
Projects described as sovereign are frequently cross-border. A 2025 framework between France and the United Arab Emirates covered investment in chips, data centres, talent and virtual "data embassies" intended to support sovereign AI and cloud infrastructure in both countries.[16] In Norway, Aker and Nscale planned a jointly owned facility using Nvidia processors, with OpenAI as its first customer.[17]
Models, data and language
[edit]Locally developed or adapted models are intended to improve coverage of languages and cultural contexts that are underrepresented in global systems and to give public institutions systems that can be adapted or operated under local control. Reporting in 2025 described government-backed or regional projects for Southeast Asian languages, Indian linguistic and legal contexts, and Swiss language use. Developers of Singapore's SEA-LION models characterised them as complements to larger international systems rather than replacements for them.[1]
Switzerland released the open model Apertus in 2025. Its source code, model weights, training data and development process were made available, and the model was trained on more than 1,800 languages using public data.[18] Portugal released Amalia, an open foundation model for European Portuguese, in July 2026. It was developed by a consortium of universities and research institutions as a base for applications in public administration, education, research and business, with access to Portuguese and European supercomputing resources.[19]
Skills, investment and procurement
[edit]Sovereign-AI policies also use grants, equity investment, public procurement, access to computing resources and training. The European Commission allocated €1.3 billion for AI, cybersecurity and digital skills under the 2025–2027 Digital Europe Programme.[4] Its later €1 billion "Apply AI" strategy supported adoption in sectors including healthcare, energy, manufacturing, mobility and defence and linked adoption to European strategic autonomy.[5]
The United Kingdom launched a £500 million sovereign-AI fund in 2026. Its first investments combined government shareholdings in domestic companies with access to publicly funded supercomputers for selected firms.[13] Technology secretary Liz Kendall later described domestic capability as a response to the concentration of global computing infrastructure among a small number of United States companies, while saying that sovereignty did not require weakening cooperation with the United States.[20]
International partnerships and competition
[edit]National programmes often depend on foreign suppliers. In 2025 Saudi Arabia's sovereign-wealth-fund-backed company Humain agreed to receive 18,000 Nvidia processors for a large data-centre project.[12] The project combined state planning and domestic ownership with imported hardware.
Sovereign-AI services have also become part of international technology competition. Reuters reported OpenAI's claim that China's Zhipu AI was offering governments private hardware and sovereign large-language-model infrastructure in partnership with Huawei; the same report noted that OpenAI was promoting an "OpenAI for Countries" programme in coordination with the United States government.[21] Russia's Sberbank similarly marketed locally adapted models to countries in the Global South, while one of its executives acknowledged dependence on foreign chips and Nvidia's CUDA software ecosystem.[22]
Criticism and limitations
[edit]Ambiguity and vendor influence
[edit]Critics argue that the term can obscure differences between local data hosting, adaptation of a foreign model, and control of a complete technology stack. The lack of a common threshold permits governments and companies to apply the same label to substantially different arrangements.[1][7]
Technology suppliers have helped popularise the concept and may benefit from the associated public spending. Reuters reported that Nvidia's campaign encouraged local infrastructure investment while keeping its processors central to many projects.[2] Associated Press described Nvidia as presenting itself as an engine of national AI infrastructure while critics questioned whether the capital, energy and specialist labour required for large systems would concentrate capability among wealthy states and companies.[9] An industry open letter signed by Airbus and more than 90 European companies and trade groups advocated a sovereign infrastructure fund and "buy European" procurement across applications, AI models, chips, computing, storage and connectivity.[3]
Cost, performance and oversight
[edit]Training frontier models and operating large computing facilities require substantial capital, electricity, cooling and specialist labour. Analysts interviewed by The Guardian argued that smaller states may gain more from adapting existing models, building expertise in selected areas or cooperating regionally than from trying to reproduce frontier systems. The same reporting warned that a poorly designed national model could be expensive and lightly used.[1]
Large investment announcements can also be difficult to evaluate. A 2026 Guardian investigation into British AI projects found that some figures originated with companies rather than independent government audits, that some announced investments involved renting space in existing data centres, and that a proposed sovereign-AI supercomputer site remained undeveloped at the time of reporting.[23]
Persistent dependencies and environmental constraints
[edit]Systems operated within national borders can still depend on foreign processors, cloud services and software ecosystems. The Taiwanese, Saudi and Norwegian projects all combined local ownership or control with technology supplied by foreign companies.[15][12][17][22] South Korea's planned national computing centre was announced in a GPU market in which Reuters estimated Nvidia's share at about 80 per cent.[14]
Computing infrastructure also creates local environmental and energy constraints. The European gigafactory plan prompted concern about electricity demand, cooling-water consumption and possible conflict with climate targets.[8] Reuters separately reported that high electricity costs and rising demand were challenges for Europe's sovereign-AI build-out.[2]
Approaches to interdependence
[edit]Proposed approaches between unrestricted dependence and technological autarky include specialising in selected layers of the AI stack, pooling resources among several countries, diversifying suppliers, and supporting open models that can be adapted and hosted locally. In the broader technological-sovereignty literature, reliable access through trusted partners and redundant supply can satisfy sovereignty objectives without domestic production of every component.[6][10]
One proposal reported by The Guardian, informally called an "Airbus for AI", would create a jointly owned public company through which middle powers could share the cost of model development.[1] Open models such as Apertus and Amalia represent another approach: institutions can inspect, adapt and host a shared base technology without reproducing every stage of frontier-model development.[18][19] These sources frame practical autonomy in terms of the dependencies reduced, the capabilities created, and the public value achieved relative to financial and environmental costs.[6][1][8]
See also
[edit]References
[edit]- 1 2 3 4 5 6 7 8 9 10 Down, Aisha (9 October 2025). "Governments are spending billions on their own 'sovereign' AI technologies – is it a big waste of money?". The Guardian. Retrieved 15 July 2026.
- 1 2 3 4 5 Mukherjee, Supantha (16 June 2025). "Nvidia's pitch for sovereign AI resonates with EU leaders". Reuters. Retrieved 15 July 2026.
- 1 2 3 "Airbus leads call for Europe to create sovereign infrastructure fund, buy European". Reuters. 17 March 2025. Retrieved 15 July 2026.
- 1 2 "EU to invest $1.4 billion in artificial intelligence, cybersecurity and digital skills". Reuters. 28 March 2025. Retrieved 15 July 2026.
- 1 2 3 "EU rolls out $1.1 billion plan to ramp up AI in key industries amid sovereignty drive". Reuters. 8 October 2025. Retrieved 15 July 2026.
- 1 2 3 4 5 Edler, Jakob (2024). Technology sovereignty of the EU: Needs, concepts, pitfalls and ways forward (Report). Fraunhofer-Gesellschaft. pp. 427–450. Retrieved 15 July 2026.
- 1 2 3 4 "Capgemini CEO dismisses calls for full European tech autonomy". Reuters. 13 February 2026. Retrieved 15 July 2026.
- 1 2 3 4 Rankin, Jennifer (9 April 2025). "EU to build AI gigafactories in €20bn push to catch up with US and China". The Guardian. Retrieved 15 July 2026.
- 1 2 3 Adamson, Thomas; Chan, Kelvin (11 June 2025). "Nvidia chief calls AI 'the greatest equalizer' — but warns Europe risks falling behind". Associated Press. Retrieved 15 July 2026.
- 1 2 "EU need for 'digital sovereignty' does not mean protectionism, German minister says". Reuters. 11 October 2025. Retrieved 15 July 2026.
- 1 2 3 "European Union launches tech sovereignty initiative to boost chips, cloud and AI at home". Associated Press. 3 June 2026. Retrieved 15 July 2026.
- 1 2 3 "Nvidia to send 18,000 AI chips to Saudi Arabia". Associated Press. 13 May 2025. Retrieved 15 July 2026.
- 1 2 "Liz Kendall urges UK public to embrace AI as government makes first £500m fund investment". The Guardian. 17 April 2026. Retrieved 15 July 2026.
- 1 2 "South Korea aims to secure 10,000 GPUs for national AI computing centre". Reuters. 17 February 2025. Retrieved 15 July 2026.
- 1 2 "Taiwan opens new cloud centre to bolster 'sovereign AI' effort". Reuters. 12 December 2025. Retrieved 15 July 2026.
- ↑ "France, UAE agree to develop 1 gigawatt AI data centre". Reuters. 6 February 2025. Retrieved 15 July 2026.
- 1 2 "OpenAI to build its first European data centre in Norway, with partners". Reuters. 31 July 2025. Retrieved 15 July 2026.
- 1 2 Welle, Elissa (3 September 2025). "Switzerland releases its own AI model trained on public data". The Verge. Retrieved 15 July 2026.
- 1 2 "Portugal launches first open-source AI model, joining Europe's sovereignty push". Reuters. 1 July 2026. Retrieved 15 July 2026.
- ↑ Milmo, Dan (28 April 2026). "UK must seize initiative on AI or be left at its mercy, Liz Kendall says". The Guardian. Retrieved 15 July 2026.
- ↑ "OpenAI says China's Zhipu AI gaining ground amid Beijing's global AI push". Reuters. 25 June 2025. Retrieved 15 July 2026.
- 1 2 "Russia's Sberbank offers AI model to Global South states keen to bridge digital divide". Reuters. 3 June 2026. Retrieved 15 July 2026.
- ↑ Down, Aisha (9 March 2026). "Revealed: UK's multibillion AI drive is built on 'phantom investments'". The Guardian. Retrieved 15 July 2026.
