Sovereign AI

Definition and scope

Sovereign AI is a national government's policy of placing the development, deployment and control of AI models, infrastructure and data in the hands of domestic actors, and is described as an offshoot of digital sovereignty. It is widely used but lacks a single agreed definition, and the projects it covers span computing infrastructure and cloud services, models and data, technical skills, public procurement and regulation. The academic literature separately defines full-stack sovereignty as control over the full AI lifecycle, from data collection and model training through deployment and governance, with data, models, GPU clusters and cloud data centres as its key layers. It notes that complete self-sufficiency is a tough call even for the United States and China.

History

The concept started gaining prominence around 2020, particularly within European policy discussions focused on digital sovereignty and data protection. In 2023 NVIDIA CEO Jensen Huang began promoting the specific expression sovereign AI to governments, arguing that countries should develop AI systems reflecting their languages, knowledge and culture. A rush of global activity followed. Reuters reported in 2025 that the message had gained support among European leaders while Nvidia sought to supply processors for the resulting projects. Pablo Chavez tracked roughly 40 government-backed sovereign AI projects across about 30 countries in 2024, rising to nearly 130 projects in more than 50 countries by January 2026; the CNAS Sovereign AI Index counted 184 projects across 67 countries as of June 2026, with infrastructure projects at 59 percent, model projects at 32 percent and data projects at 9 percent, up from a single project tracked in 2023 and with infrastructure accelerating sharply since mid-2024.

Relations to adjacent concepts

Sovereign AI overlaps with the broader concepts of technological sovereignty, digital sovereignty and data sovereignty. In the technological-sovereignty literature the goal is reliable access to critical technologies without uncontrollable structural dependence, which 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. Six drivers recur behind sovereign AI initiatives: keeping sensitive data under domestic jurisdiction and limiting exposure to foreign legal compulsion; continuity and resilience of critical AI services; security, compliance and domestic oversight under national law; economic development and domestic capability-building; reducing single-provider dependence and vendor lock-in; and preserving national languages and cultural context.

Distinctions and debate

Critics argue the term obscures the differences between local data hosting, adaptation of a foreign model, and control of a complete technology stack, and that the lack of a common threshold lets governments and companies apply the same label to substantially different arrangements. Capgemini's CEO divided digital autonomy into control over data, operations, regulation and technology, arguing that no country controls the entire technology value chain. Sovereign AI refers to state efforts to control AI capabilities, whereas AI nationalism frames AI as geopolitical competition between states and appears as a distinct concept in AI policy taxonomy. The Brookings Institution holds that full-stack AI sovereignty is structurally infeasible for most countries and proposes managed interdependence: mapping dependencies by layer, prioritising feasible interventions, diversifying suppliers and partners, and embedding interoperability through standards, procurement and governance. The CNAS index likewise concludes that no country, not even the United States, can achieve full control over the complex and varied inputs that power frontier AI systems, and that for most economies sovereign AI means managing rather than eliminating dependencies.

Examples and limits

Most sovereign AI projects rely heavily on foreign, overwhelmingly American, technology providers. The CNAS index counts NVIDIA alone as supplying GPUs for 45 percent of tracked infrastructure projects, and the academic literature identifies the chip layer as the most obvious layer where countries inevitably have to cede sovereignty, with Nvidia dominating AI GPU demand at around 80 percent of the global market. Building national data centres to escape U.S. cloud platforms only shifts exposure from one layer of the U.S. tech stack to another. Large public spending is also hard to evaluate: a 2026 Guardian investigation into British AI projects found some figures originated with companies rather than independent government audits, some announced investments involved renting space in existing data centres, and a proposed sovereign AI supercomputer site remained undeveloped at the time of reporting.

Compute examples include the EU's 2025 outline of a 20 billion euro plan for three to five AI gigafactories, each intended to contain more than 100,000 advanced processors; South Korea's announcement of 10,000 high-performance GPUs obtained through public-private cooperation for a national AI computing centre; and Taiwan's 15-megawatt cloud-computing centre opened in Tainan in December 2025 hosting the Nano 4 supercomputer. On models, Switzerland released the open model Apertus in 2025, trained on more than 1,800 languages using public data, and Portugal released Amalia, an open foundation model for European Portuguese, in July 2026. Government-backed sovereign models include Japan's Fugaku-LLM, the Netherlands' GPT-NL (about 15 million dollars), Singapore's SEA-LION family (about 52 million dollars, 11 Southeast Asian languages), Spain's foundation models on MareNostrum 5, Sweden's GPT-SW3, Taiwan's TAIDE (about 7.4 million dollars) and the UAE's Falcon family; most are open source, and protecting and promoting national languages is a common stated goal. Developers of Singapore's SEA-LION models characterised them as complements to, not replacements for, larger international systems. Taiwan launched a 7.4 million dollar project in 2023 to build TAIDE by enhancing Meta's Llama open models with Taiwanese government and media data, partly to counter Chinese AI chatbots and secure a domestic alternative aligned with Taiwanese culture and facts.

Partnership examples include a 2025 France-UAE framework covering investment in chips, data centres, talent and virtual data embassies; a jointly owned Aker and Nscale facility in Norway using Nvidia processors with OpenAI as first customer; and Saudi sovereign-wealth-backed Humain agreeing to receive 18,000 Nvidia processors. Great-power competition overlaps with all of this: Reuters reported OpenAI's claim that China's Zhipu AI was offering governments private hardware and sovereign LLM infrastructure in partnership with Huawei, while OpenAI promoted an OpenAI for Countries programme with the U.S. government; Russia's Sberbank marketed locally adapted models to Global South countries while an executive acknowledged dependence on foreign chips and Nvidia's CUDA ecosystem.