Introduction
The New America Security Center's latest "Sovereign AI Index" 2.0 shows that global sovereign AI projects have increased from 1 in early 2023 to 184, covering 67 countries and the European Union. In this movement, which has evolved from being exclusive to rich countries to a global trend, the most notable change is on the supply side: while US foundational models still dominate, signs of loosening are emerging, with Chinese models entering six countries through two laboratories. Meanwhile, the proportion of models built from scratch has risen to 40%, and industrial-level sovereign AI is accelerating its landing. For China, which is both a participant in independent construction and a supplier of technology going overseas, the opportunities and challenges under this dual role are worth deep consideration.
Sovereign AI's second wave is on the horizon, with countries such as the United States, China, and the European Union actively developing their own artificial intelligence technologies, marking a significant shift away from relying on foreign companies. This new wave of sovereign AI is driven by the need for data security, privacy, and autonomy, as governments and companies seek to reduce their dependence on foreign technologies and promote domestic innovation. In recent years, the development of AI has been dominated by tech giants such as Google, Amazon, and Facebook, which have accumulated vast amounts of data and developed advanced AI algorithms. However, with the increasing importance of AI in various industries, countries have begun to realize the need to develop their own AI capabilities to ensure national security, economic competitiveness, and technological autonomy. China, for example, has launched a series of initiatives to promote the development of domestic AI technologies, including the establishment of a national AI research center and the allocation of billions of dollars in funding for AI research and development. The Chinese government has also introduced policies to encourage the development of domestic AI companies, such as tax incentives and subsidies for AI startups. Similarly, the European Union has launched its own AI strategy, which aims to promote the development of trustworthy AI that respects European values and promotes economic
The original text comes from the Center for New American Security (CNAS)
Interpretation and Analysis: Commentary on Digital Economy Development
Sovereign AI is evolving from a game for rich nations to a global trend, with the most notable shift occurring on the supply side: the dominance of US foundational models is beginning to loosen, as Chinese models enter sovereign AI projects in an increasing number of countries.
Data from the 2.0 version of the Sovereign AI Index released by the Center for a New American Security (CNAS) in August shows that Chinese models, which were largely absent before 2025, have now appeared in six countries with the technical foundation of two laboratories. The narrative of sovereign AI appears to be a pursuit of self-sufficiency on the surface, but in essence, it is a reshuffle of the supply side: the country whose model base is chosen by more countries will occupy a favorable position in the new round of technological landscape reconstruction.
From 1 to 184 Projects: A Global Expansion of Panoramic Scale
The index provides a quantifiable panoramic scale for observing the global expansion of sovereign AI. In January 2023, the index only tracked one sovereign AI project; by mid-2026, the cumulative total had reached 184 projects, covering 67 countries and the European Union region. In the first half of 2026 alone, 41 new projects were added, exceeding the total for all of 2024; the peak of new additions occurred in the second half of 2025, with 51 projects. The growth from one to 184 over three years indicates that sovereign AI is not a policy preference of individual countries, but a global structural response.
The rollout itself has the characteristic of being "broad but shallow": in the countries tracked by the index, 40% only support one project. The threshold for participating countries is decreasing: in 2023, it was mainly wealthy economies in Europe, East Asia, and the Gulf region, but by mid-2026, it had expanded to 67 countries, including 10 that were entering for the first time, all of which are lower-middle-income economies, covering Cambodia, the Dominican Republic, Egypt, Ghana, Nepal, Pakistan, Papua New Guinea, Paraguay, the Philippines, and Uruguay.
About one-third of countries worldwide are currently advancing some form of sovereign AI, accounting for approximately 90% of global GDP outside of the US and China. This figure indicates that sovereign AI has shifted from a debate over "whether to do it" to a practical issue of "how to do it".
The rollout of sovereign AI is not advancing uniformly, but rather is characterized by pioneers continuing to increase their investment: countries that already had projects in 2023 have since added more. In the first half of 2026, about 70% of new projects came from countries already included in the index, with just 10 countries accounting for nearly 40% of all tracked projects. New entrants have expanded the scope of participation, but it is the veteran players who are continuously increasing their investment that are at the core of the supply-side game - their choices of foundation directly determine the direction of the global sovereign AI supply chain.
US Dominance Shows Signs of Loosening
At the model level, the US still dominates, but the foundation is shifting. Most countries recognize that building a competitive sovereign model from scratch is not realistic, so they still use US foundation models as their technical base. CNAS judges that countries' continued preference for the US base is "not a matter of course". The fact that a US think tank is proactively issuing a warning is noteworthy in itself.
The driving force behind the change comes from the entry of Chinese models. Chinese models were largely absent before 2025, but now they have appeared in six countries through two laboratories: Alibaba's Qwen has entered Egypt, Singapore, Thailand, Uganda, and the UAE; Kazakhstan's national model AlemLLM is provided with direct technical support by 01.AI, reportedly based on a Chinese proprietary base. The speed at which Chinese models have entered the sovereign AI supply chain, from being absent to covering six countries, may exceed the expectations of most observers.
The CNAS report does not explicitly state the specific reasons why countries choose Chinese bases, but the data points to two key factors: one is cost and availability, and the other is the guarantee of supply continuity. The core motivation of sovereign AI is to reduce dependence on external parties, and the choice of which supplier to use depends on who is less likely to cut off supplies at critical moments. This diversification strategy has opened a window of opportunity for Chinese bases.
Building from Scratch: A Pragmatic Tech Approach
Models built from scratch are significantly behind the state-of-the-art in terms of capabilities, but the CNAS report emphasizes that "sub-state-of-the-art does not mean useless." This judgment provides a reasonable basis for model projects in "Global South" countries. About two-thirds of the models built from scratch have fewer than 700 billion parameters, a huge disparity with China's Yuezhuan Kimi K3, which has 28 billion parameters.
However, as long as the capabilities meet the specific scenario requirements, they have their application value: Apertus, a model trained from scratch in Switzerland, has overall performance lower than the cutting-edge level, but after being fine-tuned by Ticino, it performs better than Meta's Llama Scout in the translation of confidential files. This case reveals a universal trade-off in sovereign model projects: accepting a downgrade in capabilities in exchange for native control over the data. In fields involving proprietary data, strict regulation, or national security sensitivity, such a trade-off is often reasonable.
The Rise of Full-Stack Differentiation and Industrial-Level Sovereign AI
The advancement of sovereign AI is showing signs of divergence. Computing power, models, and data constitute the three layers of the AI technology stack, each of which is a threshold, and most countries are only staying at the computing power construction layer. Indicators show that only a few economies, including India, South Korea, the UK, and China Taiwan, have operable projects on all three layers. The gap between full-stack sovereign AI participants and countries that only have single-layer projects may be more noteworthy than the traditional wealth gap between nations.
Meanwhile, industrial-level sovereign AI is emerging. As key industries such as shipbuilding, finance, and defense start building their own model foundations, sovereign AI is transitioning from macro strategic planning to specific industrial practices. In the first half of 2026, South Korea launched a shipbuilding foundation model, Taiwan, China launched a financial industry model, and Ukraine launched a defense data platform; earlier examples include Turkey funding specific models for the medical and financial fields, and France launching a defense AI project.
Proprietary data, strict regulations, and national security implications give key industries a strong motivation to reduce their dependence on external sources. Sovereign AI is sinking from the "national strategy" level to the "industry infrastructure" level, with industry data becoming a new focal point of competition.
In the face of this global trend, the CNAS report's advice to Washington is to promote the widespread adoption of the US AI technology stack overseas as a policy goal, which is neither realistic nor wise, as no country, including America's closest allies, would accept such a degree of dependence. The correct goal for the US should be to become the "global partner of choice for sovereign cooperation" by making access to the US AI technology stack more predictable, transparent, and reliable, thereby alleviating countries' anxiety about technological dependence.
For the domestic market, this index turns the concept of "sovereign AI" from an abstract idea into a quantifiable global trend, while also revealing China's dual role in it: as a participant, with its own computing power and autonomous model construction, and as a supplier, with technologies such as Tongyi Qianwen and Lingyiwu already being used in other countries' sovereign AI projects. As the US consolidates its ally ecosystem with a "preferred sovereign partner" positioning, the export of Chinese models has formed another supply chain.
What deserves scrutiny is not merely whether one's own sovereignty is sufficient, but what it means to serve as the supply foundation for other nations' sovereign AI—a role that carries multiple costs spanning technical standards, data governance, and geopolitical trust. Notably, the heaviest investors in sovereign AI are the first movers who launched initiatives in 2023 and have steadily scaled up since. These countries hold the strongest sovereignty demands and are the most cautious about switching infrastructure, making them both the hardest customers for China to win over and the most strategically valuable clients.