CNAS Insights | Sovereign AI’s Second Wave Is Coming Into View

Earlier this year, CNAS launched the Sovereign AI Index to track every government-backed sovereign AI project since 2023. These projects seek to improve a country’s AI self-reliance by increasing capacity and reducing dependency at the infrastructure, model, and data layers of the AI stack. Our tracking shows that, over the past three years, rapid AI progress and growing anxiety about technological dependence have produced a global surge of sovereign AI projects. In January 2023, the index highlighted just a single project. In the latest index data, covering every new project in the first half of 2026, the total is now 184 projects spanning 67 countries.

Below are key insights from the latest data, which reveal several trends about where, how, and with whom countries are pursuing sovereign AI with limited resources.

Global momentum for sovereign AI continues. The index tracked 41 new projects in the first half of 2026. Governments launched more sovereign AI projects in the first half of 2026 than in all of 2024. The only six-month period with more new projects was the last six months of 2025, with 51 new projects added. At the same time, the diffusion of sovereign AI remains broad but shallow. Forty percent of all countries tracked in the index are backing just a single project.

Sovereign AI Projects Are Surging

Source: CNAS Sovereign AI Index, Version 2.0

Sovereign AI is no longer the preserve of mostly wealthy countries. In 2023, the index tracked 16 governments backing sovereign AI projects. They were overwhelmingly wealthier countries in Europe, East Asia, and the Gulf. By mid-2026, the footprint had expanded considerably, rising to 67 countries and the European Union. They include 10 first-time entrants to the index, all from lower- and middle-income economies: Cambodia, the Dominican Republic, Egypt, Ghana, Nepal, Pakistan, Papua New Guinea, Paraguay, the Philippines, and Uruguay. At the year’s halfway point, one in three countries worldwide was pushing for sovereign AI in some form, representing 90 percent of global GDP outside the United States and China.

New regions are also putting down first markers. Papua New Guinea launched the first sovereign AI project in the Pacific tracked by the index, while the Dominican Republic launched the first in the Caribbean.

Sovereign AI leaders are doubling down. Every country that had a sovereign AI project in 2023 has since added more: Roughly 70 percent of all new projects in the first half of 2026 came from countries already tracked in the index. Countries are treating sovereign AI projects as first steps to address specific needs and dependencies, not as comprehensive programs. A smaller subset is going even further: Just 10 countries account for nearly 40 percent of all tracked projects.

The Asia-Pacific is an emerging hub for sovereign AI. For the first time, the Asia-Pacific added more projects than Europe—16 versus 13, respectively—compared to any previous six-month period. India now hosts 11 sovereign AI projects—more than any country—propelled by announcements following the February 2026 AI Impact Summit in Delhi. Over the first six months of 2026, India added four infrastructure programs, including one of the country’s largest AI compute clusters, while the Delhi summit showcased Sarvam and Param2, models launched from projects that the index has tracked since 2025.

Infrastructure projects still account for the largest share of sovereign AI initiatives. In the first half of 2026, the share of sovereign AI projects focused on infrastructure—such as AI data centers or compute-access programs—rose to 80 percent, compared to 62 percent over the previous six months. The regional diversity of the programs also expanded, with Ghana and South Africa launching Sub-Saharan Africa’s first compute-focused projects tracked by the index.

Sovereign AI Infrastructure Projects Pull Away

Source: CNAS Sovereign AI Index, Version 2.0

Data projects, which build national datasets for model training or platforms for data-sharing, remain the smallest category by far, accounting for just 5 percent of new projects in the first half of 2026. Just one in five of all governments tracked by the index are backing projects with an explicit data component.

Sovereign AI model projects still rely largely on American foundations, but the ground is shifting. Most countries recognize that building a competitive sovereign model from scratch is unrealistic. Most model projects with disclosed bases are, in fact, fine-tuned versions of foreign open-weight models. Three-fifths of these foreign bases are American. As of mid-2026, Meta’s Llama remains the most-used base model for tracked model projects. France’s Mistral and Google’s Gemma tie for the second-most common base model.

Most Sovereign AI Base Models are American

Source: CNAS Sovereign AI Index, Version 2.0

Continued preference for U.S. base models, however, is not assured. Chinese models were largely absent before 2025 but now appear in six countries through two labs. Alibaba’s Qwen appears in Egypt, Singapore, Thailand, Uganda, and the United Arab Emirates, and Kazakhstan’s national model, AlemLLM, was built with direct technical support from 01.AI, reportedly on a proprietary Chinese base.

A few countries are building sovereign AI models from scratch. Despite the technical and financial hurdles, more countries are training their own models independent of a foreign base because they worry they may become cut off from foreign providers. The total number of “from scratch” sovereign AI models tracked by the index has roughly doubled every year and now represents about 40 percent of all model projects—a new high.

To be clear, all of these “from scratch” sovereign AI models lag far behind the frontier of capabilities. About two-thirds have fewer than 70 billion parameters—the weights a model learns during training that shape its outputs—compared to the 2.8 trillion parameter Kimi K3 from China’s Moonshot AI. But sub-frontier does not mean useless: Countries are applying them in specific applications where their capabilities prove adequate. Take Switzerland’s Apertus model, which it trained from scratch. Even though it performs below the frontier, the Canton of Ticino fine-tuned it for a model that can translate confidential documents better than Meta’s Llama Scout. This case captures the trade-off in many sovereign model projects: accepting diminished capability for domestic control.

Few countries have sovereign AI projects across the stack. Most countries are building compute. Some are building models. Almost no one is building the full stack. Only a handful of jurisdictions—such as India, South Korea, the United Kingdom, and Taiwan—have operational programs across all three layers of compute, models, and data. There may be a growing divergence between the full-stack sovereign AI players and a much larger cohort of countries with one-off projects at just one or two layers of the stack, not all of which are even operational.

Sovereign AI expands at the sub-national level. Many of the most prominent efforts to strengthen national AI sovereignty are happening at the sub-national level. The index has tracked sub-national programs dating to 2023, when Spain’s Catalonia launched a regional language model and Germany’s Hesse launched a regional AI supercomputer. The first half of 2026 saw six new sub-national projects in five regions—Bavaria in Germany, Tamil Nadu and Odisha in India, Sarawak in Malaysia, and Ulsan in South Korea. As many sub-national sovereign AI projects were added in the first half of 2026 as in the previous three years combined.

Critical industries are embracing sovereign AI. Industry-specific projects began appearing in 2025—a Turkish program funding sector-specific models for health and finance, a French defense AI program—and became more common in 2026. The first half of this year saw a shipbuilding foundation model in South Korea, a financial sector model in Taiwan, and a defense data platform in Ukraine. Critical industries with proprietary data, heavy regulation, or national-security significance have strong incentives to reduce foreign dependence and maximize control and resilience. The result is growing sovereign AI stacks for critical infrastructure.

Stepping back, the global push for sovereign AI shows few signs of slowing. Given this, Washington should recognize that full adoption of the American AI stack abroad is neither realistic nor wise as a policy goal. No country–not even America’s closest allies–will accept that level of dependence. Instead, the right goal for America is to become the world’s sovereign partner of choice, specifically, by helping to shape emerging sovereign AI strategies and building partners’ AI capability through U.S. technology instead of around it. Above all, Washington must mitigate the anxieties fueling the global push for sovereign AI by making access to the American AI stack more predictable, transparent, and reliable.

The second wave of sovereign AI is coming into view; Washington would do well to ride it.

Pablo Chavez is an adjunct senior fellow for the Technology and National Security Program; Vivek Chilukuri is the director of the Technology and National Security Program; and Ruby Scanlon is a research associate on the Technology and National Security Program.