September has brought fresh momentum to AI beyond the US and China. On September 3, Saudi Arabia’s HUMAIN, backed by the Public Investment Fund (PIF), unveiled humain-m3, an Arabic-language model it commissioned from MiniMax. Days later, France’s Mistral AI raised €3 billion at a valuation exceeding €21 billion, funding model development and computing infrastructure. The Center for a New American Security (CNAS) calls this sovereign AI’s second wave.
Coverage of AI as a US–China contest often overlooks what everyone else wants. Governments increasingly hope domestic companies and infrastructure can reduce their dependence on foreign suppliers. The EU has proposed its European Technological Sovereignty Package. Canada has a sovereign computing strategy. Chile is leading the regional Latam-GPT initiative.
An international wave of AI nationalism is taking shape. Governments want personal data kept within their borders and models that understand local languages, laws and norms. They also fear that AI will reinforce the internet economy’s concentration of wealth and power.
Yet existing projects reveal a paradox. Sovereign AI can give countries more control while embedding American and Chinese technology more deeply in their economies. America’s approach is more forceful; China’s often involves a softer sell. Neither delivers complete independence, but partial autonomy still matters when access to AI can become a geopolitical bargaining chip.
Three questions explain the tension:
What drives sovereign AI: language, the internet and economic influence.
What countries are building: customized models and data centers.
How influence spreads: America’s harder approach and China’s softer one.
Why Sovereign AI? Language, the Internet and Economic Influence
Nvidia CEO Jensen Huang helped popularize sovereign AI as a market for chips and servers. It has since become a political and security priority. Before examining those ambitions, though, consider a simpler question: Why would a country need an AI model of its own?
Often, the answer is language.
Japan’s Sakana AI develops technology adapted to Japanese needs. Mistral supports French; HUMAIN targets Arabic speakers. India’s Sarvam AI builds for Hindi and other Indian languages.
General-purpose models have historically relied heavily on English-language training data. Their apparent fluency elsewhere can conceal awkward phrasing and gaps in cultural understanding. Chinese users who tried ChatGPT in 2023 immediately noticed that its Chinese answers felt like Chinese written with English logic. Each paragraph made sense on its own, but the full response felt deeply strange.
The problem extends beyond translation. A shortage of high-quality local material can leave models poorly equipped to interpret legal terms, everyday assumptions or regional references. More speakers do not automatically mean more usable training data.
Web scraping alone cannot close every gap. Countries need access to local archives, institutional records and specialized datasets, alongside infrastructure for processing sensitive material under domestic rules. What I think of as a country’s “sovereign internet”—its language communities and locally generated information—becomes a resource for sovereign AI. Common Crawl’s language statistics illustrate how unevenly the public web represents different languages.
The larger motivation is economic. Many governments watched the internet boom enrich foreign platforms without producing comparable champions at home. They see AI as another chance to secure a greater share of digital activity.
The US and China dominate the platform economy. Japan, South Korea and India have substantial domestic businesses, but fewer platforms with comparable international reach. That imbalance now shapes who supplies AI—and who can withdraw it.
June’s US restrictions on Anthropic’s newest models, followed by Reuters’ July report that Beijing was considering restrictions of its own, made that vulnerability tangible. Chip controls add another layer.
The common objective is straightforward: recover some control over how models are developed, supplied and used.
What Countries Are Building: Customized Models and Data Centers
By mid-2026, CNAS tracked sovereign AI initiatives in 67 countries and the EU, up from 16 governments in 2023. New participants included Cambodia, Egypt and Pakistan. This was no longer exclusively a wealthy-country project.
A national model is often the most accessible starting point.
On September 2, Spain’s Multiverse Computing launched Quasar 438B, an English–Spanish reasoning model with a 1M-token context window. Its announcement cited a score of 43 on Artificial Analysis’s Intelligence Index v4.1.1, claiming the highest European result in that comparison. But Quasar is based on GLM-5.2, developed by China’s Z.ai. Its contribution is compression and more efficient deployment, not an entirely independent foundation.
Other projects follow a similar pattern. Japan’s Rakuten AI 3.0 uses the DeepSeek-V3 architecture. AI Singapore’s Qwen-SEA-LION-v4 builds on Alibaba’s Qwen.
CNAS finds that most tracked model projects with disclosed foundations adapt foreign open-weight models. Meta’s Llama remained the most common base by mid-2026, followed jointly by Mistral and Google’s Gemma.
This is a practical bargain. Adapting an existing model costs less than training from scratch. Deploying its weights locally also reduces exposure to a foreign provider withdrawing API access.
But models still need chips and electricity. Infrastructure accounted for 80% of new sovereign AI projects tracked by CNAS in the first half of 2026. India added four infrastructure programs. Elsewhere, governments are exploring alternatives to American accelerators, including Huawei’s Ascend family.
Local models and computing facilities become particularly valuable in industries handling proprietary data, operating under strict regulation or serving national-security needs. South Korea’s shipbuilding sector, French defense initiatives and Germany’s industrial AI projects illustrate the opportunity. Deutsche Telekom’s Industrial AI Cloud began operating on February 4, built with Nvidia and data-center partner Polarise.
These countries are not necessarily trying to reproduce an entire foreign AI industry. They want more control over the systems that matter most.
Partnerships can help. Under Mistral’s agreement with HUMAIN, the companies plan to combine Saudi infrastructure with model development, including stronger Arabic capabilities and deployments in regulated sectors. Saudi Arabia is using capital and infrastructure to diversify its suppliers beyond the two dominant AI powers.
Existing sovereign AI projects still depend on two critical elements: U.S.-dominated compute chips and open-source model architectures developed in China.
Countries pursuing sovereign AI may have specialized data, power resources and data center clusters, but they have yet to build a technology stack fully independent of either China or the U.S.
How Influence Spreads: America’s Harder Approach, China’s Softer Sell
American firms occupy commanding positions in frontier models, AI accelerators and cloud infrastructure. Until recently, governments could treat that dominance as a relatively stable foundation for their own projects.
The Fable episode challenged that assumption. On June 12, US controls forced Anthropic to suspend access to Fable 5 and Mythos 5. Those controls were lifted on June 30, but the interruption showed that even allies could lose access because of decisions made in Washington.
Other incidents reinforced the concern. AP reported that US sanctions disrupted ICC prosecutor Karim Khan’s Microsoft email access; Microsoft denied blocking his account. The disputed episode nevertheless exposed anxiety about dependence on American services.
The Economist argues that independence from both AI powers is unrealistic. Advanced hardware is expensive and ages quickly. Governments risk spending billions on facilities whose equipment needs replacing before their ambitions are realized.
But the impossibility of complete independence is not a persuasive argument against partial control.
Sovereignty is not a switch between 0% and 100%. Germany, Japan or India can buy Nvidia chips while retaining control over models trained on domestic defense or financial data. That is also the promise behind Huang’s AI factories: infrastructure located within a country and serving its priorities.
CNAS recommends that Washington become the preferred partner for sovereign AI, helping countries build capabilities with American technology without simply deepening their dependence.
That is still a route to influence. A remotely controlled model API is a harder sell to sovereignty-conscious governments than a system they can operate locally. Suppliers offering Palantir-style integration and deployment may therefore find a receptive market. Nvidia can encourage national open-model projects while keeping their developers within its hardware and software ecosystem.
Financing strengthens that position. Huang has estimated a fully equipped 1GW AI factory at roughly $50–60 billion. That is not the cost of an ordinary data-center building, but it indicates the scale of frontier infrastructure.
Smaller projects also depend on foreign finance. Côte d’Ivoire’s planned national data center is being built by Washington-based Cybastion. The Export-Import Bank of the United States approved a $66.1 million guarantee for its construction.
China’s approach often emphasizes adapting technology to the customer’s existing system rather than selling an entire American-style stack.
Consider 01.AI. Once identified primarily with foundation-model competition, it now emphasizes enterprise deployments and technical services. Its work on Kazakhstan’s AlemLLM offers a useful example: a national AI project supported by Chinese developers. Founder Kai-Fu Lee is selling expertise, not merely access to a remotely hosted chatbot.
Open weights offer another route. Policy researcher Pranay Kotasthane describes an Indian financial company running DeepSeek on rented GPUs in India. The weights reside locally; inference need not send customer data to China. The original developer does not have an API connection, it can simply revoke.
Think of owning a movie on disc rather than depending on a streaming subscription. It may seem less convenient, but the film remains available if the service disappears. Security reviews are still necessary; local deployment does not make a model inherently trustworthy.
This delivery model fits China’s preferred language of noninterference: provide a working system without demanding that the customer reorganize its institutions around the supplier.
Weather forecasting shows the appeal. At the 2025 World Artificial Intelligence Conference, China introduced MAZU, its Early Warnings for All solution, named after the maritime guardian goddess. An Egypt-specific version under development includes aviation weather services and sandstorm warnings for tourism.
Such applications offer tangible benefits without requiring a government to hand over its most sensitive decision-making systems. They can build trust in Chinese suppliers and potentially create demand for Chinese hardware.
That next step remains uncertain. Huawei is courting Egypt, while Malaysia is considering Ascend 910C chips. Neither interest nor a proposal equals completed deployment. Domestic demand and manufacturing constraints also limit how quickly Chinese suppliers can expand overseas.
Conclusion
Sovereignty is less about independence than control. Whether through chips, models or infrastructure, American and Chinese technology remains embedded in projects intended to reduce dependence on both.
China’s present openness is not a permanent guarantee. Reuters’ reporting on possible frontier-model restrictions concerns deliberations, not an announced blanket ban. But the possibility matters.
In that uncertainty, time becomes an asset. Locally installed accelerators, adaptable models and the expertise to operate them give countries a buffer against future restrictions. They preserve room to act if AI competition divides more sharply into rival blocs.
Sovereign AI does not offer an escape from the US–China rivalry. It offers a chance to retain some agency within it—a need that may become more urgent than it was in the internet era.
