Sovereign AI: What it means when a country wants to run its own AI

Sovereign AI is a country’s ability to make its own decisions about the AI systems it depends on. That covers the data they learn from, the computers that run them, the models themselves and the rules around them. The idea is control, not isolation: a government wants to be able to keep working if a foreign supplier changes terms.

The term sounds technical, but the worry behind it is old and familiar. Countries already think about who supplies their energy, water and telecom networks.

Owning the power station instead of renting the electricity

Picture a country whose hospitals, banks and ministries all rely on a chatbot hosted abroad. The service works well until the provider raises prices, changes its rules, or is barred from serving that country. Nothing at home can replace it quickly.

Sovereign AI is the attempt to avoid that position. A country builds or secures its own data centers, trains or adopts models it can inspect, keeps sensitive data inside its borders, and writes its own rules. It is closer to owning a power station than to renting electricity, though few countries own all of it.

The layers a country can try to control

AI is a stack of layers, and sovereignty can apply to each one. A 2026 Brookings report by Brooke Tanner, Cameron Kerry, Andrew Wyckoff and colleagues lists minerals, energy, computing hardware, networks, digital infrastructure, data, models and applications. It adds talent and governance as threads running through all of them.

A 2025 arXiv paper by Shalabh Kumar Singh and Shubhashis Sengupta groups the goal into four pillars. The paper’s abstract refers to four pillars without naming them, but the paper describes them as data, compute, model autonomy and norms (the values and rules built into systems).

• Data: who holds national records, texts and languages, and who may train on them.
• Compute: the chips and data centers where models are trained and run.
• Models: whether a country can build or adapt its own, for example in its own language.
• Norms: whose rules and values shape what the systems do.

How a chip maker popularized the phrase

The idea of national control over technology is not new, but the label spread in this AI wave through Nvidia chief executive Jensen Huang. In a February 2024 interview with VentureBeat, he described sovereign AI as countries and companies keeping their data, and the language models built on it, within their own borders. He called data the raw material of AI and said it belongs to whoever owns it.

Huang also said countries will build their own AI infrastructure, which he described as hardware. He calls the facilities that turn data into model output “AI factories”. Nvidia sells the chips these factories use, so the company has a commercial interest in the idea, and readers should weigh his framing with that in mind.

Not every definition matches his. Researchers and think tanks use the term more loosely, and there is no single agreed meaning.

Who is pursuing it, and where Malaysia fits

The Singh and Sengupta paper applies its model to India, which it says has footholds in data, compute and norms but weaker control over models. It also looks at Saudi Arabia and the United Arab Emirates, where it describes large public investment in Arabic-first models and sovereign cloud. The Brookings report calls India a leader among sovereign AI initiatives.

Malaysia is a current example of a country building data center capacity at scale. Our report on SoftBank’s reported $50 billion plan for Malaysia describes a government account of Masayoshi Son’s plans, which SoftBank has not confirmed. The reports describe a goal of becoming an “AI nation” by 2030. They do not label the plan sovereign AI, and the money would come from a foreign company. That is a useful case for the next section.

Why full independence is out of reach

Almost no country can control every layer. The Brookings authors conclude that concentrated choke points across the stack make full-stack sovereignty structurally infeasible for almost any country. Advanced chips come from a small number of suppliers, and the global open-source ecosystem and international standards resist being fenced in.

The Singh and Sengupta paper reaches a similar view: the more a country opens up, the more it gains from scale, talent and shared research, and the more it is exposed to dependence. Their summary is that “sovereignty in AI needs managed interdependence, not isolation.”

Brookings also lists risks of the push itself:

• protectionism and fragmented markets and standards;
• duplicated or stranded public investment, such as data centers that sit underused;
• weaker interoperability between countries’ systems;
• use of sovereign systems for digital authoritarianism.

The Brookings authors propose “managed interdependence”: map dependencies layer by layer, diversify suppliers, and build portability into standards and purchasing.

What to watch when a country announces a sovereign plan

Announcements often blur the layers, so a few questions help. Who owns and operates the data centers, a state body or a foreign company? Where does the data sit, and under whose law? Can the country run the models if a supplier leaves? Is the money committed, or only a stated ambition?

Our Malaysia story shows why the last question matters. The only documented step is a non-binding framework agreement in Sarawak, and the $50 billion figure has no published timeline or financing. Sovereignty depends on what gets built and who controls it, not on the size of a headline number.