Late in 2023 Jensen Huang, chief executive of Nvidia, began peddling a new idea. Every country, he said, should have its own artificial-intelligence (AI) system, trained on domestic data, aligned with national values and built using local infrastructure. Appealing to policymakers’ fondness for manufacturing, the boss of the chip colossus described these systems as “AI factories”, ingesting data and churning out intelligence. He called it “sovereign AI”.
Politicians are warming to the idea. Earlier this year the European Commission unveiled plans for a €20bn ($23bn) fund to build up to five “AI gigafactories”. In just the past two months the governments of France, Germany, Saudi Arabia, South Korea and the United Arab Emirates (uae) have all been involved in deals to build local AI infrastructure (see chart). According to Nvidia, at least 20 countries are pursuing the idea of sovereign AI.
The reason for Mr Huang’s enthusiasm is clear. For Nvidia, which this month became the first company to reach a $4trn market value, governments are a potentially lucrative source of business. Jefferies, an investment bank, estimates that sovereign initiatives could generate some $200bn in cumulative revenue for the chipmaker “over the coming years”; the company believes that spending could reach $1trn over an equally fuzzy period.
Sovereign ai could also act as a welcome hedge for Mr Huang. His biggest customers—Amazon, Alphabet, Meta and Microsoft—are all developing their own chips, which could in time reduce their reliance on Nvidia. Saudi Arabia expects to purchase “several hundred thousand” of Nvidia’s top-end processors over the next five years. The UAE, with perhaps the most ambitious plans of all, intends to import half a million annually.
But will sovereign ai pay off for taxpayers? Despite growing enthusiasm, the concept remains woolly. Sovereign AI may help countries develop national models, protect sensitive information and widen access to the technology. But it will do little to reduce countries’ dependence on America and, in many cases, risks wasting vast sums of money.
There are subtle variations in how governments are pursuing sovereign AI. France’s approach centres on Mistral, a domestic model-maker, which has formed a consortium with Bpifrance, a government bank, MGX, an Emirati state-owned investment firm, and Nvidia to build what has been touted as Europe’s largest AI data campus. In the Gulf, governments have taken more direct control. In May Saudi Arabia launched Humain, a firm tasked with building AI infrastructure in the kingdom. In the UAE that role has fallen to G42, an AI firm part-owned by Mubadala, a sovereign-wealth fund.
Governments justify these projects on various grounds. A few big spenders want to catch up to America: the European Commission, for instance, hopes to propel Europe to the “forefront of AI development”. Others, such as India, worry that AI models trained on foreign, and especially English-language, data will not reflect local languages and values, a growing concern as AI systems come to shape everything from education to public services. A degree of control over domestic data is also viewed as essential. In areas such as health care, officials worry about patients’ information being plugged into foreign models. A homegrown system, some contend, makes it easier to protect such information while still allowing it to be used to train AI models.
A final concern for governments is to widen access to AI systems. Nadia Carlsten, chief executive of the DCAI, which runs Gefion, Denmark’s national AI supercomputer, says that smaller companies and research institutes are “always at the back of the line”. A domestic offering, she argues, helps ensure cheaper and more reliable access for such users. Gefion, launched in 2024, is being used for applications such as drug discovery and weather forecasting.
Yet these initiatives will not offer anything close to self-sufficiency. Cutting-edge processors are the most important requirement for advanced AI systems. Nvidia dominates that market, accounting for about 90% of all commercially available AI chips—which is why it has played a central role in nearly all sovereign-AI initiatives. Its only serious rival, AMD, is also American. The servers that house these chips are mostly built by another two American companies, Dell and Supermicro. Even China, which has built something close to a self-sufficient AI stack, has yet to develop a domestic alternative to America’s whizziest chips.
Sovereign-AI projects may well help governments achieve more modest objectives, such as keeping data secure and ensuring access to the technology for groups such as researchers. But America’s cloud giants could probably provide this more efficiently—not least because their scale gives them bargaining power over suppliers such as Nvidia. The chipmaker disagrees, but Amazon and Microsoft are already pitching sovereign clouds with enhanced controls over data and dedicated local infrastructure. National AI models could simply be built on top of these.
Kevin Xu of Interconnected Capital, a hedge fund, warns that many sovereign-AI projects risk creating something “more like a palace than a factory”. That would be a poor use of taxpayers’ money—but it might suit Nvidia just fine.
To stay on top of the biggest stories in business and technology, sign up to the Bottom Line, our weekly subscriber-only newsletter.
