Jul 16th 2026|12 min read
Listen to this story
Your browser does not support the element.
https://infographics.economist.com/2022/video-iframe-player/autoplay.html?key=graphics/Aiflag&size=1920&size=1440&size=1280&size=896&format=mp4&format=webm&credit=John Provencher
WHEN THE leaders of the biggest Western economies gathered last month in the French Alps for the G7 summit, it was not just presidents and prime ministers in attendance. The bosses of America’s leading artificial-intelligence firms—Sam Altman of OpenAI, Sir Demis Hassabis of Google DeepMind and Dario Amodei of Anthropic—were seated alongside the political bigwigs. Even the stewards of the mightiest economies, it turns out, are buttering up the potentates of AI.
Days before the summit, America’s government had barred Anthropic from making its most advanced model available to foreigners. OpenAI soon imposed similar restrictions on some of its frontier systems_._ The move turned a simmering concern into a burning one: might countries without their own, home-grown AI firms end up excluded from the world-changing technology? Emmanuel Macron, France’s president, warned that no one would buy American AI if it could “turn off the switch” at will. The American AI labs’ Chinese rivals offer only faint reassurance. On July 7th Reuters reported that Chinese officials, too, were considering restricting foreigners’ access to leading AI models.
Caught between the world’s two AI superpowers, governments are increasingly pinning their hopes on “sovereign AI”: the idea that countries should limit their dependence on foreign providers of AI by actively developing domestic AI firms and infrastructure. Announcements of such initiatives are proliferating. The European Union has unveiled a technology-sovereignty package encompassing semiconductors, AI and cloud computing. Canada has launched its own plan to reduce reliance on American technology providers. India, Japan and Singapore are backing domestic AI infrastructure and models. The Centre for a New American Security (CNAS), a think-tank in Washington, estimates that the number of state-backed AI projects outside America and China grew fivefold over the course of 2024 and 2025 (see chart 1). Governments have announced investments of $70bn-plus in such initiatives, according to CNAS’s tally.
Chart: The Economist
Yet turning these ambitions into reality will be extremely difficult. No matter what, most countries will still rely on American chips, Chinese open-source models or both to build their AI systems. Even then, most governments will find it prohibitively expensive to achieve anything remotely resembling AI sovereignty.
The incentive to seek it, however, is clear and compelling. Pablo Chavez of CNAS puts the restrictions on Anthropic in a “long line of events” demonstrating America’s willingness to use its technological dominion as a weapon. In the dying days of Joe Biden’s presidency in 2025, America proposed sweeping export controls that would have rationed access to advanced chips through a complex licensing regime. When Donald Trump became president a few weeks later he dropped the plan. But he has not hesitated to use America’s lead in AI as leverage in trade negotiations. The Department of Commerce is considering rules that would allow it to vet all sales of AI chips designed by American firms anywhere in the world. It might also insist that buyers allow inspections or monitoring to make sure chips are being used as promised. Another possible requirement could be to invest in American AI infrastructure in exchange for access.
Protecting themselves from coercion is not the only reason governments want to build their own “AI stack”, the layers of hardware and software on which AI depends. National security is another: it is awkward to use foreign, commercial AI for top-secret purposes, especially of a military nature. Governments also like the idea of models trained on their own languages, laws and norms. They worry, with some justification, about users’ sensitive, personal data falling into foreign hands. And policymakers want to reap the economic benefits of AI. They have seen how previous technological innovations—the internet, smartphones and cloud computing—entranced consumers and reshaped their economies while generating riches for American tech firms. They are keen to avoid a repeat.
The Anthropic cut-off was instructive, says Arthur Mensch of Mistral AI, a French model-maker. In an interview for _The Economist’_s Inside Tech video series, he argues that the imperative to develop sovereign AI runs much deeper than the risk of diplomatic strong-arming. The scale of the outflow of cash to America and China in a world where the pair continue to dominate AI would be so vast it would create “enormous instabilities”, he argues. Countries will find that unacceptable and so will be forced to invest in the ability to serve at least some demand for AI domestically.
Mr Mensch estimates (conservatively, he says) that within the next five or ten years, the European Union will be spending the equivalent of a tenth of its current bill for labour on AI tokens, or about around €1trn ($1.1trn) a year. If things stay as they are, the vast majority of that expenditure will flow directly to American cloud providers. Such outflows would hit the euro, increasing the cost yet more. That, he believes, would create an opportunity for a local enterprise to compete directly, even if its product is not of the same quality. Every country, he predicts, will need to make sure that whatever happens, “The AI can continue to run.”
Sovereignty looks different at each layer of the stack. Start at the bottom: computing power (“compute”, in the jargon), largely in the form of AI chips. The market for those chips is dominated by Nvidia, whose processors account for two-thirds of the world’s AI compute. Chips designed by Google, AMD and Amazon, all American firms, make up much of the remainder.
The alternatives are limited. Chinese firms such as Huawei and Cambricon have made impressive strides despite being cut off by American trade restrictions from the most advanced chipmaking equipment. But they remain some distance from the frontier. Huawei’s Ascend 910C, one of China’s most advanced AI chips, provides roughly a fifth of the performance of Nvidia’s latest offering. Performance is not the only constraint. Chinese manufacturers are prioritising domestic customers, and demand still exceeds available capacity. Even if Chinese chips continue to improve, shortages are likely to persist for years.
Aiming for self-sufficiency in hardware is therefore unrealistic. Kevin Xu of Interconnected Capital, an AI-focused hedge fund, argues that sovereignty should be understood less as independence than as control. If the machines being used to run AI models sit within a country’s borders, governments retain a degree of autonomy even if future access to chips is restricted. “Is your factory going to still run if the US government decides that suddenly you can’t?” asks Mr Mensch. The latest AI chips have a lifespan of at least five or six years.
Chips are only part of the equation, however. Running AI places huge demands on the electricity grid. According to the International Energy Agency, an intergovernmental forecaster, a typical AI data centre consumes as much electricity as 100,000 households; the largest may use around 20 times as much.