Generated by Codex with GPT 5.6 Sol XHigh

Artificial-intelligence sovereignty sounds like national control over models, chips and computing infrastructure. For most countries, that goal is unrealistic. The frontier is increasingly dominated by America and China, whose governments can shape not only how advanced models are made safe, but also who is allowed to use them. The practical challenge for everyone else is therefore to reduce vulnerability without wasting fortunes on an impossible race for total independence.

Safety rules can also become instruments of power

The immediate debate concerns regulation. The newest AI models may be capable of helping users attack critical computer systems or design biological weapons, yet America still controls access through improvised decisions. Google DeepMind chief Demis Hassabis proposes a public-private regulator modelled on the body that oversees American securities markets. Clear rules would make oversight more predictable, and an American system might eventually become an international standard.

But safety is only part of the issue. America has already used allies’ military dependence as negotiating leverage, and it could eventually do the same with AI. China has treated exports such as rare earths as strategic tools and is likely to view advanced models similarly. As AI becomes embedded in factories, public services and defence, losing access to a foreign model or data centre could become an economic and security crisis.

Trying to reproduce the American frontier is not a credible response. OpenAI and Anthropic have each raised well over 100bn dollars, and success depends on scarce researchers, costly experiments and a willingness to make risky bets. Governments have a poor record with that kind of research. A prestige project that merely imitates the leaders would consume enormous resources without providing real autonomy.

Resilience matters more than self-sufficiency

Countries can still protect themselves by building enough domestic computing capacity to handle sensitive data and keep essential systems running. Local data centres would also preserve the option of switching to open-weight models if access to frontier systems were cut off. The biggest barriers are often not subsidies but slow planning processes, inadequate electricity supplies and long waits for grid connections.

Those delays are expensive. A nine-month postponement can damage a data centre’s economics as much as doubling its lifetime electricity bill. Grid connections take about two years in America, three years in Britain and India, and three and a half in Germany and South Korea. Faster approvals, priority connection queues and permission to generate power outside the grid would do more than elaborate national AI branding.

Market access can create leverage too. Governments could trade permission to build data centres for guarantees that local users receive the same models as Americans. Encouraging businesses to adopt American AI would make foreign customers more valuable to model developers, giving those companies a reason to resist protectionist restrictions that threaten their revenue.

Some countries possess additional bargaining chips. Taiwan dominates advanced chipmaking, the Netherlands is home to ASML’s essential lithography equipment, and South Korea is strong in memory chips. These specialisms should be protected, but even their owners cannot assemble a fully sovereign AI industry.

The article’s central lesson is that sovereignty should mean optionality rather than isolation. America is a more attractive partner than authoritarian China, but dependence on any superpower carries risk. Countries that secure local compute, preserve open-model fallbacks, remove infrastructure bottlenecks and cultivate strategic specialisms will not become independent of the AI giants. They will, however, be harder to coerce and better able to negotiate.