TL;DR. Whoever controls the most capable models controls a growing share of how an economy produces, how a society is informed, and how a state coerces. The concentration risk that AI safety worries about — a small group using frontier systems to seize durable power — won't debut in Washington or Beijing. It will debut in countries that consume frontier AI without producing it, where the guardrails that protect people at home don't reach. Developing countries are the leading edge of this threat. Watching what happens to them is how the rest of the world learns what's coming for it.
As gradual disempowerment dynamics take place across the world, developing countries can end up locked in to frontier models built in the US or China, ceding autonomy over strategic capabilities (from the economy to the military) with no capacity to reverse the dependence or course-correct what those systems do.
There's historical precedent for powerful companies, based in superpower countries, leading or facilitating power concentration in poorer nations. The United Fruit Company's land holdings in Guatemala directly fed a power grab there (the case that gave us the term "banana republic") while being perfectly legal in the US. Similarly, the Wall Street Journal has reported on Huawei technicians helping Uganda and Zambia track opposition leaders.
Even if a US lab fully complied with every political and legal standard in its own country, that compliance is defined by that country. A model that refuses to help a coup in Washington can still enable one abroad. Anthropic's standoff with the Department of War/Defence over mass surveillance made the line very much clear: the protection it insisted on for Americans didn't extend to people elsewhere.
Geopolitical competition, a rush for critical minerals, or the pursuit of military advantage can all push a lab or a foreign government to seize power from abroad, or to work with a small domestic group to grab it in exchange for mineral concessions, market access, or political alignment.
Then everywhere
Treat the developing world as the frontier's testing ground. Techniques that work - single-actor controllability over critical systems, the automation of coercion and surveillance, access to hyper-advanced political strategy tools and intelligence, get proven somewhere with weak guardrails and high stakes before they're mature enough to deploy against their home country. The mechanism normalizes. The playbook gets written. After the coups, the aligned governments vote alongside their backers in international fora.
The jurisdictional wall that protects citizens of frontier countries is thinner than it looks: the same capabilities, once they exist and once someone has run them at scale, don't respect the border they were first used across.
