AI that answers to your country
Every government will use AI. The question is whose. A rented model carries its makers' values, keeps your words under another country's law, and can be withdrawn from abroad. Sovereign AI means your country runs, tests and answers for the models it relies on.
A country that only rents AI pays every year and learns little. The people who can run, test and fix these systems learn by doing it, and they move to where the computers are. A national service gives them work worth staying for, and gives ministries people who understand what they buy.
of the world's shared data-centre capacity is in low-income countries. High-income countries hold 77%.
World Bank data, June 2025, in UN report A/80/817A model that drafts your position shapes it. People who write with a slanted AI assistant move towards its view, and most never notice. Every text that leaves the delegation needs an owner with a mandate, who can see how the model got there.
or fewer noticed the slant in an AI writing assistant. Their own views still moved towards it, and warnings did not stop that.
Williams-Ceci et al., Science Advances, March 2026A model learns its values from its data and its makers. In a 2023 test, chatbots asked about environmental problems cited more authors from the Global North than the South in all 1,512 answers. China requires its models to uphold "Core Socialist Values". US agencies may only buy models that meet Washington's "Unbiased AI Principles". Test a model against your own positions before it helps you negotiate.
leading AI models advised more escalation for the US, UK and France than for China or Russia, in the same crisis scenarios.
CSIS Futures Lab, Critical Foreign Policy Decisions benchmark, 2025Attackers now use AI agents against ministries, embassies and the hotel Wi-Fi delegates use. Defending your systems will take AI that reads your own records today, and the code that runs your grid, water and elections tomorrow. Give that access only to a model that runs under your law. Abroad, access to the most capable cyber model is vetted: Anthropic reviews who may use it on grids and government networks "in collaboration with the US government".
national identity records taken from a North African government agency by a spy group that used AI at every stage.
Anthropic threat intelligence report, September 2026What a delegation types into a foreign AI service is stored under that country's law. US providers can be compelled through US legal process to hand over data wherever they store it, and a judge can bar them from telling you. In 2025 a US court made OpenAI keep users' deleted ChatGPT chats for over four months. A model you host answers to your courts.
said Microsoft France's legal director under oath, asked to guarantee that French data would never reach US authorities without French consent. He added that it had never happened yet.
French Senate inquiry on public procurement, 10 June 2025Access to a capable AI model should be a basic right, and nobody should pay for it with their data. ChatGPT now shows ads matched to your conversation on its free and cheapest plans in over 60 countries. Public models show another way: Switzerland publishes Apertus with its weights and training data, and publicai.co offers it free without training on chats.
GPUs that India rents to startups, researchers and public bodies at about 65 rupees an hour each.
Government of India, answer to the Lok Sabha, 1 April 2026Run open models for government work on national or regional machines. Start with translating and drafting anything not yet public. Field sheet 01 shows the hardware and what it costs.
Train people, then hire them. Fund fellowships tied to public service, and let the fellows run the systems they trained on.
Buy with conditions. Data stays in the country, no training on government data, a right to audit, and a way out. Test every model against your own positions.
Offer citizens a public AI service that is free or cheap and keeps their conversations private. Pool computers with neighbours where one country is too small.
Back it at the UN. The Secretary-General proposes a Global Fund for AI Capacity-Building of about $3 billion. Resolution 78/311 already asks for technology transfer and open-source AI.
Never paste a position that is not yet public, instructions from your capital, or anything said in confidence into an AI service your country does not control. Use online models for public material and practice.
The best open models still trail the strongest closed ones on the hardest tasks. A national model can carry its own bias: in one 2026 test, models built in a country predicted its UN votes no better than any other. You choose, test and answer for the model, whoever built it.
Sources