Nations are concluding that intelligence is critical infrastructure, and that it should live at home. What sovereign AI programmes mean for where the next gigawatts get built, and who gets to build them.
For a decade the received wisdom was that compute was a commodity, rented by the hour from three companies, located wherever the accountants preferred. That era is closing. Governments have concluded, quickly and almost simultaneously, that the models their economies will run on, the data those models are trained on and the compute beneath both are strategic assets, in the same category as ports, grids and spectrum. From the European Union’s gigafactory programme to national compute plans across the Gulf, South Asia and Latin America, sovereign AI has moved from white paper to procurement.
The instinct is sound. A country that cannot run its own models on its own soil under its own law has outsourced a piece of its future. The question is what sovereignty actually requires, because most of the early answers have been about flags rather than infrastructure.
A sovereign AI capability is not a badge on a rack. It rests on four things, and none of them can be imported at short notice. Jurisdiction: the data, the model weights and the operator sit under domestic law, and stay there. Power: domestic, preferably clean, because sovereignty built on imported gas is a contradiction in terms. People: engineers, operators and researchers who live in the country and intend to stay. And capital patient enough to fund an asset that will still be running in 2050.
“Sovereignty is not a sticker on a server rack. It is who owns the land, who controls the power and whose law applies at three in the morning when something goes wrong.”Jamie MacDonald-Murray, Chairman & CEO, Eppur
Of the four, power is the one that decides geography. A model can be trained anywhere the electrons are cheap and clean. That single fact is redrawing the map.
Compute follows power and policy, not population. The countries and regions pulling ahead are the ones with surplus clean generation, grids with headroom and consenting regimes that can say yes: Iberia and the Nordics, the Baltic states, the Balkans with their hydro and their untapped solar, Brazil with one of the cleanest large grids in the world, Paraguay with Itaipú on its doorstep. Some of these were peripheral to the old internet. They are central to the new one.
Our pipeline is drawn along that map deliberately. Greece and Latvia in the European Union. Albania, North Macedonia and Montenegro in the Balkans. Brazil and Paraguay in South America. Twelve markets, three continents, chosen not because they were fashionable but because each has power, policy and a government that wants intelligence to live at home.
Governments will not build this themselves, and they should not. They will partner with developers who can bring three things at once: secured power, delivery at gigawatt scale, and the credibility to stand beside a ministry for twenty years. Local enough to be trusted, global enough to be bankable. That is a narrow field, and it is the field we have chosen to compete in. We treat sovereign capital and national programmes as equals, not as customers to be managed, because the partnership outlasts any single project.
“The countries that win this decade will not be the ones with the most data. They will be the ones with the cheapest clean electron and a government that can say yes in under a year.”Jamie MacDonald-Murray, Chairman & CEO, Eppur
The geography of compute is being decided now, in grid plans and planning committees and cabinet rooms. The gigawatts will go where they are wanted and where they can be powered. We intend to be there first.
Data Centres are becoming the grid’s largest new customer. The developers who arrive with generation, flexibility and additionality, not just load, will be the ones communities and regulators say yes to.
Read the essay →The AI era wants capacity yesterday, financed for thirty years. How to structure capital that matches the duration of the asset, and why the mismatch is where most developers will fail.
Read the essay →