AI Energy and Sovereignty: What Governments Need to Plan Now

Talking about AI sovereignty without talking about energy is like discussing food sovereignty while ignoring soil. A country can have the best researchers, the most sophisticated algorithms and a well-designed industrial policy — and still be completely dependent on another nation to run anything at scale. The computational power that underpins modern AI models does not exist in the ether. It exists in racks of servers that consume electricity in volumes that rivaled, ten years ago, that of medium-sized cities.

The logical chain that national strategies ignore

All AI capabilities are ultimately reduced to a question of computation. And computing at scale comes down to a question of energy. A frontier language model like GPT-4 or Claude, during training, consumes energy equivalent to the annual residential consumption of hundreds of families. Inference — the act of answering a question, generating text, processing an image — is most efficient per call, but multiplies across billions of daily requests. The total is not trivial.

The strategic implication is straightforward: a country that does not control the energy layer does not control the physical layer of AI, regardless of what it does in the upper layers. It can regulate, it can finance startups, it can create national laboratories. But if the data centers running the models depend on imported energy, the price and availability of which are determined by third parties, the entire architecture of sovereignty has a fissure at its foundation.

The double exposure of dependent countries

Most countries that today talk about a “national AI strategy” actually face two overlapping structural problems. The first is energy dependence: countries that import hydrocarbons to generate electricity have their computing costs linked to the price of oil and natural gas in international markets — markets subject to geopolitical shocks, sanctions and cartel decisions. The second is the dependence on cloud infrastructure: most of the computing capacity available globally is operated by three American companies and, to a lesser extent, by one Chinese company.

When the two problems combine, the result is a compound vulnerability. An energy crisis increases the cost of running local models. A geopolitical crisis could restrict foreign cloud access. A scenario in which both occur simultaneously — and recent history in Europe in 2022 has shown that this is possible — can paralyze entire digital infrastructures without any adversary needing to directly attack a server.

What separates countries that take energy seriously

Some countries realized the logical chain early and built their AI strategies with energy as a central variable, not as an afterthought. France is the most obvious case: decades of investment in nuclear energy have produced an electrical matrix with low marginal cost and high predictability. When France decided to bet on AI sovereignty, the energy infrastructure was already there. The problem was different—and more manageable.

The UAE has built an even more deliberate position. The Abu Dhabi Investment Authority and G42 have developed parallel energy and data infrastructure strategies, with data centers powered by local renewable energy and GPU capacity agreements negotiated directly with manufacturers. Saudi Arabia, with the NEOM Project and Saudi Aramco's investments in technology, follows a similar trajectory. What these countries have in common is not the energy model — nuclear, solar, sovereign fossil fuel — but the decision to treat energy and computing as integrated variables of the same industrial policy, not as separate ministries that never talk to each other.

The Brazilian case: real advantage, real problem

Brazil enters this conversation with a genuine advantage that few countries have: more than 85% of its electrical matrix generated by renewable sources, with a predominance of water complemented by expanding wind and solar capacity. This means that a data center built in Brazil has, at its base, one of the cleanest matrices in the world — and, structurally, less exposure to fossil fuel price shocks.

The problem is not in the power source. It is in the geographic concentration of the infrastructure that uses it. Almost all of the installed capacity of data centers in Brazil is located in the São Paulo–Rio de Janeiro corridor. There are historical reasons for this — proximity to the largest consumer markets, presence of submarine cables, availability of technical labor — but the result is structural fragility. An interruption in the transmission system that supplies this corridor, whether due to an extreme weather event or systemic failure, has the capacity to disproportionately impact the national digital infrastructure. Furthermore, the operational dependence on large American clouds — AWS, Azure and Google Cloud dominate the Brazilian corporate market — means that even data processed within the national territory often depends on control, authentication and orchestration systems that reside outside of it.

What digital infrastructure decision makers need to understand before 2030

The issue for public policy decision makers is not simply "build more data centers." Building more data centers in the same corridor, with the same dependence on foreign cloud providers, solves the capacity problem without solving the sovereignty problem. The right question is another: how to distribute computing infrastructure across the national energy matrix so that cost, resilience and control improve simultaneously?

This implies specific regulatory decisions. Encouraging data centers in regions with a surplus of renewable generation — the Brazilian Northeast has solar irradiation among the highest in the world and underutilized wind capacity — requires a combination of tariff policy, regulatory frameworks for long-term energy contracts and investment in transmission. It also involves reviewing the legal framework so that companies that process data from Brazilian citizens cannot do so exclusively in infrastructure whose final control is in another country. And it implies treating technical qualifications in data center operations as part of the professional education policy, not as an externality that the market will resolve on its own.