Artificial intelligence is often presented as software available everywhere through the same window. But advanced models depend on chips, data centres, energy, networks, cloud services, data and specialised people. A country that controls none of this infrastructure can use AI while remaining dependent on other actors’ industrial and political decisions.
Brazil made that material dimension visible in August 2026 with a package of roughly 2.3 billion reais. It includes a new supercomputer in Rio Grande do Norte, advanced infrastructure in Rio de Janeiro with Huawei and iFlytek, a national cloud, training, technology transfer, open RISC-V chip research and an algorithmic-transparency centre.
The distribution of alliances is striking. Brazil is working with Chinese companies while a separate tender is expected to use US chip technology. Avoiding a single power across the whole stack is an attempt to create room for manoeuvre in a sector dominated by a few firms and fragile supply chains.
Digital sovereignty does not require making every component domestically. No country is fully self-sufficient. It means retaining the capacity to choose, negotiate, replace suppliers, keep strategic data under appropriate control and build skills that survive after a contract ends.
The Macaíba machine is intended for training and running advanced models for government, universities, research and industry. Placing it in the north-east reflects energy, connectivity and cooling requirements, but also a territorial choice to distribute scientific capacity beyond the country’s established centres.
Domestic compute can support models in Portuguese and Spanish and applications in agriculture, climate, health, public services and industry. Language representation is not cosmetic: it shapes which questions are understood and which communities remain peripheral.
Infrastructure alone does not guarantee independence. If hardware, maintenance, foundational software and expertise remain concentrated abroad, a country may own machines without controlling their evolution. Technology-transfer clauses, training and open research can matter more than a ranking headline.
The material cost also needs scrutiny. Systems at this scale require continuous electricity, cooling, water, networks and long-term investment. A national symbol must still disclose consumption, benefits, access rules and results. Sovereignty loses meaning if costs are shifted onto host communities.
A national cloud adds another layer. Keeping strategic public data under closer legal and operational control may reduce some dependencies, but it also concentrates responsibility. Security, audits, resilience and clear limits on model training are essential. Sovereign infrastructure must also be accountable.
Brazil’s initiative shows that the AI race is no longer only about companies and products. Governments are deciding where compute will sit, which languages models will represent, which geopolitical blocs to engage and which capabilities to retain.
Success will not be measured only in petaflops, but in what remains: trained researchers, software, access for universities and businesses, useful public models and the ability to change direction without asking a single supplier for permission. Sovereignty is not isolation; it is the practical capacity not to inherit every choice made elsewhere.