Gilles Chemla, ³Ô¹ÏºÚÁÏ College Business School and Harvard University; Tahmid Quddus Islam, Consultant and Independent Advisor; Simeng Li, ³Ô¹ÏºÚÁÏ College Business School. 'Strategic and Financial Requirements for a Sovereign AI Infrastructure.' Published in IOS Press, 2026.
Problem
As AI becomes critical to national security and economic prosperity, the ability to design and maintain systems under geopolitical stress has become one of the defining strategic challenges of our time. AI now sits inside power grids, hospitals, defence systems and public services. Most governments assume that because they use these systems, they control them, but they don't. The chips that power AI are made by a handful of companies. The cloud platforms that run AI applications are owned by a small number of technology firms. The raw materials that go into the hardware are mined in regions subject to export controls and geopolitical tension. When any one of those links is disrupted, the critical national infrastructure that governments and public services depend on can stop working.
Science
Our research showed governments where a single decision in the AI supply chain by a foreign government could bring critical systems to a halt. We reviewed the full chain of dependencies that underpin modern AI systems and identified six layers where sovereign control matters: physical and energy infrastructure, data and cloud architecture, models and algorithms, financing and investment, supply chain governance, and human capital. Drawing on evidence from geopolitical conflicts, export control regimes and real-world infrastructure failures, including cyberattacks on Ukraine's power grid and US chip export restrictions, we mapped the conditions under which each layer becomes a point of failure. We also examined how countries can fund the transition to greater AI self-sufficiency, including through public-private partnerships, sovereign wealth funds and development finance institutions.
Outcome
A country's AI capability is only as strong as its weakest link. Owning capable AI models means little if the computing power that runs them belongs to a foreign firm, or if the chips that power them can be cut off by a single export control decision. Only a small number of companies globally manufacture the advanced semiconductors that AI systems require. That concentration is a significant vulnerability for critical national infrastructure. We show that when AI infrastructure fails, the people with the fewest alternatives bear the most disruption. Pursuing greater control over AI infrastructure often means accepting higher upfront costs.
Change
We gave governments the first comprehensive map of where their critical AI infrastructure is exposed and what it would cost to protect it. Our research calls on policymakers to diversify the supply of critical materials, build domestic positions in chip manufacturing, create sovereign computing capacity for the most sensitive workloads, and train the scientists and engineers needed to sustain these systems over time. The European Parliament passed a resolution on technological sovereignty and digital infrastructure in January 2026, reflecting the policy questions our research addresses. Our paper was submitted to NATO as the first comprehensive framework for AI sovereignty across energy, hardware, data, finance and human capital, directly informing the strategic policy debate among NATO allies on the resilience of critical AI infrastructure.