Article published in Revue Défense Nationale, 2026

The race for artificial intelligence is not being decided where most observers are looking. What will shape future power relations is not the complexity of algorithms, but the understanding of the physical system that makes them possible: energy, computing capacity, semiconductors, critical raw materials. Whoever controls compute controls power. And the bottleneck of this race is energy.

AI as a Physical System

Most decision-makers, including those in military headquarters and ministerial offices, still think of AI as an essentially software phenomenon: algorithms, code, language models. In reality, it is first and foremost a physical and industrial phenomenon. Each new generation of models requires orders of magnitude more computation—exponential growth, not linear. This computation relies on specialised chips (GPUs), accelerated servers, and very large data centres that can be described as the true factories of the 21st century. The electricity consumption of a single hyperscaler can reach 300 to 1,000 MW, roughly that of a medium-sized city, with construction costs measured in millions of dollars per megawatt. These needs place considerable pressure on power grids, which must evolve rapidly to avoid congestion or outages.

In 2024, global data centre electricity consumption was estimated at approximately 415 TWh. According to projections, this demand is expected to continue rising in the years ahead, following an upward trajectory through 2030. In the United States, Morgan Stanley anticipates a deficit of 49 GW of accessible power, while nearly 70% of the grid is approaching the end of its useful life. Interconnection queues are lengthening, and certain jurisdictions are now imposing moratoria, reflecting the growing difficulty of keeping pace with demand. Goldman Sachs estimates that an additional $720 billion in grid investment will be required by 2030.

The figures speak for themselves. AI-related investments contributed 39% of US GDP growth over the first three quarters of 2025. The four leading hyperscalers (Meta, Alphabet, Amazon, Microsoft) plan to deploy over $650 billion in capital expenditure in 2026. The associated infrastructure supercycle could reach up to $3 trillion by the end of the decade.

Current economic fundamentals do not suggest the emergence of a speculative bubble, as these expenditures rest on consolidated and robust balance sheets. Nevertheless, substantial risks remain, particularly those related to power grid constraints and US export controls on advanced chips.

Three Competing Models

In Washington, the strategy aims to export the complete technology stack to achieve dominance. The Stargate programme envisions $500 billion in AI infrastructure over five years. The policy on advanced chip exports to China, oscillating between restrictions and targeted relaxations, illustrates the doctrine.

On the Chinese side, Xi Jinping regards high technology as the “primary battlefield.” China dominates 57 of the 64 critical technologies identified worldwide, combines civil-military fusion, and explores AI-assisted cognitive warfare—the “intelligentisation of war” (zhìnénghuà).

Europe has adopted the AI Act but remains dependent for infrastructure: non-European hyperscalers control 70% of the cloud market. The InvestAI programme allocates €200 billion, and the EURO-3C project federates 70 organisations. The AI Act formally excludes military use from its scope, but this exemption resolves nothing: in an ecosystem where technological building blocks are inherently dual-use, civilian regulation de facto hampers defence innovation.

What This Changes for Defence

Data centres are now strategic centres of gravity, which several countries are reclassifying as “critical infrastructure” in the national security sense. Armed forces rely on expanding computational capabilities for intelligence, operational planning, cyber defence, and the deployment of autonomous weapons systems. If this capacity depends on infrastructure controlled by foreign powers or non-European private companies, it constitutes a structural vulnerability whose operational consequences could be devastating.

The semiconductor supply chain represents a major strategic vulnerability. TSMC, based in Taiwan, produces virtually all advanced chips, while China concentrates the refining of 95% of the world’s gallium. Data centre demand could, in turn, account for up to 10% of global supply by 2030.

A conflict over Taiwan or a reaction from Beijing could paralyse global computing capacity—a scenario for artificial intelligence comparable to what the Strait of Hormuz represents for oil. Without sovereign infrastructure, the very notion of cognitive sovereignty remains theoretical: the contamination of training data through state-sponsored influence operations is already a concrete and documented threat.

France’s Position

France possesses real strengths: Mistral AI—valued at €11.7 billion—an exceptional research ecosystem, a sophisticated DTIB, and a nuclear sector capable of providing baseload energy. Europe has produced Helsing, valued at over €5 billion, which builds AI defence systems. The 2025 National Strategic Review explicitly acknowledges the need to mobilise all levers of power. But the weaknesses remain concerning. Europe captures only 5% of global venture capital flows. The AI Act’s military exemption does not prevent civilian regulation from constraining dual-use innovation. Non-European hyperscalers dominate 70% of the cloud market. Grid connection procedures remain cumbersome. Above all, there is no integrated strategy linking AI infrastructure, energy policy, and defence. France appears as the principal actor capable of catalysing this convergence.

Lines of Effort

Five lines of effort are essential for French and European decision-makers: treating compute as critical defence infrastructure, with a sovereign cloud backed by nuclear power; massively accelerating grid connection procedures; building a European semiconductor and critical minerals supply chain through the EU Chips Act and diversification away from China; adapting the European regulatory framework so that the AI Act’s military exemption does not remain theoretical in the face of dual-use constraints; and investing in talent at the intersection of AI, energy, and strategy—a domain in which France’s grandes écoles hold an advantage still underexploited by defence.

The history of technological revolutions shows that power does not reside in technology itself, but in the mastery of the physical system that makes it effective. Yesterday, it was steel and coal; today, it is energy, silicon, and data centres. France and Europe possess considerable assets, but these will only translate into power if they are mobilised without delay. Time is running out.