Energy as a competitive advantage in the AI ​​race

Artificial intelligence is no longer just a matter of algorithms and performance. It has become an infrastructure issue, with considerable physical impacts that are reshaping economic competition.

Explosive electricity consumption

Data centers dedicated to AI are undergoing a radical transformation. Where a traditional data center consumes a few tens of megawatts, AI installations regularly exceed 100 MW, equivalent to the consumption of a large medium-sized city.

Globally, data centers still represent a minority share of electricity consumption (around 1,5% in 2024), but this share could more than double by 2030. In France, consumption could reach 4% of national electricity in 2035, creating significant local tensions on the networks.

Electricity is becoming a strategic advantage

Access to the electricity grid is becoming a barrier to entry. Connection times are measured in years, and in the most sought-after areas, securing land and electricity capacity early on is becoming a decisive advantage.

Energy costs represent between 30% and 50% of a data center's operating expenses. Large players secure their supply through long-term contracts for decarbonized electricity, stabilizing their costs and consolidating their dominant position. Smaller players struggle to access these mechanisms.

Beyond electricity, other resources are under pressure.

The footprint of AI is not limited to energy. Water used for cooling becomes critical in regions facing water stress. The rare metals needed for chips and equipment create geopolitical dependencies. The carbon footprint remains significant, even with decarbonized electricity, due to emissions related to hardware manufacturing.

Frugality as a new arena of competition

Faced with these constraints, the concept of AI efficiency shifts from a peripheral topic to a competitive imperative. It's no longer just about doing better technically, but about doing better with less.

There are many levers to pull: using smaller, more specialized models, optimizing inference rather than just training, orchestrating multiple levels of models depending on complexity, and limiting unnecessary calls. Inference often accounts for the majority of computing power consumed once the models are publicly available.

For startups, frugality becomes a double advantage: lower unit costs and a response to a growing demand for sobriety, particularly in public markets and regulated sectors.

Transparency and standardization: new competitive challenges

Without robust methodologies and shared standards, the risk of greenwashing skyrockets. The opacity of environmental impacts can become a competitive advantage based on information asymmetry rather than on actual merits.

Standardization is progressing, but it must itself avoid becoming a tool for market control. Standards must be representative, transparent, scientifically robust, and accessible to smaller players.

A new performance paradigm

AI is entering a phase where performance is also measured in kilowatt-hours, tons of CO₂, cubic meters of water, and grams of critical metals. This is not a constraint that slows innovation, but rather one that redirects it toward systems that are smarter in their decision-making, more transparent in their impacts, and more competitive because they do more with less.

The next major advances will come as much from resource optimization as from model improvement.

To go further: this race for energy extends our article on cognitive dependence and the ecological cost of AI et our article on Claude's safety and compliance in business.