The image has a paradoxical flavor. At the G7 working lunch on artificial intelligence, the head of Europe's largest AI lab found himself discussing not transformer architectures or context windows, but kilowatt-hours and supply contracts. France holds a card that few outside its borders had realized: cheap, low-carbon electricity, thanks to the most extensive nuclear fleet in Europe. As American giants race to build data centers on French soil, the government must now decide who gets that computational power: national companies developing LLMs and foundation models, or hyperscalers promising billions in investment and jobs.

The stakes are underestimated by those who view AI only through the lens of algorithms. Training a large model requires GPU clusters consuming as much electricity as a small town, and inference, though less intensive, scales linearly with users. With industrial electricity prices in Germany or Italy remaining high, the French advantage translates into a TCO differential that can exceed 30% over a plant's lifetime. For a startup wanting to self-host an LLM with periodic fine-tuning, this means more sustainable margins and less reliance on cloud credits.

The sovereign core

This is not just about energy accounting. The tension is between two deployment visions: on one side, a continental on-premise ecosystem where European actors retain control over data, training pipelines, and GDPR compliance without resorting to foreign cloud regions. On the other, the classical expansion of cloud service providers, bringing efficiency, scale, and technological lock-in extending from accelerators to orchestration frameworks. If Paris prioritizes grid connections for American data centers, energy costs for local competitors will rise, pushing them toward renting GPUs in the public cloud — paradoxically from the same operators benefiting from the favorable tariffs. A vicious circle would emerge, with digital sovereignty retreating in favor of short-term economic efficiency.

Second-order implications extend to hardware supply chain dynamics. Major accelerator vendors tend to favor clients with large, multi-year orders; if the French electricity backbone ends up powering cloud regions, those supply agreements strengthen, reducing GPU availability for smaller operators and public labs pursuing on-premise deployment. On a European scale, this means slowing down the very independent innovation that the AI Act aims to stimulate.

Then there is a third-order consequence, less visible but more structural. Whoever controls energy distribution for AI loads implicitly sets the pace of the transition toward extreme quantization, efficient fine-tuning, and low-power architectures. If foreign capital grabs the cheapest grid nodes, the incentive to develop models that run on modest hardware — the real challenge for enterprises wanting to perform inference on-site — weakens, because compute costs remain artificially low only for a select few. The alternative is an industrial policy that ties energy access to transparency requirements, open weight models, and local spillover clauses: a kind of "locational grid edge" not unlike carbon leakage debates, but applied to tokens.

For those evaluating an on-premise LLM deployment in Europe today, the French match is a signal that energy cost is becoming a variable comparable to VRAM or memory bandwidth. AI-RADAR follows these developments closely, because every decision on who gets the kilowatt-hours directly ripples into inference service pricing and the real convenience of self-hosting over paying an API provider. The answer is still hanging in the balance, but the very fact that the question is being asked publicly — and at the G7 level — shows that the era ethereal AI is over. Now it's about the meter readings.