France Wants 5 Gigawatts of AI. I Thought About Servers. I Was Wrong.

The number that stayed with me wasn’t €75 billion. It was 5 gigawatts.

When I first saw the news about SoftBank’s plan to invest up to €75 billion in French AI infrastructure, I thought about the usual things: chips, models, data centers and the race for compute. Then I noticed another number attached to the project — 5 gigawatts. For some reason, that figure stayed with me much longer than the investment itself.

https://www.businesstimes.com.sg/companies-markets/softbank-plans-75-billion-euro-investment-french-ai-centres

The first phase alone is expected to absorb around €45 billion and deliver 3.1 GW of capacity by 2031. On paper, it reads like another technology investment. In the physical world, it means substations, transmission lines, cooling systems, transformers, backup power and years of construction.

The more I looked at those numbers, the harder it became to see AI as a software story. A single prompt feels weightless. Billions of prompts do not. At that scale, AI starts looking less like code and more like industrial infrastructure.

We call it the cloud, but clouds don’t require substations.

Every new AI campus needs electricity. Electricity needs networks. Networks need materials. Among those materials, copper remains one of the least visible and most important. Nobody talks about copper during product launches, yet enormous parts of the digital economy depend on it moving energy through cables, transformers, motors and distribution equipment.

That is where the story became more interesting to me.

While investors focus on compute capacity, another race is unfolding much farther from the server halls. New copper supply cannot be scaled the way software can. Discovering, evaluating and developing a resource often takes 10 to 15 years. AI moves in quarters while geology moves in decades.

One example sits in British Columbia, where NovaRed Mining combines copper-gold exploration with MetalCore AI-assisted mineral evaluation. The idea is simple: use historical datasets, geological modelling and modern analytics to identify targets that might otherwise remain overlooked.

The Trojan-Condor package added 5 mineral tenures covering 4,573.82 hectares. Compared with €75 billion, that number looks tiny. Yet many industrial stories begin this way — a land package, a survey grid, a drilling target, a field season. Years later, cables, substations and industrial facilities appear downstream.

Compute can scale in months. Supply cannot.

The longer I looked at the French AI investment story, the less it felt like a story about servers. What stayed with me was the growing gap between how fast technology wants to move and how long the physical world takes to catch up.

Perhaps the most important part of the AI boom is not sitting inside a data center.

It may still be underground.