GPU and Electrical Power: Can France Power Generative AI?

NVIDIA just announced it delivers 100,000 H100 GPUs per quarter. Each H100 consumes 700W. A cluster of 10,000 GPUs — the minimum to train a competitive language model — consumes 7 MW. That's equivalent to a city of 15,000 inhabitants. And France has 18 GW of datacenter demand in queue at RTE. The question is not technological. It's energetic.

France is an energy paradox. It produces 95% decarbonized electricity — nuclear, hydro, wind, solar. It exports 20% of its production. And yet, it cannot connect the datacenters that demand it. Why? Because the distribution network is not sized to absorb 40 MW loads in rural areas. And because transformation stations require 2 to 3 years of work.

RTE, the transmission network operator, published a report in early 2026 that sends chills: 18 GW of demand is reserved for datacenters, but only 2.7 GW is achievable in the short term. The rest is blocked by network constraints, permits, and land availability. In other words: 85% of demand cannot be satisfied in the next 3 years. This is a structural shortage.

The GPU War: A War of Power

Hyperscalers no longer fight for GPUs — they fight for MW. Microsoft, Google, Amazon and Meta each announced $10 billion investments in AI infrastructure in 2026. But these investments are not in chips. They are in electricity. Hyperscalers buy power plants, negotiate 20-year decarbonized PPAs, and invest in distribution networks. Because they know: without electricity, GPUs are useless metal.

The French case is exemplary. EDF, which produces electricity, has excess capacity. But Enedis, which distributes electricity, doesn't have the transformation stations to deliver it to datacenters. This is the bottleneck. A 40 MW datacenter in Paris is not a problem — the network exists. But a 40 MW datacenter in Amiens, Rouen, or Limoges is impossible without 3 years of work and 50 million euros of network investment.

France: Energy Exception, Infrastructure Handicap

France has a unique advantage in Europe: its nuclear power. 56 reactors produce low-carbon electricity, 24/7, regardless of wind and sun. This is the major asset to attract AI datacenters, which consume constantly — not intermittently. A GPU doesn't stop when the wind drops. It needs stable, predictable, decarbonized power. France is the only European country offering this combination.

But this advantage is wasted by obsolete distribution infrastructure. The French electrical network was designed in the 1960s to power factories, cities, and households. Not 40 MW GPU clusters. RTE is modernizing the network — but work takes 5 to 10 years. And datacenters don't have 5 years. Language models evolve every 6 months. Whoever doesn't have GPUs in 12 months is outdated.

The Investment Thesis: The Site with the Right Connection

It is precisely in this tension that the investment opportunity lies. Sites with good electrical connections — existing, sized, decarbonized — become strategic assets. FINXIA Capital has identified 23 sites in France that combine three advantages: an existing connection of 5 to 20 MW, a decarbonized energy source (nuclear or hydroelectricity), and an obtained or quickly achievable building permit.

These sites are not in Paris, Lyon or Marseille — they are already saturated. They are in industrial zones in reconversion, in former free zones, in regions seeking to attract tech employment. A site 100 km from Paris, with a 10 MW connection and a nuclear PPA, can host a cluster of 2,000 GPUs. That's enough for fine-tuning, inference, and enterprise AI workloads. And that's where value is created.

What Investors Must Understand

The power shortage is not a French problem — it's a European problem. Germany depends on coal and gas for 40% of its electricity. Spain and Italy depend on natural gas. Only France and Scandinavia offer stable decarbonized electricity. This is why hyperscalers are investing massively in France: Google in St-Ghislain, Microsoft in Paris-Saclay, Amazon in Haute-Savoie. They know French electricity is a geopolitical asset.

The investor who understands this dynamic doesn't bet on the GPU — they bet on the MW. GPUs are a commodity. The MW is a rare asset. And whoever owns the MW owns the value chain. FINXIA Capital has structured its TITAN DC AI strategy around this conviction: invest in electrically advantageous sites, transform them into certified datacenters, and lease them to AI operators desperately seeking capacity. The window is 2026-2027. After that, available MW will be locked up.