Anthropic, Mistral and French AI: Who Really Powers European Datacenters?

Anthropic just raised $3.5 billion. Mistral AI, the French unicorn, is now worth $6 billion. Language models are exploding. But behind this euphoria, a question very few ask: where are the datacenters that run them hosted?

The answer is unsettling. Anthropic, like OpenAI and Google, depends heavily on American infrastructure. AWS, Azure, Google Cloud — the three hyperscalers dominating 65% of the European market — host the bulk of generative AI workloads. In Europe, the ratio is even more brutal: 92% of European corporate data transits through servers located in the United States.

Mistral AI claims to break this dependence. Its Mixtral model, fully open-source, can theoretically be deployed on any infrastructure. Theoretically. In practice, even Mistral uses public clouds to scale its training. The promise of digital sovereignty hits a material reality: there aren't enough certified datacenters in Europe to absorb the demand.

The Gap Between Promise and Concrete

Europe has about 1,200 datacenters. The United States has over 5,000. But the number doesn't tell the whole story. What matters is certified capacity — Tier III or IV, PUE below 1.30, connected to a decarbonized energy source. On this criterion, Europe has only 300 viable infrastructures for generative AI.

Anthropic, for its Claude model, needs clusters of at least 10,000 GPUs. A single cluster consumes 40 MW. In Europe, only three countries can absorb this demand in the short term: France (thanks to its nuclear power), Iceland (geothermal), and Norway (hydroelectricity). The rest of the continent is saturated.

Why Hyperscalers Remain Masters of the Game

Microsoft, Amazon and Google invested $25 billion in Europe in 2025. But these investments are strategically targeted: Frankfurt, Amsterdam, Dublin, London. Primary markets. Markets where building permits are already obtained, where electrical networks are sized, where hyperscalers have government relations.

What these investments don't solve: market access for medium-sized European companies. A French SME wanting to host its fine-tuned Mistral model can't afford to lease 10 MW from Equinix. It turns to OVHcloud, Scaleway or — more likely — stays on AWS Paris. Digital sovereignty remains a luxury reserved for large accounts.

The Investment Thesis: Brown-to-Green as the Answer

It's precisely in this gap that the opportunity lies. Brown datacenters — existing, under-optimized, poorly connected — can be transformed into certified AI infrastructures for 40% less than a greenfield. The method: electrical retrofit, breakthrough cooling (immersion or ORC), and EU Taxonomy alignment.

FINXIA Capital has identified 47 brown assets in France, Spain and Italy that, once transformed, could host workloads from Mistral, Anthropic or European open-source models. The angle isn't technological — it's geopolitical. Who owns the infrastructure, owns the sovereignty.

What Investors Must Understand

Generative AI is not a bubble. It's a structural leverage effect on datacenter demand. But value creation won't happen in language models — margins are low there, competition fierce. Value creation will happen in the physical infrastructure that runs them. Real assets, not intellectual assets.

The entry window is 2026-2027. After that, hyperscalers will have locked up the last available MW, regulators will have tightened EED standards, and brown-to-green transformation costs will have risen. Those who enter now, with a structured thesis and an execution team, will capture the transformation premium.