From Models to Compute and Compliance

Sovereign AI is turning into a local infrastructure business, not a national model champions business. Once European buyers accepted that a Chinese open weights model could run inside European data centers with European data controls, the scarce thing stopped being the model itself and became GPUs, hosting, deployment, and compliance work that lets banks, manufacturers, and governments use AI without sending prompts, data, or logs outside the region.1 2 3 4

• Mistral already showed the economic pull of this shift. It moved from open model competition into private deployments, tooling, and implementation support, reached $400M ARR by February 2026, and is now extending into dedicated compute so it can sell a full local stack instead of just model access.1 2 3
• This changes who Mistral competes with. The relevant rivals are no longer just OpenAI or Anthropic on model quality, but neocloud and sovereign compute providers like Nebius, Fluidstack, and Nscale that win by offering regional GPU capacity, predictable latency, and data residency inside Europe.1 2 5
• Cohere points to the other viable path. Rather than owning as much compute, it has pushed upward into enterprise applications and workflows, and its tie up with Aleph Alpha and deployment on Schwarz Group's STACKIT suggests Germany's answer may be software, integration, and local control layered on shared infrastructure.2 4 6

The next phase of sovereign AI will look more like regional cloud procurement than frontier model racing. The winners will be the companies that can lock in GPU supply, package compliance and deployment services, and become the trusted local operator sitting between global model supply and regulated enterprise demand.1 2 3 4 5