Digital Sovereignty

An office suite just set an all-time download record for what it left out. LibreOffice 26.8 hit 1,031,162 installs in a week with no AI, and a million people were escaping the feature Microsoft keeps adding to yours. In rented software, the next change is never yours to make.

A frontier open model now runs on a desktop. It writes five words a minute. I tested a smaller one on my own machine and it ran slower than someone reading aloud. Running these models yourself is real. It is a fire escape, not a workplace, and there is no third state between working and gone.

Open, capable, sovereign: in AI you get to pick two. The frontier models are proprietary and revocable, the open ones you cannot run or cannot match, and even Europe's champion runs on chips it will never make. Sovereignty is not the model you pick. It's the weakest layer beneath it.

A Dutch 'sovereign' cloud almost became American overnight, and only a last-minute government veto stopped it. Real digital sovereignty was never just about where your data sits. It is about whose courts can reach it, and Europe's billion-euro plan still misses the point.

The NHS survived hostile state attacks on its open-source pandemic app with zero critical breaches. Now it ordered every team to hide their public code in one week. AI was the justification, not the reason. When a taxpayer-funded health system abandons transparency, that's a democratic decision.

Anthropic pulled Claude Code from its $20 plan and reversed course within days, but the signal matters more than the reversal. Every major AI tool is sold below cost. When subsidies end, every workflow built on rented AI becomes a cost you didn't budget for or a capability you lose overnight.