It just stopped being enough. Open, capable or sovereign. In AI, you only get to pick two.
Whenever a software vendor stops playing nice, my advice is usually: go open source, and you can always walk away.
For AI, that advice does not really cut it. The exit still opens. It just no longer leads anywhere as good.
In June, the US government used export-control law to order Anthropic to cut foreign users off from its newest model, Fable. The company could not separate its American users from everyone else, so the rest of the world lost access overnight. Weeks later the same pressure hit OpenAI's new GPT-5.6 Sol. Two of the world's leading AI labs, both American, both switched off abroad by decree.
So the instinct kicks in: go open. Moonshot AI, a Chinese lab, released Kimi K3, an open-weight model anyone can download, that beats the American leaders at writing code. DeepSeek and others followed within weeks. Free of Washington, free to run yourself. Except you cannot.
Open weights look like the escape hatch. They are not. Three walls stand behind them.
The first is compute. Kimi K3 carries 2.8 trillion parameters. Almost no company could buy the hardware to run it, let alone house it. "You can download it" is not "you can operate it". The second is transparency: open-weight is not open source. You get the finished model, not the data or method behind it, so you cannot audit it or rebuild it. The third: the door swings both ways. Beijing is now weighing the same limits on foreign access that Washington just imposed.
That leaves the fully open alternatives: weights, code, and training data published in full. Switzerland's Apertus and Germany's Soofi drew fanfare this month as sovereign answers to the two superpowers. Both sit a step behind the frontier. Honest and sovereign is not the same as competitive.
Underneath sits the wall no one clears. Mistral, Europe's one serious contender, took criticism this month over a deal with Microsoft. Look closer: Microsoft owns under one percent of it, its biggest shareholder European, and it builds its own data centers. More independent than the headlines said. Yet it still runs on chips designed in America and built in Taiwan, at the one factory on earth that can make them. If the company that did everything right cannot reach the bottom of its own stack, neither can you.
The rest of the world is not caught between two superpowers with a clean third way. It stands on a stack of dependencies, each layer held by someone else: who can revoke your access, who runs the platform, who holds the weights, who owns the compute, who makes the silicon. You are only as sovereign as your weakest layer, and for almost everyone that layer sits far below the model they thought they chose.
Before the next export order or license change decides for us: if the AI we all depend on can be switched off in a capital we do not vote in, and runs only on machines we cannot build, what is the plan for the next year, the next three, the next ten? Pretending we have already escaped is not one.
