Ecosia added Chinese open models alongside Mistral. What that says about Europe's AI strategy.
Ecosia, the Berlin search engine that spends its profits planting trees, moved its AI from OpenAI to Mistral in May. Europe's AI champion, the sovereign choice. In October, Ecosia added Chinese models alongside Mistral's, and its CEO Christian Kroll told POLITICO Mistral was "a year behind".
A defeat for Europe's AI strategy? Or a lesson in what sovereignty in AI means?
An AI model is a file: billions of numbers, called weights, set during training. Rent it through a provider's service and it can be taken away. On 12 June, a US export order made Anthropic shut off its newest model, Fable 5, for every customer, for 19 days.
The Chinese models Ecosia uses are open-weight, as some of Mistral's are: anyone can download the file and have it run wherever they choose. Ecosia says it prefers European hosts and never sends users' questions to the US or China. No developer can switch a downloaded copy off.
Some ties remain. A license can fence Europe out. MiniMax's open video model H3, released in August, may not be used in the EU, while MiniMax offers it there as a service.
Whatever training puts into a model travels with the file. In November 2025, the security firm CrowdStrike found that DeepSeek's R1 wrote less secure code when a prompt mentioned Tibet.
And the next version is someone else's decision. On 16 October, China proposed export controls on large AI models, exempting what is already public.
Open weights remove the off switch. They don't remove the supplier.
Look at AI as a chain: chips, computing power, data, models, and the organizations that use them.
Europe holds a chokepoint in the first link: the most advanced chips can only be made with the Dutch company ASML's machines. It is weak in computing power: in September the Brussels think tank Bruegel put the EU at about 5% of the world's AI capacity, the US at 78%. Its industrial and health data is rich but scattered. It is behind on models. And it holds the last link: 450 million people, and rules that bind whoever sells to them.
The EU's headline AI programs aim at the frontier, where the gap is widest: gigafactories to train the largest models and an open European model to rival them. US private AI investment in 2025 was $286 billion, Europe's $21 billion, according to Stanford's AI Index.
Europe won't win every link. The aim is that nobody builds AI without the links it holds. So it should put its money where control is won: computing power to run open models, not only to train new ones, and European data to adapt them with. And build open European models as insurance, so no single country supplies the file. Switzerland's Apertus also publishes its training data, so anyone can check what went into it.
So was Ecosia's move a defeat for Europe? Half of it is the strategy working: open models, run where Ecosia chooses, with questions kept at home. The other half is the gap: a European model a company like Ecosia would pick on quality alone.
