A supplier that changed twice in five months
Start with the timeline, because it is the whole story in miniature. In May, Ecosia swapped OpenAI for Mistral in a move that read as a statement of European independence: a Berlin company buying from a French lab. Five months later the same customer is leaving, and the exit interview is unusually blunt. Quality disappointed, servers buckled under the load, and the company decided it was simply too large a customer for its own continent's flagship AI lab.
That last detail deserves a pause. This is not a customer complaining about a model being mid at release. It is a customer saying the supplier could not absorb real production traffic. Model quality gets the headlines, but capacity is the unglamorous part of enterprise AI, and a European champion failing at it is the kind of signal procurement teams across the EU will read twice.
The sovereignty argument, read twice
The most interesting part of the interview is not the engineering complaint. It is the definition of sovereignty on offer. Kroll argues that Mistral's reliance on international investors means it is, in his words, not truly sovereign. Fair enough as a definition. But follow it to its conclusion: the replacement weights come from Alibaba, Zhipu AI and Moonshot AI. Under Kroll's own test, European money backing a French lab is disqualifying, while the trained parameters of Chinese companies are the answer.
The environmental objection is stranger still. Ecosia has climate concerns about France's nuclear-heavy energy mix. Nuclear power is the lowest-carbon baseload energy in Europe. Whatever Kroll means by environmental here, it is not a carbon argument, and POLITICO notes the article does not elaborate on how the energy mix bears on the company's criteria. When a company plants trees for a living, the brand requires an environmental sentence. Read it as branding, not analysis.
None of this means the decision is wrong. It means the stated reasoning is a wrapper around a simpler calculation: the Chinese open-weight models were good enough, cheap enough, and available through a hosting arrangement that worked. Half the cost, better quality, no overloaded servers. Sovereignty enters after the numbers, to make a procurement decision sound like a strategy.
The small print on open
Two footnotes before anyone cites this as a victory for open weights. First, "open-weight" is not "open source." The parameters are downloadable; the training data and code usually are not. And Mistral also publishes weights, so this dispute was never about openness. It was about which models performed and which company could stand behind them at scale. Ecosia's move is a procurement verdict on a vendor, not an ideological one on licensing.
Second, the bias question is real and will not be solved by the hosting layer. A German public-sector-facing search engine now routes queries through models whose governments treat certain topics as non-discussable. Kroll says that can be addressed technically. The expert POLITICO asked says it cannot, not at the level where it matters. For a consumer chatbot the cost of getting this wrong is a misleading answer. For a search engine used by ministries and health services, the cost is measured in trust, and trust is the product.
What builders should take from this
Strip away the sovereignty theatre and there is a plain signal for anyone choosing models this quarter. A company with 20 million users and institutional clients just put open-weight models into production over a flagship proprietary lab, and its stated outcome was lower cost with better quality. However it is defined, that is open-weight models clearing the "good enough for production" bar in a real deployment, with the EU's AI Office able to demand weight-level access to general-purpose models since August.
The uncomfortable part is the supply chain that comes with it. Europe's open-model strategy now runs on weights trained in Shenzhen and Beijing, served through a European hosting platform. If the goal was independence from American tech, mission accomplished. If the goal was independence at all, the asterisk is doing most of the work.
Sources
- [1] ppc.land summary of the POLITICO interview, "Ecosia drops Mistral AI for open models, citing a 1-year quality gap" (Oct 6, 2026)Read source
- [2] Techmeme, "German search engine Ecosia... ditches Mistral for open-source models, including Chinese ones, citing Mistral's disappointing quality (Politico)" (Oct 8, 2026)Read source
- [3] Mistral AI, Large 4 announcementRead source
- [4] NewsGuard, "Crack in the Mirror: Chinese Chatbots and the Western Benchmark" (Aug 13, 2026)Read source
- [5] Internal link to our earlier piece on Mistral Large 4Read source