The State of European AI

European companies have largely stopped asking whether they should use AI. The open question is how to use it without sending every prompt, company data and users' personal data to the United States or China. At the same time, EU regulators are calling for more "European AI" initiatives.

But what is "European AI"? The press regularly mixes up three things: European models in the sense of "Made in EU" LLMs, AI inference running on European infrastructure, and AI assistants that don't send European users' prompts overseas.

Aleph Alpha, a German AI lab once pitched as a potential frontier-model contender, raised a €500 million investment package from a consortium of German enterprise firms (including SAP, Bosch, and Schwarz Group). Roughly a year later, it quietly exited the frontier-model race and repositioned as an enterprise sovereign-AI software vendor.

The only European lab left with serious funding for general-purpose LLM development is Mistral AI in Paris, founded by former Google and Meta researchers. Mistral Large 3 was competitive among open-weight models at launch but trailed the best proprietary frontier models. The latest release, Mistral Medium 3.5 from April 2026, narrowed the gap on coding benchmarks; overall it remains below the global frontier, which now includes open-weight models from Chinese labs.

Mistral recently announced it will offer third-party models, starting with GLM-5.2, an open-weight model from the Chinese lab Z.ai. A pessimistic reading: the last remaining EU champion has stopped trying to win the frontier-model race and now sells tokens for hosted Chinese models. Whether Mistral will eventually stop spending on training its own models is, for now, speculation.

The AI inference layer is where already trained models run and answer user requests. If "European AI" means anything practical today, it is probably there. EU-based infrastructure providers host open-weight models in European data centers. The models may not be "Made in EU", but at least the data stays in Europe, under EU privacy and data protection regulations. Sixteen of these inference providers are currently listed in our AI APIs category.

So the EU has decent inference capacity; the most capable models running on it, however, still come from American and Chinese labs.

Most inference providers in the list fall into two camps: pure-play AI clouds (examples include Nebius and Nebul) and general-purpose EU cloud providers. The pure-play providers usually lead on model choice: new releases typically show up within a few days, while the general-purpose clouds often run a slightly outdated model catalogue.

The closest thing to a B2C, mass-market product like ChatGPT is Mistral's Le Chat. A few smaller, specialized AI assistants exist (focused on certain niches or optimized for non-English languages), but their footprint is tiny. ChatGPT remains the dominant AI brand among casual consumers, and Gemini and Copilot have a massive distribution advantage over any new European contender. That is unlikely to change: consumer AI is won as much through distribution as through model quality, and the big US tech companies already own the browser, the phone, and the operating system.

For companies wanting a general-purpose, ChatGPT-style assistant, the realistic path is a small frontend interface, hosted internally, combined with a European AI inference provider and whatever open-weight model fits the task — rather than a managed ChatGPT clone.

Europe's likely place in AI is hosting and inference: EU-based infrastructure with strong privacy commitments rather than home-grown frontier models. Frontier models will come from the US and China, and that will not change without massive new investment in Europe.

That is a workable position. AI inference is a growing and profitable business, and it is the one AI market European providers can still compete in, if they position themselves as the privacy-first alternative.