Hosting other people’s AI: is that sovereignty?

Not entirely. Running an American or Chinese open-weight AI model in Europe protects your data and guarantees that no one can take it away from you. But Europe decides neither what these models have learned nor what their next versions will be.

Has Europe lost the model race?

For the most powerful models, yes. The leaderboard of Artificial Analysis, an independent company that evaluates AI models, shows this clearly. As of 14 October 2026, the top ten places are all held by American models (Anthropic, OpenAI, Google, Meta), with scores from 52 to 58. China comes next: 45 for Qwen, 44 for Kimi K3. Mistral, the French company that carries Europe’s hopes, scores 38 with the preview version of its latest model.

The gap is first a matter of money. In September 2026, Mistral raised €3 billion from investors. At the end of March, the American company OpenAI alone had raised $122 billion. And the four American tech giants (Microsoft, Google, Amazon, Meta) plan to invest more than $700 billion in 2026, mainly in data centres and chips.

With gaps like these, catching up with the Americans is not a realistic goal in the short term.

Closed, open-weight, open-source: what is the difference?

It all depends on what the model’s creator agrees to publish.

The closed model. You only have access to the service, through a website, an app or a subscription. The model itself stays with the company that created it. This is the case for ChatGPT (OpenAI), Claude (Anthropic) and Gemini (Google).

The open-weight model. The company publishes the model’s “weights”: the billions of internal settings fixed during its training, which hold everything it has learned. Anyone can download them and run the AI on their own computers. However, the way it was built remains secret, in particular the data used to train it (the texts, images and code it was given to read in order to learn). A licence sets out what you are allowed to do with it.

The open-source model. The company publishes the weights, the code and a description of the training data precise enough for a specialist to recreate an equivalent model. This is the definition adopted in October 2024 by the Open Source Initiative, the reference organisation for open-source software. Such cases are rare.

To put it another way: open weight is a ready-made dish you are given and can serve at home. Open source is the dish with the recipe and the list of ingredients. Many articles say “open source” when they actually mean open weight.

What happened this summer?

At the end of July 2026, the Chinese company Moonshot AI released Kimi K3 as an open-weight model. In the overall ranking, it remains behind the best American models. But on some programming tests published by Moonshot, it matches them. And it can be downloaded for free.

Behind this release, two strategies are at odds. The large American companies rent out access to their best models. Several Chinese companies, such as Moonshot, DeepSeek and Alibaba, release some of their models for free. Their reasoning seems clear to me: if the whole world builds its tools on Chinese AI, it will find it hard to do without it.

Does Mistral’s new model change the game?

It changes something, but not the ranking. On 6 October 2026, Mistral unveiled Mistral Large 4, nicknamed “Le Chonk”. It is very large: around one trillion internal settings, or parameters. The company claims that it clearly outperforms every open-weight model developed in the United States or Europe.

What it does not change: it remains behind the best American and Chinese models.

What it does change: Europe now has its own large open-weight model. Mistral published its weights at the end of October 2026. The first independent tests also rank it well on specific tasks, such as cybersecurity.

Why does open weight change the sovereignty question?

Because it changes the nature of the dependency.

With a closed model, you rent a service. The provider sets the price and the terms. It can change the service, withdraw it or cut off access. And if the provider is American, the American authorities can require it to hand over the data it holds, even when that data is stored in Europe: this is what the CLOUD Act, an American law from 2018, allows.

With open weight, you have the model itself. Once it is installed on your computers, or on those of a European hosting provider, no one can take it away from you. Not the company that created it, not Beijing, not Washington. And your data does not leave Europe.

In practice, few companies will install it themselves. Kimi K3 needs more than a terabyte of memory, dozens of times that of an ordinary laptop. They will therefore go through a hosting provider. This is where Europe finds its place: being the safe place where AI runs, with its data centres and its rules.

The European Union is preparing for this. On 3 June 2026, the European Commission presented a “tech sovereignty” package: a legislative proposal on cloud and AI, another on chips, and a strategy for open-source software. The two legislative proposals still have to be negotiated by the European Parliament and the Member States.

Is hosting the same as controlling?

No. And this, I believe, is where Europe needs to stay clear-sighted.

Hosting someone else’s AI provides one real guarantee: no one can switch it off. That is why open weight matters so much. But three things remain beyond the control of whoever hosts it.

The next versions. You keep the version you downloaded. Improvements depend on the maker. If a Chinese lab stops publishing its models, or if new rules prevent it from doing so, European companies are left with a version that grows older.

What the AI has learned. The training data remains secret. So you do not know exactly what it knows, or what it ignores or avoids. For example, some Chinese AI models refuse to answer on subjects censored by Beijing, such as the 1989 Tiananmen Square crackdown. There is a workaround: connecting it to your own documents, known as RAG. It then answers from sources you have chosen. This limits the problem on your own subjects, without removing it.

The licence. Open weight does not mean free of rules, and each licence has its own. Kimi K3’s is a good example. A company that uses it internally is not affected by its commercial terms. But a hosting provider that offers it to its customers must sign a separate agreement with Moonshot as soon as its revenue, together with that of its group, exceeds $20 million over twelve consecutive months. Even when installed in Europe, Kimi K3 remains partly subject to its maker’s terms.

My position is therefore simple. Europe is right to become the place where AI runs safely: it is realistic, and it is useful. But it will not build the best AI. Its strength lies elsewhere: in its rules. Any AI used in Europe is subject to the EU Artificial Intelligence Act (AI Act), and to the GDPR as soon as it processes personal data, wherever it comes from.

And for a business, in practice?

First, do not confuse two situations. Using a Chinese AI through its Chinese app means sending your data to China. Using the same AI, installed by a European hosting provider, means keeping your data in Europe. The model is the same; the risk is entirely different.

Before choosing an AI tool, ask three questions:

1. Who hosts it, and where? A European hosting provider, in Europe, protects you better than an American provider, even if its servers are in Europe.
2. What does its licence allow? Especially if you are building a tool that you will later sell.
3. What data are you entrusting to it? The GDPR applies whatever the tool’s country of origin. Customer or employee data calls for more care than a draft social media post.