The geopolitics of AI models and Europe's sovereignty

The best large language models today come from two camps: from the US as closed cloud services – and from China as open weights you can run locally, but trained under state oversight. In the race for the top model, Europe lags behind. What does this model dependence mean for digital sovereignty – and for the public authorities and companies that want to use AI?

The landscape: three blocs, one bottleneck

The market for capable language models is clearly split three ways today. The US supplies the technical frontier – but only as closed services. China supplies comparable models as open weights – but under its own state's content rules. In the race for the absolute top model, Europe is so far represented mainly by one name: Mistral. The picture can be plotted on two axes – openness of the weights and ties to European control:

USA: leading, but closed

OpenAI, Anthropic, Google, Microsoft, Meta and Amazon build some of the strongest models in the world – and their commercial flagship models (e.g. GPT, Claude, Gemini, Copilot, Nova) are mostly closed source. The weights never leave the company; access is only via the cloud API, billed per token. That has two consequences for sovereignty:

• No availability: you can't run the model in-house, can't inspect it, can't freeze it. Every request goes to the provider.
• Foreign jurisdiction: the US CLOUD Act lets US authorities demand data – even when the servers sit in Frankfurt. No in-force EU-US agreement resolves this tension with the GDPR.
• Supplier risk: the provider can change prices, availability, terms and model behaviour at any time. A model dependency is a critical supplier dependency.

The picture isn't entirely black and white, though: US providers also release models as open weights. Meta offers freely downloadable models with the Llama family, Google with the Gemma series, Microsoft with the compact Phi models and NVIDIA with open models such as the Nemotron family – and even OpenAI has added open weights with GPT-OSS. For sovereignty these behave like the open models discussed below – yet the provider's US jurisdiction remains a factor as soon as you use its cloud services rather than running locally.

China: open, but imprinted

Chinese providers have caught up with models like DeepSeek and Qwen – and take a different route: they often release their models as open weights, freely downloadable and runnable locally. For sovereignty that's an advantage at first: whoever holds the weights can run them in their own data centre.

The field is now considerably broader than just two names. The internationally noted Chinese open-weights providers today include, among others:

• DeepSeek – powerful reasoning and general-purpose models that drew international attention.
• Qwen (Alibaba) – a broad model family from compact to very large, strongly multilingual.
• MiniMax – efficient models with a very long context window (including the MiniMax-M series).
• Kimi‑K2 (Moonshot AI) – a very large mixture-of-experts model with strong agentic and coding performance.
• MiMo (Xiaomi) – compact, openly released models focused on reasoning and efficiency.

The catch lies in the knowledge base. Chinese models are built under state regulation (including the Cyberspace Administration of China) that restricts content on certain topics – documented ones include Tiananmen 1989, Taiwan, Tibet, Xinjiang and criticism of the leadership. Research shows that part of this alignment sits not only in the cloud filter but already in the model weights themselves and so persists even when run locally. It can be mitigated through evaluation and targeted fine-tuning, but not guaranteed to be removed entirely.

Europe: late, but not out of the game

In the race for the latest, best model, Europe has lost touch with the frontier. The most visible provider is Mistral (Paris), known for efficient architectures and several freely available open-weights models. Alongside it are serious efforts:

• Aleph Alpha (Heidelberg) – focused on sovereignty, explainability, on-premise and regulated industries.
• OpenEuroLLM – an EU initiative for truly open models across all EU official languages, publicly funded, first models from mid-2026.
• OpenGPT-X / Teuken – a multilingual, European-trained model from German research.

These projects are strategically valuable – but they don't (yet) close the gap to the absolute performance frontier. Anyone wanting to deploy AI productively today can rarely avoid models of non-European origin. So the interesting question isn't only "which model", but "operated how".

The comparison

What is digital sovereignty?

Digital sovereignty means keeping control over your own data, systems and critical dependencies. For AI models, that breaks down into three questions:

• Data sovereignty: do prompts and documents leave your building? Who can access them?
• Availability & control: can we run the model ourselves, freeze it, inspect it – or does everything hang on an external service that can change or be switched off?
• Values & transparency: what content rules are baked into the model and can we understand and influence them?

None of the three blocs, on its own, satisfies all three criteria optimally. That's exactly why the naive question "US or China model?" falls short.

Implications for authorities & companies

What sounds like abstract geopolitics gets very concrete in daily life – especially in the public sector and in regulated industries:

• Public authorities: an administration processing citizen requests through a US or China cloud hands sensitive personal data into foreign jurisdiction. If a content-imprinted knowledge base is used in citizen contact, the neutrality of state action is affected too.
• Companies: trade secrets, contracts and customer data in a foreign cloud are a compliance and competitive risk. A critical model dependency can also become more expensive, restricted or discontinued overnight.
• Strategic resilience: the deeper AI is woven into processes, the more dangerous an external "kill switch" becomes – via price, politics or export controls.
• Democratic dimension: if information and text production increasingly run through a few non-European-imprinted models, that is also a question of how public opinion is formed.

The lever: the operating model

The good news: sovereignty depends less on a model's origin than on its operating model. Running an open-weights model locally defuses the biggest risks at a stroke – even for non-European origin:

Run locally, no prompt and no document leaves the building, there is no external kill switch and the value imprint can at least be inspected and adjusted through fine-tuning. This is exactly the path we describe technically in our articles on the local LLM server with vLLM and on GDPR-compliant self-hosted RAG search. On top of that: back European models wherever you can – and integrate and secure the best of the open models in a European way.

Recommendations

• Treat model dependency like a critical supplier: assess risks, keep alternatives ready, secure exit options.
• Prefer open weights and – where data protection or confidentiality matter – run them locally rather than through foreign cloud APIs.
• Check the use case against the value imprint: for sensitive, socially relevant applications, evaluate the model's content alignment.
• Plan for European options: use Mistral and the upcoming EU models where they hold up – and actively support their development.
• Establish governance: clear rules on which data may go into which model, with logging and regular review.
• Authorities first: in the public sector, local operation on European infrastructure takes priority over convenience.

Frequently asked questions

Sources

External sources, as of August 2026 (open in a new tab):

• Stanford HAI & DigiChina – China's diverse open-weights AI ecosystem and its policy implications (PDF)
• arXiv – Information Suppression in Large Language Models: auditing and quantifying censorship in DeepSeek
• TechCrunch – Open-source LLMs hit Europe's digital sovereignty roadmap
• MarkTechPost – Europe's top AI models of 2025 (incl. Mistral, Aleph Alpha, OpenEuroLLM)
• Wire – What the CLOUD Act really means for EU data sovereignty
• EU Cloud and AI Development Act / EU AI Act – overview and deadlines

This article is a snapshot as of August 2026 and is for general information. Providers, models and regulation are evolving fast.