Following a record summer for AI in terms of model releases and security incidents, Vincent Luciani (X05), co-founder and executive chairman of Artefact, a European leader in AI consulting, puts the debate on AI sovereignty into perspective. According to him, Europe still has cards to play to guarantee both competitiveness and resilience.
Why is the notion of sovereignty so prevalent today in discussions around artificial intelligence?
Sovereignty, first of all, is not autarky. It is the ability to maintain a choice. Being able to decide, at any given moment, which technology to use, where it runs, under what conditions, and having the capability to switch.
Sovereignty has always been a top-tier issue for governments, but this is the first time that businesses are also feeling so strongly how crucial this concept is. Two things have changed over recent months: technology has largely become weaponized (in June, Washington classified two models as military technology), and certain countries have begun locking out others’ access to top-performing tools (Anthropic was forced to temporarily pull Fable at the explicit request of the US administration).
How have recent events brought sovereignty issues even further into focus?
This summer’s news was particularly rich in takeaways regarding model performance and associated risks. While the last three months saw the highest volume of model releases ever recorded (75 models across 36 laboratories), several incidents demonstrated that frontier AI models are not like other technological tools.
We must first remember that in 2025, every organization faced nearly 2,000 cyberattacks per week—a 70% increase compared to 2023. From July 11 to 21, models being evaluated for their cyber capabilities escaped OpenAI’s evaluation sandbox, accessed the internet, and infiltrated Hugging Face’s production systems.
When American proprietary models refused to read the offensive code to resolve the incident, engineers even had to resort to a Chinese model, GLM-5.2, to organize the defense! Then in late August, OpenAI suspended the training of its most advanced models because it could not rule out that its Astra model posed a “critical cyber capability” risk.
How do you think this race for AI model performance will evolve in the future?
My central hypothesis is that the commoditization of models will continue to accelerate, offering accessible intelligence at ever-lower costs. This is the direction the market is taking with the recent explosion of open-weight solutions (where model weights are shared), which show a 2% to 3% lower performance for a dramatically reduced price (roughly 80% cheaper). Keep in mind that all these models are already ultra-performing, and the marginal incremental gain of proprietary models can only be justified for highly critical or niche applications. Finally, consider that an open model backed by contextualized data and a custom harness will outperform a standalone frontier model!
This concept should not be confused with open source, which refers to models whose entire training pipeline is made public. A wave of Chinese open-weight models has launched recently beyond the hyper-hyped DeepSeek, including Qwen (Alibaba), GLM (Zhipu), Kimi (Moonshot), and Hunyuan (Tencent).
However, a divergence in model performance cannot be ruled out, with a frontrunner emerging among the AI Labs that manages to achieve the elusive Artificial General Intelligence. For example, in early September, OpenAI presented GPT-6 Astra as “the beginning of the AGI era,” featuring a leap in capabilities on long-horizon tasks.
For companies seeking sovereignty while accessing top-tier models, open weight is the solution! Concretely, for an enterprise aiming to deploy AI, we recommend integrating these models using an architecture that keeps critical value in-house (the company’s context and harness) and considering the deployment of open-weight models on French-based infrastructure. While perfect sovereignty does not exist, controlled dependency is achievable.
What should be done at the French and European levels to safeguard our leeway against these technological shifts Is Europe still in the race to build sovereign solutions?
Europe must take a pragmatic view of its positioning across the primary technological layers of AI. Broadly speaking, there are two: infrastructure and models.
Regarding compute infrastructure, we obviously need to accelerate the deployment of data centers in Europe and France, and growing public support is moving in the right direction. I would add a nuance by reminding everyone that sovereignty is not limited to a physical presence on European soil. The collective pride that followed SoftBank’s €45 billion investment at the Choose France summit sometimes overshadows the fact that decarbonized electricity is one of our major competitive advantages and must be allocated with restraint. Unfortunately, the adoption of open weight is currently not ingrained in European culture, unlike what we observe in China and the United States. The data from Mozilla’s 2026 developer survey is unequivocal: Western Europe and South America are the only regions where closed models outperform open ones, with open adoption rates dropping from 90% in China to 70% in Europe… There is an urgent need to shift this mindset at the European level, as these technologies may well represent our best hope for the future!
“If perfect sovereignty doesn’t exist, managed dependency is possible!”
It is a matter of culture, marked by a genuine lack of familiarity, but also a matter of service availability: closed models always come with multiple operational layers that facilitate deployment, unlike open alternatives. We must therefore rapidly build the software building blocks required to deploy these open models with the same level of service as closed ones (API reliability and standardization, orchestration, fine-tuning, etc.). On this front, the September 30 announcement of the launch of AI infrastructure startup Blackfuel is excellent news, with this neocloud player specializing in AI inference.
Even so, Europe must not abandon the battle for intelligence. A political decision could easily ban the use of open models (which would quickly be rendered obsolete), as they always carry the blind spots of their training, along with potential misalignment issues. The signal sent by Mistral this summer should trigger a collective wake-up call. Lacking a competitive top-tier model since Mistral Large 2 (July 2024), Mistral had to integrate a Chinese model, GLM-5.2, and largely repositioned itself as a cloud operator and integrator of foreign models. The announcement of Mistral Large 4 is welcome news, but it took a considerable amount of time, and its performance remains far from the generalist model frontier (Vals Index). We must reinvest massively right now in the race for frontier models!
What role does Artefact play in this debate?
It has been over 10 years since I launched Artefact, and we have focused exclusively on deploying Data and AI in companies worldwide. Our role is to drive economic competitiveness through the adoption of new technologies. We have now reached a scale that gives us unmatched field experience, with 2,600 people across 26 offices globally.
Our motto has always been “AI is about People,” reflecting our conviction that transforming ways of working and corporate value capture strategies must always keep humans at the center. Regarding Mistral’s current core activities, we are essentially competitors serving the same client advisory needs!
We are committed to contributing to national sovereignty issues. This is why we created an AI Academy years ago to train as many people as possible in AI, and it is also the driving force behind our participation in the AION consortium to respond to the European Commission’s AI Gigafactories call for tenders, which will build sovereign compute infrastructure in France and across Europe.
