Digital sovereignty
The kill switch makes the case for open models
In June, the American government restricted the use of Anthropic's two most capable AI models. Anyone who is not an American citizen could get access only with the government's permission. Anthropic switched both models off for all its customers.
At NVIDIA GTC in Berlin, Anne Bouverot called this a kill switch, a way to switch off access to a model from outside. Bouverot is France's special envoy for AI and chairs the board of the École Normale Supérieure. In her account, the company had no way to tell citizens from non-citizens, and that is why access closed for everyone.
By the start of July, the restriction on one of the two models had been lifted. The effect stayed. According to Bouverot, the reality of Europe's situation is sinking in, and every bank, every telecom operator, and every company that depends on very strong cyber protection is very worried.
Bouverot spoke in a panel on open models with Jörg Bienert, the managing director of the Center for Sovereign AI in Germany. Julien Demouth of NVIDIA led the conversation. An open model is a model that anyone can download and run on their own hardware. The panel showed that an organization has to know what must keep running when access stops, and that open models can do most of the work.
Organizations are working out their minimum
According to Bouverot, every company and every government she knows is working on the same question. What are the minimum viable operations, the part of the work that an organization has to be able to carve out, secure, and keep running, whatever happens?
For her, this is the resilience side of sovereignty. She spoke of increased sovereignty, of a Europe that is more resilient and better able to build its own tools across the whole chain.
In her advice, the models of today stay in use. Europe remains a customer of the American and Chinese models, because they are what is available now. She does take into account that they may at some point be paused or switched off, or that Europe gets the version before the latest.
In an earlier blog, Alec Fong of NVIDIA named the same risk for a single team. The maker of a closed model controls the execution and can downgrade a user.
Is an open model good enough?
An open model that runs on an organization's own hardware keeps running when access to an outside service stops. That leaves the question whether such a model is good enough.
Bienert leads Soofi, an open model that is being built in Germany. It began early last year, when the news about the model DeepSeek showed that a good model can also come from working efficiently, with far less computing power and money. Within a week and a half, universities, research institutes, and startups had written a position paper together. The German Federal Ministry for Economic Affairs and Energy funded the project.
Training started in March of this year on the new industrial AI cloud of Deutsche Telekom. A few months later, the first model was ready, with 30 billion parameters. Parameters are the values inside a model that are set during training, and their number is a rough measure of its size.
According to one of the companies that tested the model, running Soofi is 95 times cheaper than running the large models of OpenAI or Anthropic, and for 95 percent of the applications it is good enough. That is the feedback of one tester, and Bienert presented it as such.
For the remaining five percent, the model has to grow. A model of 120 billion parameters has finished its first phase of training. For next year, Bienert announced a model of 550 billion parameters, which comes close to the most capable models. Soofi has received more than 3,000 requests to test it.
Bouverot added that many applications have no need for the most capable model. They can be built on Soofi and on models like it, and the more specific to a sector those models are, the better.
Is a secret model safer?
Demouth put a common objection to Bouverot, that AI is safer when its makers keep it closed and secret.
She answered from her time as CEO of Morpho, a company that worked on the chips in payment cards, passports, and SIM cards. Those chips are secured with cryptography. In the early days of that technology, the idea was that secret is safe, so the algorithms stayed unpublished. That did not work out. The whole industry moved to public algorithms with private keys, and the technology became much more secure.
For AI, she sees a benefit in openness. An open model is looked at by more people, so problems are spotted before they become real. She placed the truth somewhere in the middle, and held that nobody can say in general that secret is safer.
What has to stay closed?
Research and general models can be open. A version that is trained on specific data, such as cyber data, is closed. Both speakers drew the line there.
For that closed part, Bouverot pointed to arrangements that defense and cybersecurity have had for a long time. There are specific contracts and confidentiality agreements. The model runs in a hardened environment on the organization's own premises. For defense, staff are screened before they get access. These arrangements exist, and they now have to be defined for AI.
According to Bienert, companies already work this way. They train a model of their own on the open base, with their own data, and that model stays their property. Demouth explained that a general model can be trained further on a new field, and then on the tasks it has to carry out.
Who checks the quality?
Demouth asked who is responsible once such a model has been built and has to be maintained. Bouverot placed the check with the application, sector by sector.
She sits on the board of Safran, an aerospace company, and her example was Boeing. After serious accidents, Boeing slowed down its production and built a quality check into every step, where it used to check everything at the end. Production is now going up again, with higher quality and safety.
Aviation has its own authorities and standards for this, in the United States and in Europe. According to Bouverot, the same can be done sector by sector, with the mechanisms those sectors have. A supplier that sells a product stays liable to its customers when that product causes damage.
Linux shows what a shared base does
For Bienert, the example is Linux. In the nineties, every hardware maker had its own operating system. An organization that chose a supplier was tied to that supplier's system, and getting out was hard. Then Linux arrived. Today everyone uses it, it has become much cheaper, and nobody asks which hardware runs underneath.
He wants the same for foundation models, the general models on which AI applications are built. Teams across Europe would work on a shared base of code, data, and models. He called that base the operating system of the AI economy.
Europe needs enough for itself
Bouverot's goal for Europe is to have enough for itself, for its industries and for its cybersecurity. Leading the race and spending the most are outside that goal.
The talent is there. Of the authors of the 2017 paper behind today's language models, she said, only three came from the United States, and the rest came from Europe and Canada. Of the fourteen authors of the Llama paper of 2023, eleven came from Europe. The innovation that followed is now led by the United States and China.
Her advice consists of three steps that run side by side.
- stay a customer of the American and Chinese models in use today, and make sure those contracts continue
- build own capability as fast as possible, starting from what exists, such as Soofi, Mistral, and Black Forest Labs
- speed up research, so that the next generation of AI companies starts in Europe
No European country can do this alone, she said. France and Germany are middle powers that need each other, public and private parties alike, and that also need countries outside the European Union such as Switzerland, the United Kingdom, Norway, and Canada. Governments and companies can pool their demand in joint contracts, so that more computing capacity is built in Europe.
Her rule is to start a project with the countries and companies that want to take part, and to accept that the next project has a different coalition. Speed matters, and waiting until everyone has a say takes too long.
One or two years ago, Bienert said, most people thought that the train had left and that the large models were a matter for the Americans. That view has turned 180 degrees. He asks for investments that are larger and also wiser, coordinated so that every group builds a module of one large rocket.
An AI strategy draws the line
Bouverot's three steps are advice for a continent. For one organization, they come down to one line, between the work that runs on the models of an outside supplier and the work the organization can run itself.
The minimum that has to keep running belongs on the organization's own side of that line, together with every model that is trained on sensitive data. For most other applications, an open model deserves a place in the comparison, on cost and on quality. The most capable models of outside suppliers stay in use where the work asks for them.
Quilyx takes AI ambitions from the boardroom to working results. In an AI strategy, this line is one of the choices that a board adopts, with an order, an owner, and a first step. Our starting point is sovereign AI, which means that an organization decides for itself where its data is processed and by whom. Where a hosted service is good enough, we say so too.
In our own work, only our own AI, on our own hardware in the Netherlands, handles the data of clients, suppliers, and professionals, and a hosted service does the knowledge work without sensitive data. Quilyx is also a Registered Partner in the Claude Partner Network of Anthropic, the company in Bouverot's example.
Bouverot set the goal for Europe at having enough for itself. For one organization, enough is the minimum that keeps running, whatever happens.
About this article: This article is our own reading of the session How to Scale With Open Source at NVIDIA GTC Berlin on 21 October 2026. The ideas from that session belong to the speakers and are credited to them. The translation to clients, the positions and the choices in this article are ours. We wrote and edited it together with AI, which helped with the research, the working out and the wording. Quilyx carries the editorial responsibility.