Applied AI
Europe is not too late for AI
The train has left, and the large AI models are a matter for the Americans. One or two years before the autumn of 2026, that was how most people looked at Europe and AI.
Jörg Bienert recalled that view at NVIDIA GTC in Berlin on 21 October 2026. Bienert leads Soofi, an open AI model that is being built in Germany. He spoke in the panel on open models that we described in our earlier blog The kill switch makes the case for open models. In his account, the view has turned 180 degrees. He called it common sense that Europe can build such models, has the resources for it, and has to do so.
The old view and the new one are both about models. At the keynote of the same conference, NVIDIA drew a larger picture, in which AI is a new industry and the models are one part of it. This blog follows that picture. In our reading, it shows why Europe is still in time for AI, and why the same holds for an organization that has yet to begin.
The models are one of five layers
The AI industry is built up in five layers, and each layer builds on the one below it.
- energy, the bottom layer
- compute, the computing power that advanced computers deliver
- AI factories, the data centers built for AI in which those computers stand
- models, which run in the factories
- applications, the products and services that are built on the models
The large models of the old view are the fourth of those five layers, and companies and jobs are growing at all five. The advice to Europe was to produce AI as well as buy it, at every layer. Europe has already started. The slide with the five layers showed European companies at each of them, with Deutsche Telekom and Nebius among the AI factories and Mistral AI and Aleph Alpha among the models.
The lower layers come first. Infrastructure has to be in place long before the models that run on it, just as the networks had to be there before the internet. AI factories are being built all over Europe. By its own account, NVIDIA invested nearly 10 billion dollars in Europe in the twelve months before the keynote, and it described the European AI economy as starting and really accelerating.
AI went to work in 2026
In 2026, AI went from answering questions to doing work. The turn came at the beginning of that year with OpenClaw, an AI agent that anyone could download.
An agent carries out a task from start to finish. It reasons, calls on tools, and repeats its steps until the task is done. Because that is useful, more agents are built and more people use them.
That phase is young, and it began with software that was open to anyone. An organization that has yet to start is therefore still early.
The knowledge and the tools belong to the organization
An agent is a system with the following parts.
- the model, which does the reasoning and the planning
- the harness, the software that steers the model through the work and calls the tools
- the sandbox, a closed environment that sets what the agent can reach
- the organization's knowledge, in its documents and databases
- the tools of the trade, the programs the organization does its work with
The first three are on the market. Models and harnesses can be bought as a service, and all three exist in open versions that anyone may download and run on their own computers.
Both kinds of model have their place. General work can go to a general model that is bought as a service, because such a model does that work very well. For work that rests on its own knowledge, a company takes an open model and has it trained further on its own data. NVIDIA calls the result sovereign intelligence.
The last two parts are what an organization brings itself. The companies in the hall heard that the knowledge and the tools belong to them, and that this is their expertise.
One example is a supplier that reports a delay. The message arrives as a document, and the orders are in a database. An agent that can read both predicts which orders will be late, and it searches the engineering documents for a replacement part.
What Europe makes sets it apart
The same holds for Europe as a whole. NVIDIA told its European audience that what they make and how they make it sets them apart. Europe designs machines, runs factories, and discovers medicines, and it does so with extraordinary craft.
Every sector therefore has to start using AI, from the car industry and manufacturing to healthcare and financial services. The slide that went with it listed the public sector as well.
The example of what that looks like came from the car industry. A door that was designed for the Volkswagen Golf has to be adapted for the Volkswagen Tiguan. One agent plans the work, a second reshapes the design, and a third tests the new door in a simulation. The engineer reviews the results and chooses the final design.
The next wave is physical AI. Physical AI is AI that understands the physical world and acts in it. An example from healthcare is equipment for radiotherapy that follows a tumor during the treatment.
Within physical AI, NVIDIA expects the humanoid robot to become the next large industry. Today every kind of task has its own kind of robot, and small numbers keep each of them expensive. A robot in the form of a person suits many tasks, so it can be built in large numbers.
The conclusion at the keynote was that this is Europe's great opportunity, because Europe has dedicated itself to mechanical and electronic systems with more precision and craft than any other region.
The first decision is where AI goes to work
The same reasoning holds for one organization. Every organization can buy the model, the software around it, and the closed environment. Its own knowledge and its own trade are the part that only it has, and in our view AI has the most to add there.
An AI strategy by Quilyx starts with the question where AI earns its place in the organization, and in which order. That is a decision for the board, and the picture from Berlin gives it a direction. The first place to look is the work that rests on the organization's own knowledge and on the tools of its trade.
The knowledge in that work is the organization's own, and so is the decision where it is processed and by whom. That is what Quilyx means by sovereign AI, and NVIDIA's sovereign intelligence is close to it. Sensitive data belongs on own hardware, under own control. Where a hosted service is good enough, we say so too.
We advise from our own practice. We build and run AI systems ourselves, on our own hardware in the Netherlands.
The phase in which AI does the work began in 2026, and it builds on what an organization already knows and makes. That is why Europe is still in time for AI, and why an organization that starts now is in time as well.
About this article: This article is our own reading of the keynote of NVIDIA GTC Berlin on 21 October 2026, with one remark from the session How to Scale With Open Source of the same day. The ideas from the keynote and from that session belong to their speakers and are named as such. 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.