ITARR Lab
Sovereign AI for a strong, independent Europe
What sits on a European workbench in 2045 depends on the choices we make today. Will the next shift still be able to repair the machine? Can the AI we add run on the computer already there? Who gets to build its replacement? At ITARR in Breda, we are designing our workspace, I (pronounced One), for the people who will inherit those decisions.
In this article
On a European factory floor, a maintenance engineer can tell from the sound of a machine which bearing is wearing out. That knowledge belongs to a person. It took years to acquire, and an apprentice can learn it at the engineer’s side. Suppose the factory captures that experience in an AI assistant. Who holds the knowledge then, if the next engineer can only read the maintenance history for as long as an overseas supplier keeps the account open?
Europe must be able to pass its knowledge to the next generation even if the supplier no longer exists.
What the car industry can teach us
Europe trained generations of engineers to get more power from a combustion engine while burning less fuel. Toolmakers, test facilities and specialist suppliers grew up around that skill. Electric vehicles shift some of that work into battery cells, power electronics and software. An engine maker can deliver the finest engine it has ever built on the very day its biggest customer places the next order elsewhere.
Europe still builds cars. In 2025 the EU produced nearly 3.2 million electric cars and remained the world’s second-largest producer. Tesla builds cars in Europe too. Yet Chinese producers supplied more than half the battery capacity deployed in electric cars in the EU that year. The assembly line tells only part of the story. Who designs the cell? Who learns from a failure? Which factory puts that knowledge into the next generation? IEA on manufacturing and trade · IEA on battery supply
In AI, those choices are still ours to make. Every European organisation using a model brings knowledge the model lacks. A maintenance firm knows its machinery. A hospital knows how its wards actually work. That knowledge can become the foundation of tools designed here, whose successors can be built here too.
The electricity meter belongs on the workbench
Suppose the night shift needs the same three pressure readings at every handover. A language model can turn them into three polished sentences. The machine already records the readings in its log. If a few lines of software display them with their timestamps and units, the crew can check the source and keep the existing computer in service for years. A language model might earn its place on a different question. What can an engineer learn from ten years of fault reports, when every shift described the causes in different words?
That is why our work begins by watching, sometimes for a whole shift. What does the crew already know? Where do they lose the thread? Only then do we choose how to combine their knowledge with software and computing power. In our next comparison we intend to measure electricity at the socket per successfully completed task, including retries. We also want to establish how much longer an organisation could keep its equipment in use. That puts care for the climate in the same design decision as quality and cost.
A continent of inventors and builders
In 1989, Tim Berners-Lee devised a way for researchers at CERN to find one another’s documents across different computers. In 1993 CERN made the web software freely available. A team elsewhere could run its own server and publish something without going back to Geneva for permission. That freedom turned a solution for one research institute into something the world could build on. CERN on the birth of the Web
Europe also has knowledge gained only by making exceptionally difficult things. The European Court of Auditors identifies equipment for advanced chipmaking as a European strength. A complete computer also depends on processor design, fabrication, memory and the software that makes those parts work together. It is at the junction of these disciplines that a European computer could be designed to stay in service for a generation. European Court of Auditors on the semiconductor supply chain
How much computing power does one question deserve?
‘Write the handover for the night shift’ sounds clear. ‘Use today’s maintenance record, carry over the three outstanding checks and leave the missing measurement blank’ gives the next engineer something to test. They can see whether the system invented a reading. In Meetlat, our model evaluation experiment, we give language models the same fictional task and agree in advance what makes an answer usable. A reviewer can assess the answers without knowing which model wrote them. That gives a small specialist model a fair hearing.
Some tasks need only a fixed rule. Our Dutch-language software turns different phrasings of a request to create a project into the same editable proposal. We observed zero external network requests in 45 browser checks of that demonstration. For open-ended questions we can run a language model on our own Mac. Our model switchboard, Wissel, checks whether there is room to switch models. On a machine with a configured limit of 48 GB, it refused a switch estimated to require 51 GB. For a hardware engineer, those three gigabytes raise a design question. Which model fits alongside the applications people already use, how long will they wait, and how much energy does each answer take? ITARR Lab on fixed rules and thinking colleagues
Who decides which chip is available?
In October 2023, NVIDIA reported that the United States had extended licensing requirements for exports of certain high-performance chips to destinations including China. The H800 was among them. For a European manufacturer, the precedent is concrete. A part can still be in the catalogue while a government elsewhere already has a say over its delivery. NVIDIA’s disclosure on the 2023 export controls
In late 2024, China’s DeepSeek described how it had trained its V3 language model on NVIDIA H800s. For each token it processes, V3 activates about 37 billion of its 671 billion parameters. Its designers found a way to direct computation to a fraction of a large model at a time. NVIDIA still supplied the chips. Together, those facts pose two different engineering challenges for Europe. One is to do more useful work with the equipment we have. The other is to be able to supply its successor. The DeepSeek-V3 technical report
The day after deliveries stop
Running a local model on American hardware gives us a choice today about what data leaves the machine. If Washington were to restrict sales of a processor to Europe tomorrow, our computer would still work. The dependency would become tangible when its motherboard failed or the next factory needed the same workspace. A backup would be worth having if someone here also held the replacement hardware, drivers, model files and keys needed to restore it.
The decisive test for European AI begins when a second team can replace the first machine.
That is the standard we have set for I. First, another team should be able to disconnect the prepared workspace and complete the same task. Next, the task should meet the same quality threshold on a second kind of computer. Time, memory and energy per completed job would show where a different chip design makes a difference that someone can measure. Finally, that team should be able to restore the workspace from a backup with its permissions intact. If the buyer has to sign a purchase order, a restored assistant must still have only the authority to draft it.
The engineer at that workbench may be someone we will never meet. She has the drawings, can replace the board and can establish for herself whether the machine is safe to use again. She can do all of that even if ITARR itself is gone. What Europe passes on to her is a working knowledge of her own machines, and the freedom to decide what to build with them next.