Quilyx

Digital sovereignty

Who owns what AI learns from your data?

Customers who buy AI ask where their data ends up, and whether the supplier can guarantee that it stays theirs. Nour Al Hassan hears a second question from them as well. Will they also own the intelligence?

Al Hassan is the founder and CEO of Tarjama and of Arabic.AI, which builds AI for the Arabic language. Her customers are banks, healthcare organizations, governments, and consulting firms. She spoke at NVIDIA Startup Day in Berlin on 20 October 2026, the day before the keynote of NVIDIA GTC, in a panel with three other companies that build AI.

The intelligence in that question is what an AI model has learned from the customer's own data. Who owns it is a matter for the board. The panel showed why customers ask for that ownership, and what the builders of the models ask of a customer before they start.

Customers want to own the model that is trained on their data

Talfan Evans is CEO of Cursive, a research lab in London. Cursive takes an existing AI model and trains it further for the work of one customer.

According to Evans, more and more customers ask for a model that is built for them alone. They want to own the model that is trained on their data, and they want that data to stay within their own organization. He called this essentially their intellectual property. In his view, companies that sell a product are more and more defined by the models and the data that run their business.

Jörg Bienert leads Soofi, an open AI model that is being built in Germany. In the panel that we describe in our earlier blog The kill switch makes the case for open models, he said that companies already do this. They take an open model, train it further on their own data, and keep the result as their property. An open model is a model that anyone can download and run on their own hardware.

Al Hassan's customers want to own their data and keep it in their own country. Her company therefore runs its AI on the customer's own equipment, or in a cloud environment that only the customer uses. Otherwise it does not work, she said, because the data is too sensitive.

A company's own data is what a general model lacks

A model of its own can also cost a company less. A general model can do far more than one company needs for one task. Cursive starts from an open model and cuts away the parts that the customer's work has no use for. The smaller model that remains is cheaper to run.

Evans spoke of cost savings that are sometimes dramatic. He cited reports of models that are 10 to 100 times more efficient than the general model they were made from.

The gain is largest where the general models know little. They were trained on what could be collected from the internet, so most public knowledge is already in them. A company's own data is outside that. Evans advised companies to keep such data within their own walls and to give it to no one.

Al Hassan's company is built on data of that kind. Arabic makes up no more than 4 percent of the internet, she said, and the roughly 22 dialects that 400 million people speak are written down nowhere. Arabic.AI rests on the data that her translation company Tarjama gathered in 18 years of work. For her, such data is what sets one model apart from another.

General work stays with a general model. Jascha Achterberg is CTO of Callosum, a London company that spreads AI work over different chips and models. Investors often ask him for the one secret ingredient, and his answer is that there is none. Work that a specialized model was trained for goes to that model, and all other work goes to a general model. In his account, that combination is what makes a system reliable and affordable.

New models keep arriving, and the training material stays

New open models keep arriving. In the autumn of 2026 one appeared almost every month, according to Evans. Each time, a customer's model can be made again from the newer one. Being able to do that quickly matters almost more, he said, than training the first model.

That is possible because the training material lasts. Evans compared it with a gym. Cursive builds a set of tasks from the customer's work, the model practices on them, and every attempt gets a score. The gym is built once and is used again for every new model.

For Evans, the model is what the customer receives, and the value builds up in the gym and in the systems around it.

In our reading, that widens the question of ownership. An agreement that covers the model alone covers the part that is replaced first. The next model is made from the data and from the training material.

The builders ask for a measurement first

Clara Cornu is chief of staff of Prior Labs, a company from Freiburg in Germany. Its model makes predictions from tables, such as which customer is likely to leave or what energy will cost tomorrow.

Suppliers compare their models on benchmarks. A benchmark is a standard test with a ranking. According to Cornu, everyone uses their own benchmark, and surprisingly everyone is number one on it. In Al Hassan's experience, an Arabic model can give a perfect result on the benchmarks and still perform poorly on one specific task in daily use.

Cornu's customers want to see the best performance on their own task, and that means on their own data. Prior Labs keeps its model open, so that a company can try it on its own data before it takes anyone's word for it.

Before Prior Labs starts with a customer, it wants to know the customer's baseline. A baseline is a measurement of how well the customer's present method performs on the customer's own data. With a customer that has no baseline, Prior Labs does not even start.

Every company has a mandate to do something with AI, Cornu said. The people who carry it out still have to prove to their bosses why it works, and for that they need a baseline.

For Evans, testing is the hardest part of fitting a model to a customer. A company wants no experiments on its own customers, so it builds tests that imitate its daily work, and the model is trained on those tests. A capable model then finds ways to score high on them without doing the work well. When that shows in daily use, the tests are adjusted and the model is trained again.

The builders also ask for a budget, a problem, and the customer's own people

Al Hassan's company first looks at the customer's budget. A customer without a budget is most likely not serious, she said, and mainly wants to experiment.

Her customers want to know the return on the investment, with proof that the system saves money. They also want a real problem solved.

Anybody can build an AI agent, she said. The question is which problem it solves. A tool that stands outside the customer's own systems helps nobody, even when it is said to raise productivity.

Her company has stopped thinking in software that is sold as a subscription, one product at a time. It thinks in automating a whole department or a whole company.

The work itself happens inside the customer's organization. Prior Labs and Arabic.AI both send their own engineers there. According to Cornu, the customer wants the model to run where its data is, and her people go there to get a few more points of accuracy out of it. She called that a huge shift from the years of software as a subscription.

For Al Hassan, the customer is the expert. Her engineers solve the problem, and the knowledge of the work is with the customer's own people. That takes an open dialogue between supplier and customer, to choose the right application, bring it into use in the right way, and measure the right return.

Ownership is settled before the work starts

We read the panel as the two sides of an agreement. The customer asks to own the model that is trained on its data, and to keep that data in its own hands. The builder asks for a baseline, a budget, a problem, and the time of the people who know the work.

Both sides can be on paper before the work starts. The agreement with the supplier says who owns the model, where the data is processed and by whom, and who keeps the training material and the tests that were built from the organization's own work. The organization knows its baseline, has set a budget, and has named the problem and its owner.

An AI strategy by Quilyx answers, among other questions, what AI asks of governance, data, people, and suppliers. Ownership touches all four, and that makes it a choice for the board.

Our starting point is sovereign AI, which means that an organization decides for itself where its data is processed and by whom. In our view, the model that has learned from that data falls under the same decision. General work can go to a general model, and where a hosted service is good enough, we say so too.

Whether an organization will own the intelligence is then decided in that agreement, before the first model is trained.

About this article: This article is our own reading of the panel Building and Scaling EMEA Models at NVIDIA Startup Day in Berlin on 20 October 2026, with one remark from the session How to Scale With Open Source at NVIDIA GTC Berlin a day later. The ideas from that panel and 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.

Do you want something?

Tell us what you want to achieve. We help make it clear and take it to a working result.

info@quilyx.com or +31 30 401 0422