How do we ensure that the artificial intelligence of the future understands the Danish language, Danish institutions, and Danish society – while handling data within frameworks that we control ourselves?
This question lies at the heart of the research project Danish Foundation Models (DFM), where the University of Southern Denmark, Aarhus University, the University of Copenhagen, and the Alexandra Institute are collaborating to develop open Danish language models. The project has been underway for some time, and researchers have already released the first models, established benchmarks, and initiated collaborations with external partners. Now, the project is entering its next phase.
With access to the new national AI supercomputing facility BITTEN, inaugurated in Sønderborg in May, and accessible through the research platform UCloud, the pace of development can increase significantly.
Culture, norms, and societal understanding are lost
Within just a few years, language models have become a strategic technology. They are already used for text generation, search, decision support, automation, and analysis. However, the most widespread models have been developed by global companies and trained primarily on English and other major languages.
This creates limitations when such models are expected to operate within a Danish context.
“If you look closely at the details, many international models are actually poor at Danish. The way they formulate themselves often resembles English translated into Danish. That is not how we speak or write,” says Professor Peter Schneider-Kamp from the University of Southern Denmark, who leads DFM on behalf of SDU.
The challenge extends beyond language. It also concerns culture, social norms, and an understanding of society.
DFM has, among other things, developed Danish benchmarks that test models on knowledge of Danish culture. Here, even the largest international models often perform poorly.
“They lack an understanding of how Denmark works – our literature, our public institutions, our healthcare system, and our cultural points of reference,” Schneider-Kamp explains.
Denmark cannot remain on the sidelines
The development of AI is moving so rapidly that access to domestic expertise and infrastructure is becoming increasingly important. According to Schneider-Kamp, it is risky to assume that other countries will continue to provide the services and models that Europe needs indefinitely.
“We cannot simply rely on American or Chinese companies to provide the right solutions for us forever. We need to take part in the development ourselves,” he says.
This is not about replicating Silicon Valley at the same scale, but about being able to develop solutions tailored to Danish needs – and doing so on a transparent and responsible foundation.
“We want models where we know what they have been trained on, that comply with GDPR and the AI Act, and that understand the Danish language, Danish culture, and Danish norms,” he says.
UCloud is the backbone
Behind the project lies a less visible, but crucial part of the story: research infrastructure.
DFM is developed to a large extent using UCloud, the national platform for interactive high-performance computing developed by SDU eScience Center together with partners. It provides researchers with access to storage capacity, GPUs, software, and collaboration tools in one integrated environment.
For Peter Schneider-Kamp, UCloud is absolutely central.
“UCloud is our secure environment where we develop models, train models, store data, and evaluate them. It is absolutely central to everything we do,” he says.
The security perspective is essential. When researchers work with large amounts of data – and in some projects also sensitive data – it is critical that the data can be handled within a controlled environment.
“If we receive new datasets from, for example, libraries, media organisations, or other sources, we can store them securely, work directly on them, and maintain control over the data,” he says.
The alternative is often more complex and fragmented solutions where data has to be moved between systems and countries.
New national supercomputer BITTEN
On 5 May 2026, a new national supercomputer was inaugurated in Sønderborg: BITTEN. The facility was established by the University of Southern Denmark in collaboration with Danfoss and Hewlett Packard Enterprise (HPE) and forms part of the Danish research infrastructure for artificial intelligence, advanced computing, and data-intensive research.
The supercomputer is made available through UCloud, allowing researchers and students at universities across Denmark to access it through their existing systems and workflows.
The collaboration combines SDU’s experience in research infrastructure, Danfoss’ expertise in energy-efficient cooling and heating solutions, and HPE’s knowledge of supercomputing and data centre technology.
The facility has also been designed with a strong focus on energy-efficient operation and the reuse of excess heat.
Lack of computing capacity slows down research
For AI research, access to computing power is not a luxury – it is a prerequisite.
Previously, the DFM group regularly experienced bottlenecks when training models.
“Sometimes we have had to wait two, three, four, or even five days to get access to a GPU. Meanwhile, PhD students and postdoctoral researchers are sitting ready with ideas and code but are held back by a lack of resources,” says Peter Schneider-Kamp.
This is precisely where the new capacity can make a difference.
More computing power means faster experiments, larger models, more iterations, and a shorter path from idea to result.
“We hope it will allow us to turn ideas into research results and concrete use cases much faster. We are incredibly excited to gain access to the many additional GPUs,” he says.
More than technology
DFM is therefore about more than software and hardware. The project illustrates how research, digital sovereignty, data security, and innovation are closely connected.
If Denmark wants to use AI in healthcare, the public sector, education, and industry, it requires solutions that can be understood, adapted, and trusted.
With UCloud as its operational backbone and the new national supercomputing capacity of BITTEN, Danish Foundation Models now stands at a point where the work can move from a promising development phase to broader implementation.
The question is no longer only whether Denmark can develop its own language models.
The question is whether we can afford not to.
