Recently, Logi Einarsson, Minister of Culture, Innovation and Higher Education, announced plans to develop a large language model (LLM) specifically tailored to the Icelandic language and conditions and hosted in Iceland. The plans have been associated with “Sovereign AI” and have parallels in other countries that are concerned about their security and independence from foreign tech giants, particularly American ones.

Most of those who have publicly commented on these plans seem to support the policy of Iceland developing its own AI model. However, there has been considerable discussion about the different paths that can be taken to achieve this goal.

From the minister's article, one can conclude that the government intends to develop the model in close cooperation with a European partner. The company is not named, but it is presumably the French company Mistral, which is a leader in this field in Europe and has specialized in such collaborations with nation-states around the world.

If an agreement is reached with Mistral, it would have several advantages: training could take place on Mistral's hardware, the model would come with a training and operational environment , reducing the complexity of making the model usable and accessible, an experienced team would work with local parties to ensure both knowledge transfer and a successful outcome, data and data pipelines in Mistral's possession would be utilized instead of extensive data cultivation taking place in Iceland, and so on.

On the other hand, the price tag for such a service is likely quite high, and judging by leaderboards for the general capabilities of large language models it is clear that Mistral's current generation models do not fully compete with open models such as Kimi 2.5, Qwen 3.5, and DeepSeek (note, however, that Mistral has commenced training on the fourth generation of its models with Mistral Small 4 already released and larger versions on their way). Furthermore, according to Miðeind's leaderboard for the Icelandic capabilities of large language models, some of these Chinese open models are better than or similar to Mistral Large 3 in Icelandic proficiency, although none of these models come close to the best closed models in that regard.

It is therefore natural for people to wonder whether it would be more sensible to use the funds earmarked for this cause to fine-tune one of these powerful open models. Such a project would broadly involve preparing Icelandic fine-tuning data that would be open to everyone, fine-tuning a model (after continued pre-training assuming that the goal is to maximize Icelandic skills) on purchased or leased computing power, building an operational environment and various software needed to make such a model accessible on a large scale, and managing maintenance, updates, and operations.

The goal of this essay is not to pass judgment on which approach is most sensible in these matters. The purpose is simply to provide a realistic insight into the scope and complexity of fine-tuning an open-source model, from the perspective of a company with the most experience in training language models in this country.

To better understand the true scope of such a project, it is wise to look for precedents. One such example can be found in a Japanese project called Swallow LLM, a collaboration between the Tokyo Institute of Science (a private initiative) and AIST, Japan's National Institute of Advanced Industrial Science and Technology. The goal of the project, now in its third year, is to develop a powerful Japanese reasoning model based on an open model, most recently Qwen3, and the results so far are promising. It is an iterative process where the team has continued pre-training ever-larger models (8B -> 32B) with a specially prepared Japanese dataset of about 200 billion tokens, and then post-trained and fine-tuned the model on specific data, such as instruction tuning data. Training has taken place on a supercomputer owned by the Japanese government, as well as on subsidized computing power through the university community. The outcome is a competitive model with good Japanese language skills, but it should be noted that the project is still in the research phase and has not yet entered general operation.

What would a project like this cost? First, it should be mentioned that the data was prepared using various methods: web scraping, translating foreign datasets, and custom-built synthetic data pipelines, to name a few. An Icelandic team would need to do the same, but in the case of Icelandic, it is not realistic to collect 200 billion tokens for pre-training, so it would have to be a mix of Icelandic, Nordic, and possibly Polish data. In addition, specific fine-tuning and post-training data would need to be prepared, either from scratch or by translating foreign datasets. If the intention is to complete data collection in one year, 8-10 full-time positions can be expected for this work, mainly engineers. It is not clear which Icelandic parties have both the desire and the capability to undertake such a project, but the cost would likely be around 200 million Icelandic krónur based on these assumptions.

Next is the computing power. Iceland does not have its own supercomputer, so there are two options available here: either buy the equipment or rent it. Assuming that continued pre-training requires a compute cluster of 32 nodes, the cost of buying and operating the computer during training would run from one and a half to two and a half billion krónur, depending on how advanced the equipment purchased is (DGX or not). The current wait time for that amount of hardware is several months long. The rental cost through cloud solutions would be much lower, probably around 100 million krónur for this part of the project (based on a price of around 400 krónur for one rented GPU hour, assuming a minimum requirement of 100,000 GPU hours for pre-training alone, plus post-training and fine-tuning, and a relatively inexperienced team not achieving full utilization of the compute).

.

Labor costs for training would be at least as high, and likely higher, than the cost specified for data processing. This is based on an 8-person team working full-time for 12-18 months. This is not a standard software development project; rather, it requires deep learning experts with experience in distributed training on large compute clusters, as well as specialized knowledge of methods such as Reinforcement Learning from Human Feedback (RLHF) and model alignment. Such expertise is in short supply globally and is priced accordingly, as designing complex training pipelines requires significant experience to navigate challenges like 'catastrophic forgetting'—where a model might lose its existing capabilities and reasoning skills while being adapted to the Icelandic language. On top of this, there would be electricity costs, housing costs, etc.

The initial cost then runs to 0.5-3 billion krónur, and it can be assumed that the timeline will be at least  2-3 years, similar to the Japanese project.

To run such a large and powerful model, you need software, personnel, and computing power. The computing power needed to run the model (inference hardware) is much less than what is needed for training, and here it would be possible to use the services of data centers that already exist in Iceland. The cost would likely run to 40-70 million krónur per year, based on considerable and general use. Alternatively, it would be possible to buy equipment for operation for 150-300 million krónur.

To build all the software around operations (interface, access control, APIs, security measures, billing system), a team of 8-10 software engineers would be needed for one year, which would likely cost around 150-200 million krónur. A team of the same size would then be needed to maintain the model and operations.

In all of the above, it is assumed that the knowledge and skills to train, handle, and operate large language models are generally available domestically. If that is not the case, it would be necessary to either hire a team of foreign experts or extend the project timeline and increase the cost estimate accordingly to allow for more iterations and costly mistakes.

As one might imagine, the risk with this approach is much greater than if working with an experienced solution and team (like Mistral). On the other hand, the outcome comes closer to what could be considered sovereign AI (assuming that the Chinese origin of the base model is not a deal-breaker in that regard). Whether the price tag is lower than with the Mistral route is not clear, as those figures are not yet public.

It is a cause for celebration that a public policy on artificial intelligence is taking shape in this country, and it is even more gratifying to see the apparent consensus on the need for sovereign AI. Important national interests are at stake, and there is every reason to discuss the pros and cons of different paths to this goal, as long as it is done objectively and the discussion is based on knowledge and facts, not shouting and innuendo. Technological independence is a precious ideal, but we must also approach the task with our eyes open to the fact that true sovereign AI requires not only political will and funding but also, not least, deep expertise that is not available domestically and therefore needs to be built up in one way or another. Such knowledge-building is in itself a valuable investment and should not be underestimated when evaluating the pros and cons of different methodologies.