The Nordic countries want to shift their role in the European digital landscape: from a fertile breeding ground for startups to a platform capable of hosting computing infrastructure, applied research, and artificial intelligence-based products. This is the picture outlined by NVIDIA, which points to a set of favorable conditions in the region—renewable energy, stringent data regulations, and close cooperation between industry and academia—converging to attract investments in cloud, supercomputing, and AI software.
This premise is far from marginal for the startup ecosystem. According to NVIDIA, the Nordics lead Europe in unicorn density—private companies valued at at least one billion dollars relative to population. However, producing new high-growth companies does not automatically translate into having the resources required to train, serve, and keep advanced models competitive. In the view of the region's industrial players, the next phase hinges precisely on the availability of compute power and the ability to deploy it within controlled environments.
Compute becomes part of industrial policy
Infrastructure growth is one of the most tangible aspects of this transformation. NVIDIA estimates that by the end of 2027, one gigawatt of capacity dedicated to AI clouds could be deployed across the region. The operators cited include Bitdeer, CoreWeave, Crusoe, Hyperstack, Nscale, Verda, and Volta. While differing in business models and geographical footprint, these players share the race to build data centers and provide GPU services in high demand from enterprises, developers, and research centers.
The figure should be viewed with caution: it is a projection presented by the US chipmaker, not capacity that is already online or a definitive measure of actual future demand. Nevertheless, it signals the scale the Nordics are striving to achieve. In Europe, where the debate over technological sovereignty has become more concrete with the rise of generative models, having local access to infrastructure can influence where products are developed, talent is retained, and sensitive datasets are managed.
In Sweden, the Sferical One project exemplifies the direction being taken. The consortium brings together Ericsson, AstraZeneca, Saab, SEB, and Wallenberg Investments, with SKF as a strategic partner, around a shared supercomputer. The infrastructure utilizes two NVIDIA DGX SuperPODs, totaling 16 racks of DGX GB300 systems and 1,152 GPUs. The stated goal is to develop services based on NVIDIA Nemotron open models, allowing researchers and companies to adapt and deploy them within an environment deemed secure.
The value of such a platform does not rely solely on the number of installed GPUs. What matters is who can access it, under what economic conditions, and with what data processing guarantees. A shared infrastructure can bridge the gap between large corporations, innovative SMEs, and research, but it does not automatically eliminate hurdles: the cost of inference, the availability of specialized talent, data quality, and competition from major global cloud providers remain decisive factors.
From music to robotics, AI enters operational processes
The Nordic strategy is not limited to building capacity. NVIDIA highlights cases where infrastructure is already translating into enterprise applications. Spotify uses TensorRT-LLM to make the inference behind podcast, music, and audiobook recommendations more efficient. For a platform serving hundreds of millions of listeners, lowering the cost per request while maintaining fast response times is an operational imperative, not an experimental exercise.
ABB Robotics, on the other hand, developed RobotStudio HyperReality leveraging NVIDIA Omniverse libraries. The solution combines simulation with a virtual controller—a controller that uses the same firmware as physical robots—with the goal of testing and validating configurations before factory installation. ABB claims up to 99% simulation accuracy and up to an 80% reduction in commissioning times. These results are reported by the company and naturally depend on the facility type and use case, but they demonstrate where AI and digital twins can make a measurable impact: in commissioning times and production-phase risks.
Also coming from Finland is Moontalk AIRI, an enterprise assistant that Elisa is preparing in beta. The system is designed to retrieve information, summarize key takeaways, and document decisions during meetings and calls. It is powered by Nemotron models and a low-latency distributed infrastructure. In this case, the challenge is less about the model's flashiness and more about reliability in enterprise environments: access to accurate data, protection of internal information, and useful answers within the workflow.
Startups carve out space in local use cases
The availability of open models and computing resources can carry particular weight for startups, which can rarely sustain the development of proprietary foundation models on their own. Instead, the ability to adapt existing models to specific languages, documents, and procedures enables the creation of vertical products, where the differentiator is not just the algorithm, but the quality of integration with a concrete domain.
Stockholm-based Stilta uses Nemotron for intellectual property analysis. The company aims to expand monitoring from individual patents to much broader portfolios, a capability valuable to industrial and technology groups that must track thousands of assets. ConfidentialMind has developed SampoTron, based on NVIDIA Nemotron-3-Embed, and claims an improvement of over 20% in multilingual retrieval accuracy across English, Finnish, and Swedish.
Agaton focuses on real-time customer conversations and employs Nemotron 3.5 Lightning for its Revenue Mining system during sales calls. The stated goal is to reduce latency when the software retrieves actionable knowledge to support sales. Klang, a Swedish laboratory specializing in voice, has instead adapted the Parakeet v3 model to build Pianissimo, a Swedish transcription system. According to the company, the model cuts errors by 75% and is 64 times faster than the most accurate alternative used as a benchmark; it was trained on over 50,000 hours of Swedish speech and is available on Hugging Face.
These projects highlight an important characteristic of the regional market: a focus on Nordic languages and on sectors where local data and regulatory frameworks can become a competitive barrier. Patents, commercial communications, transcription, and industrial automation do not necessarily require a larger model; they often require reliable, fast models tailored to a defined linguistic or professional environment.
Sweden, in particular, raised $2.8 billion for startups in the first half of 2026, with NVIDIA projecting $5 billion for the full year. Here too, it is worth distinguishing between capital raising and the ability to build enduring businesses. Yet the financial inflow, combined with infrastructure investments, suggests that the region is trying to connect venture capital, major industrial groups, and research rather than treating them as separate silos.
A race not decided by GPUs alone
For Europe, the Nordic case could become a testing ground. The region starts with tangible advantages, but it also faces fierce competition from the United States and Asia, where capital, hyperscalers, and semiconductor supply chains operate on a larger scale. Having data centers powered by renewable energy or shared supercomputers does not, on its own, guarantee technological autonomy: it requires industrial demand, clear rules, skills, and products capable of winning customers beyond national borders.
What emerges from the projects outlined by NVIDIA is therefore less a simple hardware expansion and more an attempt to turn advanced computing into a pervasive economic infrastructure. If industrial consortia, regional clouds, and startups can translate this capacity into reliable services, the Nordics could cement a European role not just in the birth of scaleups, but in their most difficult phase: growing without having to move data, workloads, and strategic expertise elsewhere.
