DCAI Gefion: Denmark’s sovereign AI supercomputer on DGX SuperPOD

Before Gefion, Denmark had no GPU-accelerated supercomputer, according to the Novo Nordisk Foundation. The foundation adds that a national consultation with stakeholders had identified access to computing power as a key obstacle for Danish AI research.¹

Gefion is built on the NVIDIA DGX SuperPOD architecture with 1,528 NVIDIA H100 Tensor Core GPUs connected by NVIDIA Quantum-2 InfiniBand, as stated by the Novo Nordisk Foundation and NVIDIA. The TOP500 entry lists the interconnect differently, as octo-rail NVIDIA HDR100 InfiniBand. Digital Realty hosts it in a data center running on renewable energy, and Eviden assembled it.

DCAI is a company formed by the Novo Nordisk Foundation (about DKK 600 million toward initial costs) and the Export and Investment Fund of Denmark, EIFO (DKK 100 million and a 15 percent stake). The system opened on 23 October 2024 with a pilot phase for six projects, including quantum circuit simulation at the University of Copenhagen, a genomic foundation model, startups Go Autonomous and Teton, and weather prediction at the Danish Meteorological Institute.

By November 2025 DCAI was serving industrial customers: Siemens Gamesa selected Gefion for wind farm AI projects, and DCAI described the platform as NVIDIA DGX systems with a WEKA data platform hosted in Denmark.¹²³⁴

What others can learn: a foundation and a state investment fund can jointly fund a national GPU system, run it as a company, and open it to academia, startups and industry through a short pilot phase. Limits: Gefion depended on a very large private foundation grant; few countries or regions have an equivalent funder. The TOP500 placement measures the HPL benchmark, not AI workload performance, and most application results in the cited sources are goals or expectations; NVIDIA reports one capability gain, a quantum circuit simulation that grew from 36 to 40 entangled qubits on Gefion. The cited sources report no utilization figures or measured customer results. Organizations needing occasional capacity can rent it from a national system like this or a cloud provider rather than build one.