Norway's National Library is training a large language model specifically for Norwegian language processing, leveraging 2 PB of Huawei OceanStor Dorado flash storage in its data pipeline. Marius Husnes, Head of IT Platform, presented the project at Huawei's ID Forum 2026, noting that commercial providers ignore local language models.

The library's 20 PB unique digital collection, gathered since 2005 through legal deposit mandates, requires significant preprocessing before training. Husnes identified data quality and pipeline throughput—not compute—as the primary bottlenecks. Their infrastructure combines an Nvidia DGX H200 system with 384-core CPU clusters and multiple Huawei flash arrays for initial processing stages.

Training runs on Norway's Sigma2 Olivia supercomputer, an HPE Cray system with 448 GPUs and 64,512 CPU cores. A major challenge involves bridging two storage systems: the 60 PB preservation archive optimized for durability versus the low-latency AI pipeline storage designed for parallel I/O operations.

The initiative highlights broader questions about sovereign AI development for non-English speaking nations. Norway's experience with evaluation frameworks, governance structures, and multi-system orchestration offers valuable lessons for countries seeking culturally relevant language models.