# Nordic AI Projects Span a 1,152-GPU Supercomputer and Local-Language Tools

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/32304
Published: 2026-10-15T05:52:14.000Z
Updated: 2026-10-15T05:52:14.000Z
Section: Technology

> Sferical AI operates a 1,152-GPU system in Sweden, while a separate Icelandic GPU deployment is planned for 2026. Developers in Sweden and Finland are adapting models for language tasks.

Sferical AI operates a 1,152-GPU system in Sweden, while a separate Icelandic deployment plan and Nordic language tools show AI activity spanning infrastructure and software.

Sferical One is a shared supercomputer built from two NVIDIA DGX SuperPODs. The equipment spans 16 racks configured for GB300 NVL72 systems and totals 1,152 GPUs. AstraZeneca, Ericsson, Saab, SEB and Wallenberg Investments founded Sferical, with SKF as a strategic partner.

Sferical announced its launch on Aug. 22, 2025, with Linköping, Sweden, named as the system’s intended site. The Iceland project has different operators and its own 2026 deployment target.

Nscale Limited (Nscale) and Verne announced the Iceland project on Nov. 20, 2025. The plan calls for 15 megawatts of infrastructure at Verne’s Icelandic campus and about 4,600 NVIDIA Blackwell Ultra GPUs, with deployment targeted during 2026. Its cooling design assigns 85% to liquid cooling and 15% to air cooling.

The infrastructure projects center on GPU capacity, while the software work adapts existing NVIDIA models for specific language tasks. Speech-to-text turns speech into text, while retrieval systems search for relevant information across languages.

Klang AI AB (Klang), a Swedish company, developed Pianissimo as an open speech-to-text model fine-tuned from NVIDIA’s Parakeet v3. Customers can deploy Pianissimo on their own infrastructure or access it through Swedish provider Berget AI. Pianissimo records a 4.5% word error rate on Common Voice and performs comparably to KB-Whisper large at 64 times the speed.

ConfidentialMind Oy (ConfidentialMind), based in Espoo, Finland, developed SampoTron, an adapter for NVIDIA’s Nemotron-3-Embed-1B. The tool is designed to improve retrieval across English, Finnish and Swedish. SampoTron’s routed adapter scored 0.2908 on cross-lingual FIQA nDCG@10, up 20.9% from the base model’s 0.2406.

The efforts span Sweden, Iceland and Finland, from an operating supercomputer and planned infrastructure to software for speech and multilingual retrieval. Nscale and Verne’s 2025 announcement set 2026 as the Iceland project’s deployment target.