Nordic AI Infrastructure Surge: New DGX SuperPODs and Open‑Model Deployments for Engineers

The Nordic region has moved from planning to concrete deployment: a consortium in Sweden launched the Sferical One supercomputer built on two NVIDIA DGX SuperPODs, and AI cloud providers are targeting a total of 1 GW of GPU capacity across the Nordics by the end of 2027. For AI, cloud, and DevOps teams this means a new, high‑density compute pool that can be accessed for large‑scale model training, inference, and simulation workloads.

New Compute Assets in the Nordics

Key hardware additions include:

• DGX SuperPOD clusters with 16 racks of DGX GB300 systems and 1,152 GPUs, hosted by a consortium of Ericsson, AstraZeneca, Saab, SEB, Wallenberg Investments, and SKF.
• Projected AI cloud capacity of 1 GW of GPU power from providers such as Bitdeer, CoreWeave, Crusoe, Hyperstack, Nscale, Verda, and Volta, slated for 2027.

These assets provide a regional backbone for workloads that previously required private on‑premise clusters or distant public clouds.

Model Deployment Patterns Emerging

Several Nordic companies are already leveraging NVIDIA software stacks on this infrastructure:

• Spotify uses TensorRT‑LLM for podcast recommendation inference, achieving lower cost per query at scale.
• ABB Robotics integrates Omniverse libraries into RobotStudio HyperReality, enabling near‑real‑time simulation with up to 99% accuracy and an 80% reduction in commissioning time.
• Startups such as Stilta, ConfidentialMind, Agaton, and Klang are fine‑tuning NVIDIA Nemotron open models (including Nemotron‑3‑Embed and Nemotron 3.5 Lightning) for patent analysis, cross‑lingual retrieval, revenue mining, and Swedish speech transcription.
• The Embedl prototype runs on a single Jetson Orin NX module to monitor protein crystallization without continuous ground connectivity.

These examples illustrate a shift toward open‑model pipelines that run on shared GPU clusters, reducing the need for bespoke hardware while demanding robust model versioning and data isolation.

Operational and Security Considerations

Running workloads on a multi‑tenant supercomputer introduces several practical concerns:

• Isolation: Shared GPU resources require careful scheduling and namespace isolation to prevent cross‑tenant interference.
• Data Governance: The region’s strict data regulations mean that any model training or inference must respect locality and access‑control policies, especially when using proprietary datasets.
• Latency and Cost: TensorRT‑LLM and Nemotron‑based inference can lower per‑query cost, but engineers must benchmark end‑to‑end latency to ensure real‑time requirements (e.g., Elisa’s Moontalk AIRI).
• Observability: High‑throughput GPU clusters benefit from centralized telemetry (GPU utilization, temperature, job queue depth) to avoid bottlenecks and to support SRE alerting.
• Model Security: Open models are customizable, so teams need to enforce model provenance checks and guard against inadvertent leakage of sensitive prompts or training data.

Related CloudNinjas coverage: AI engineering.

What This Means For Practitioners

Engineers should start evaluating access to the Nordic AI fabric as a viable alternative to traditional public clouds. Immediate actions include:

1. Prototype a workload on TensorRT‑LLM or Nemotron to measure cost per inference versus existing deployments.
2. Review GPU scheduling and namespace isolation mechanisms in the DGX SuperPOD environment to align with internal security policies.
3. Integrate observability pipelines that capture GPU metrics and model latency for SRE dashboards.
4. Map data residency requirements to the Nordic providers’ compliance statements before ingesting regulated datasets.

By treating the new Nordic compute pool as a regional edge for AI workloads, teams can achieve higher performance, lower inference cost, and tighter control over data governance.