Case 01Tensorfield2026 · Pre-training
A 70B foundation model, trained in 19 days on 4,096 H200s.
−34%Energy cost vs. previous region
Energy / run€1.42M €0.94M
Time to train24 d 19 d
Press M to toggle
Training a frontier model now draws as much power as a small city. Up to forty percent of it goes to cooling, in places that are already too hot. We chose a place where the air does that work for free.
68.4385° N · Narvik, NorwayOur first hall sits in a decommissioned mine above the Ofotfjord. Outside air below 10 °C for 300 days a year means no chillers, only fans. Waste heat warms the town's district network.
Dedicated H200 and B300 pods from 256 GPUs, on one-to-three year terms.
From $1.62 / GPU-hrSpin up 8 to 512 GPUs in minutes. Billed per second, no commitment.
From $2.24 / GPU-hr
Private halls for governments and labs. EEA jurisdiction, your keys, your staff.
By arrangement
Case 01Case 02Case 03Average power usage effectiveness across 2025, metered hourly.
Hydro and wind from the NO4 grid, matched by the hour.
Free-air cooling. No evaporative towers, no chilled water loops.
Delivered to Narvik's district heating network last winter.
“We moved pre-training to Narvik in March. Same throughput, and our energy bill fell by a third.”
For research teams, fine-tunes and evaluation runs.
Dedicated pods for production training, 256 to 16,384 GPUs.
Isolated halls under EEA jurisdiction for governments and labs.
Prices exclude VAT. B300 pricing on request. Power and cooling included.
Something went wrong. Please try again, or email hello@example.com.
