Arctic Circle Expansion: Heating Svalbard One Inference at a Time
We're excited to announce the launch of polar.sv1.arctic—our northernmost region, located on the remote Norwegian archipelago of Svalbard at 78°N latitude. This is the first GPU cluster in history to be classified as "polar-adjacent infrastructure."
Why Svalbard?
When our site selection team proposed building a data center above the Arctic Circle, we had several concerns. Then they showed us the heating bills for Svalbard's research stations, and suddenly it all made sense.
The polar.sv1.arctic region now provides dedicated, high-speed, AI-enhanced compute to:
- 89 arctic dwellings including research stations
- The Global Seed Vault (they declined our heating services)
- 3 polar bear observation posts
- 1 very confused weather station
Technical Specifications
Despite its remote location, polar.sv1.arctic offers enterprise-grade capabilities:
- 1,024× B200 GPUs — our first Blackwell deployment
- 192 TB of combined vRAM
- 480 Tbps throughput via arctic fiber cable
- 0.8ms latency (to mainland Norway)
- -40°C ambient cooling (free!)
The Unique Challenges of Arctic AI
Operating a GPU cluster in the Arctic presented some unique challenges:
- 24-hour darkness in winter: Our engineers developed seasonal affective disorder within weeks. We've since installed many sad lamps.
- Polar bear security: Our physical security team now includes Norwegian wildlife rangers. Badge access is mandatory; being a polar bear is not acceptable ID.
- Supply chain: GPU replacements arrive by icebreaker ship twice per year. Plan your hardware failures accordingly.
- Internet connectivity: We laid a dedicated fiber cable from the Norwegian mainland. It was expensive. Please use the region so it was worth it.
Eco-Friendly by Necessity
In Svalbard, our heat recovery isn't just eco-friendly—it's survival-critical. The cluster's thermal output provides 100% of heating needs for the connected research facilities. When inference demand drops, scientists literally get cold.
We've implemented "thermal SLAs" guaranteeing minimum heat output. If your AI workload isn't keeping researchers warm enough, we'll run dummy computations until it does. You'll still be billed, of course.
Pricing
Due to the heroic logistics required to operate in the Arctic, polar.sv1.arctic pricing includes a modest "polar premium":
- B200 compute: $129.00/GPU-hour (+45% polar premium)
- Storage: $0.89/GB-month (icebreaker shipping costs)
- Egress: $0.15/GB (arctic fiber isn't cheap)
We accept that this pricing is particularly unaffordable. Consider it a donation to arctic research.