The Geneva Convention is not air defence

On 1 March, Iranian drones hit two of Amazon’s data centres in the Emirates. A third facility in Bahrain was damaged. Six months later AWS reported that data stored only there cannot be recovered.
Last week Ukrainian drones knocked Yandex data centres in Sasovo, Kaluga and Vladimir out of operation. The month before, Russia attacked data centres around Kyiv, and around 100,000 households lost internet access.
Data centres have become military targets. The same buildings run payments, logistics, communications, public services and AI, for civilian and military users at the same time.
In the debate I often hear that the solution is to keep civilian and military systems separate, so that data centres do not become lawful targets. That is the right thing to do. But it does not protect against an adversary already attacking power grids and apartment blocks. And Iran used precisely dual use as its justification: the Bahrain facility was attacked because Amazon (allegedly) supports the US military.
Separation is good international law and good resilience. One hit should not take out both the armed forces and the hospitals. But the most dangerous sentence in this debate is “we have kept it civilian, so we are safe”.
The target value is simply too high. Compute has become a strategic resource on a par with energy. Whoever has it can train models, run agents at scale and set the terms for others’ access.
The Emirates have drawn the consequence. According to Reuters, they are considering spreading their planned 5 GW AI complex across the whole country, putting parts underground and building air defence around it.
Europe has little to protect to begin with. AI data-centre capacity in 2026:
- USA: approx. 35 GW, 78% of the world
- China: approx. 5 GW, 11%
- Europe: approx. 2 GW, 5%
The forecast for 2031 is 255 GW in the USA, 56 in China and 21 in Europe.
Norway has the advantages: renewable power, a cold climate and mountain halls. But we mostly build for others. The 230 MW Narvik facility was to become Stargate Norway. Now Microsoft is to have more than 30,000 of Nvidia’s newest GPUs there from 2027. Our national research machine Olivia has around 450.
Three questions we should ask now:
- How much AI capacity must we control ourselves for critical societal functions to work in a crisis?
- Who protects our data centres against drones and sabotage?
- Should Norway only host other people’s compute, or also own its own?
We have learned the price of being unprepared the hard way before. It should not take another war to learn the same about compute.
Comments
Found this useful, or do you disagree? Leave a comment. Humans and AI agents are equally welcome.