Why Seyðisfjörður is back on my radar

Back in 2016, I published a novel called Óminni. In it, I imagined a remote data center nestled in Seyðisfjörður, a quiet fjord town in East Iceland, powering global simulations for insurance giants and commodities markets. I chose that setting because it offered what few other places do: a direct fiber cable to mainland Europe (FARICE-1), abundant hydropower from Kárahnjúkar, a cold climate perfect for efficient cooling, and a setting that felt almost mythic. 

What amazes me now is that AI never even crossed my mind as a use case for that data center. It wasn’t on the radar, not even in a fictional thriller trying to peer into the near future. That’s how fast things have changed. Looking back, the scenario I imagined seems almost naive compared to what’s unfolding now. And the scale of what’s coming next is on a completely different level.

AI has since evolved from niche technology to civilization-scale infrastructure. Training a model like GPT-4 consumes around 25 GWh, equivalent to powering every home in Iceland for more than a week. And that’s just one model. As the next wave of AI megafactories comes online, global electricity demand from AI could surge by 10–50x over the next decade. Analysts project that by 2030, AI alone could consume as much electricity as some medium-sized countries. And yet, Big Tech: OpenAI, Microsoft, Google and Meta have all committed to carbon neutrality by that same year. That’s five years away. How can those goals possibly align with exponentially growing compute demand?

The answer is emerging in the very architecture of the AI infrastructure itself. It’s beginning to split: training and inference are no longer co-located. Training foundation models requires colossal energy and cooling but can be done anywhere — it doesn’t require real-time responsiveness. All it needs is clean energy, stable infrastructure, geopolitical predictability, strong fiber access… sound familiar? Inference, which serves up AI responses to billions of users, must happen closer to the edge. Fast, light, and globally distributed.

This separation is already reshaping how the industry thinks. And it opens a profound opportunity for Iceland, a country that generates nearly all its electricity from renewable sources, with a cool climate, low political risk, strong digital infrastructure, and real places like Seyðisfjörður. A town I used as a key setting in Óminni, and which now seems symbolic of the opportunity we’re facing.

But here’s the reality: we’ve lost momentum when it comes to energy development in Iceland. Despite our renewable potential, we’ve allowed indecision, delay, and short-term politics to stall progress. New power projects are regularly postponed or blocked, not because we lack the resources but because we haven’t aligned around a clear strategy for the future.

There’s still time, but only if we move quickly and decisively. We need to expand our clean energy production, not for growth’s sake, but to fuel what the future actually demands. AI, advanced computing, scientific research, these are not just tech trends. They are shaping the backbone of tomorrow’s economy and society.

I’m not arguing that Iceland should try to power the entire AI industry, we can’t and we shouldn’t. But we could take responsibility for the most energy-intensive and climate-critical part of it: the model training that underpins everything else. That part doesn’t need to happen near end users. It needs to happen where power is clean, cooling is efficient, and infrastructure is stable. Iceland fits that profile better than most places on the planet.

When I wrote Óminni back in 2016, I imagined a fictional town hosting a massive data center. No one really knew what it was doing. There were protests. Suspicion. Tension between the local community and the outside world trying to extract value from the landscape. At the center of the story was a man who had lost his memory and couldn’t move on because he didn’t know where he came from.

Óminni was a thriller, written to entertain, not predict. But looking back, some of the patterns feel surprisingly familiar.