Iceland went into production on 6 June. KEF-01 sells the Ryzen line and storage nodes, on a 100 Gbit/s uplink, with 4 Tbit/s of scrubbing at the edge, IPv6 on every instance and custom ISO support like everywhere else.
It is the first site we have opened primarily because of an electricity bill.
The context
European wholesale power went up several-fold over the last twelve months. Our Frankfurt and Amsterdam electricity costs roughly doubled inside a year, and unlike rack space or transit, power is the input we cannot hedge by being clever about procurement. A node draws what it draws.
Hosting margins are not wide enough to absorb that quietly. The options were to raise prices, to reduce what a plan contains, or to put some of the fleet somewhere the arithmetic is different. We picked the third.
Why Iceland
Electricity there comes from geothermal and hydro on long fixed-term contracts, which decouples the price from the market that has been misbehaving. Ambient air does the cooling for most of the year, so the overhead of running the room falls dramatically.
| Measure | Iceland | Our warmer European sites |
|---|---|---|
| Power overhead ratio | About 1.08 | About 1.40 |
| Cooling active | A few weeks a year | Continuously |
| Power price volatility | Fixed, multi-year | Repriced quarterly |
| Cost per node per month, all in | About 44% of a Frankfurt node | Baseline |
The overhead figure is the one that surprises people. Every watt of overhead is a watt you pay for twice, once to run the chip and once to remove the heat, and a site at 1.08 is spending eight percent on the room where a warmer site spends forty.
Latency, honestly
Reykjavík sits 39 ms from Amsterdam and 68 ms from New York on our own probes, which makes it closer to North America than most of continental Europe. That second figure is the useful one and it catches people off guard.
Going the other way is worse. Singapore is 218 ms and São Paulo 231 ms. Anything serving Asia or South America from Iceland is a bad idea and we would rather say so on the location page than let the map imply otherwise.
Why no EPYC and no GPU
Transit. Everything leaving the island travels on submarine capacity, and per gigabit that is expensive in a way that does not improve much with volume. A hundred-gigabit uplink is what the economics support today, against 400 Gbit/s in Amsterdam.
The server line and the GPU line are both bandwidth-hungry in ways that would make the site unprofitable or make us price it strangely. Ryzen instances and storage nodes are a better fit: storage in particular tends to be written once and read occasionally, which is exactly the traffic shape a costly uplink can carry.
We would like to be wrong about this. If demand justifies a second uplink, the server line follows.
Who this suits
- Backup and archive targets, where the electricity bill genuinely is the workload.
- Batch processing that nobody is watching in real time.
- Anyone who wants a European site outside the usual three or four cities, for reasons they do not have to explain to us.
- Cost-sensitive Ryzen workloads that mostly talk to Europe and North America.
Who it does not
Latency-sensitive front ends for European users. Frankfurt, Amsterdam and Paris are all under ten milliseconds from the core of the continent, and 39 ms is a real difference for anything chatty. Buying Iceland for a European web front end is buying a power saving you will pay for in round trips.
Details, latency figures from all four reference hubs, and current stock are on the location index.