Europe’s new AI data centres are being built 175km from anywhere
Facilities planned for 2026 to 2028 average roughly 175km from major hubs, against 46km historically. Latency lost the argument to electricity.
New AI data centre projects planned across Europe for construction between 2026 and 2028 sit an average of roughly 175 kilometres from major population hubs. The historical average was about 46 kilometres. Northern Sweden, rural Spain and Portugal are absorbing the investment.
A near-four-fold increase in distance-from-demand is not a preference shift. It is a constraint being obeyed.
What changed: the workload stopped caring about latency
Conventional cloud regions hug cities because interactive applications are latency-sensitive and light travels at a fixed speed. Every 100km of fibre adds roughly a millisecond each way, and for a web transaction doing several sequential round trips, that compounds into something users feel.
Large-scale AI training does not have that property. A training run is a batch job measured in weeks. Its internal communication is intense but local — between accelerators inside the building, over interconnect where the relevant distances are metres. Its external traffic is checkpoint and dataset movement, which is throughput-bound and entirely indifferent to a few extra milliseconds.
Remove the latency constraint and the siting calculation collapses to one variable: where can you get hundreds of megawatts, soon, at a price that survives a ten-year model?
Which is a question about interconnection queues, not generation
This is the part usually reported badly. The scarcity in European power is not primarily a shortage of electricity — it is a shortage of permission to connect. Grid interconnection queues near major metros run for years, because urban transmission is already congested and reinforcing it means building lines through places where people object to lines being built.
Northern Sweden is the archetypal answer to that problem, and not because of the cold, despite what the cooling-cost story usually claims:
- Large hydroelectric generation sited far from the load centres that consume it, so local capacity is genuinely spare.
- Existing high-voltage transmission built to export that power south, which means substations and rights-of-way already exist.
- Low population density, which shortens permitting rather than merely reducing complaints.
- Cheap, low-carbon, and price-stable — the combination that lets you sign a fifteen-year power purchase agreement without hedging yourself into a loss.
Rural Spain and Portugal offer the solar-and-wind version of the same argument: abundant generation, weak local demand, and land where a 200MW campus is not a planning fight.
The trade-offs this creates
Moving compute 175km from population centres is not free, it is just cheaper than the alternative.
| Gained | Given up |
|---|---|
| Interconnection in quarters, not years | Inference serving needs separate metro-edge capacity |
| Power at a price that supports long PPAs | Long, expensive private fibre to reach backbone |
| Permitting through low-objection jurisdictions | Thin local labour markets for skilled operations staff |
| Low-carbon supply that satisfies reporting | Regional political exposure — a single large load in a small grid |
That last row deserves attention. A 300MW facility inside a small regional grid is a politically visible object. When retail electricity prices rise for any reason, a foreign-owned data centre consuming a measurable percentage of regional supply becomes the explanation, whether or not it is the cause. Ireland has already lived this, and it ended in connection moratoria around Dublin. The economics that make northern Sweden attractive today are the economics that generate a local backlash tomorrow.
The structural consequence
The industry is bifurcating its footprint: enormous, remote, power-sited campuses for training, and a separate, smaller, metro-adjacent layer for inference — because serving tokens to users is latency-sensitive in the way training is not.
That split is the physical-world echo of the same divide showing up in accelerator design and in who is buying routing layers. Training and inference are becoming genuinely different businesses, and they are now being built in different places.