LONDON, 20 AUG 2026 — Hyperscale data centre campuses planned across Europe, the Middle East and Africa for 2026 to 2028 sit an average of 175 kilometres from major hub cities, against 46 kilometres for projects delivered between 2022 and 2025, according to research from JLL.
The industry has spent two decades building as close to the network as it could get. It has now stopped, and electricity is the reason.
The finding
JLL attributes the shift to developers pursuing the electricity that AI infrastructure requires, while grid constraints, planning restrictions and a shortage of suitable land make further building around Europe's established data centre hubs progressively harder.
The same research puts 2026 capital expenditure by the four largest hyperscale cloud providers at about US$725 billion, up 77 per cent from US$410 billion in 2025, with much of it directed at AI compute and the buildings to hold it.
Nearly four times the distance is a change of kind, not degree
A data centre 46 kilometres from a city is in its metropolitan area. It draws on that city's grid, its fibre, its labour market and its planning regime, and a technician can reach it in under an hour.
At 175 kilometres, none of that holds. The site falls under a different planning authority, often a different grid constraint zone, and is outside the commuting radius of its engineering staff. Fibre must be built, not bought. Everything about the operating model — staffing, maintenance response, physical security, spare parts — changes when the nearest city is two hours away.
Developers have accepted all this in exchange for power. That trade-off is a precise measure of how binding the electricity constraint has become. Nobody moves a facility 130 kilometres further out to save on land if the electricity is available where the customers are.
What the distance costs, and what it does not
The obvious objection is latency, and it is mostly wrong. A hundred and thirty extra kilometres of fibre adds well under a millisecond of round-trip propagation, which is immaterial for training runs and irrelevant beside the tens of milliseconds that routing, queuing and the last mile contribute.
The siting shift is affordable precisely because AI training is latency-indifferent. That work can happen wherever there is enough power, and so it does.
The workloads that cannot follow are the interactive ones. Inference serving a user typing into a chat interface, real-time video, financial matching engines and anything with a regulatory requirement to be in-country still want to be close, and those remain in the expensive constrained hubs. The result is a two-tier geography. Training and batch work migrate outward to cheap power, while interactive serving stays put and pays for the privilege.
The same pressure, three continents, one week
We reported yesterday on American regulators giving all six grid operators sixty days to justify their large-load connection rules, on Amazon raising a Louisiana commitment to US$18 billion while funding its own water infrastructure, and on a US$1.37 billion green financing for a Johor campus.
These are four views of the same constraint: a regulator rewriting connection rules; an operator paying for its own utility upgrades; a developer financing capacity in a power-rationed corridor; and an entire region's projects moving physically away from demand. None of these is a story about computing. They are all stories about electricity.
We also wrote about ASEAN hitting a wall of power and water, with Johor turning projects away, which is the same wall from the other side.
Why Southeast Asia cannot simply copy this
The European answer is to go further out, and the region has less room to do that than the map suggests.
Singapore has no hinterland at all — the entire country is smaller than the distance JLL is describing — which is why capacity has moved to Johor and Batam rather than to the countryside. Malaysia and Indonesia have land, and the constraint there is not distance but grid: moving 175 kilometres from Kuala Lumpur or Jakarta does not reliably reach a stronger connection point, because transmission was built to serve population rather than to distribute industrial load.
Europe can do this because its grid has strong high-voltage transmission lines running to sparsely populated areas with power generation. Much of Southeast Asia lacks that geometry, so distance alone buys less. The regional equivalents are new generation built alongside the load, transmission investment on a decade-long timescale, or accepting the constraint and rationing.
The exception worth watching is Vietnam, where industrial parks along the coast sit near both wind resource and existing transmission corridors. That is closer to the European pattern than anywhere else in the region.
The workforce consequence is often skipped in these siting discussions. A campus two hours from a city has to recruit and retain technicians who will live near it, or pay for rotation and accommodation, and neither is cheap in a labour market already short of data centre engineers. Operators have absorbed that cost silently so far because the alternative is not building at all, and it is one of the reasons the delivered average distance may end up shorter than the announced one.
What we could not establish
Whether the same shift is measurable in Asia-Pacific. The finding is EMEA-specific, and the comparable regional figure would say whether this is a global reordering or a European response to a particularly constrained grid.
JLL's research does not specify the sample of campuses behind the averages, or whether it includes uncommitted projects — which are often more speculative the further out they sit. We also do not know how many of these distant sites have secured grid connections versus merely applied for them, what the shift does to costs, if the pattern differs for hyperscaler-owned sites, or how the US$725 billion capex figure is defined.
What to watch
Watch whether the distant projects actually get built. A campus 175 kilometres out is a bet that a grid connection arrives on schedule, and connection queues are precisely what pushed developers there in the first place. Announced average distance and delivered average distance may turn out to be different numbers.
Then watch for the equivalent measurement in this region. If somebody publishes it and the distance has not moved, that is evidence the constraint here binds differently — not that it binds less.
Finally, watch the split between training and serving capacity. If the two-tier geography holds, the interesting scarcity in five years is not remote megawatts but close-in capacity for interactive workloads, and that is the market where a small, well-connected country with a constrained grid has to decide what it wants to host.
View author profile → · Editorial policy
