Data centre siting was a latency problem for thirty years, so operators built near people. AI training has no latency requirement, so European capacity is migrating toward wherever the grid can supply it - splitting the market in two rather than moving it wholesale.
Hyperscale data centre campuses planned across Europe, the Middle East and Africa for delivery between 2026 and 2028 sit an average of 175 kilometres from a major hub city. For projects actually delivered between 2022 and 2025, the figure was 46 kilometres. That is not drift and it is not a rounding artefact. It is nearly four times further out, and the change happened inside a single planning cycle.
The figures come from JLL's EMEA data centre mid-year 2026 report. It is worth stating at the top that JLL is a commercial real estate brokerage publishing research about a market in which it earns fees, so this is its dataset and its methodology rather than a neutral census. It is also the most granular public read available on where European compute is physically going, and the direction it describes is consistent with the siting decisions visible in the public record.
The framing this will mostly receive is that Europe's grid cannot cope. That is true, and it has been true for roughly three years, which is precisely why it explains nothing about what changed this year. Grid constraint is the background condition. The 46-to-175 kilometre jump is a decision taken repeatedly by different operators in different countries, and decisions of that consistency are usually responding to a change in what the binding constraint is, not to its severity.
The scale of the sites involved makes the point harder to dismiss as noise. JLL counts nine proposed European data centres at a gigawatt or above. Exactly one of them is near a major city, outside Paris. At that size the facility is no longer a real estate project that happens to consume electricity; it is an industrial load that happens to contain servers, and industrial loads have always been sited at the power source.
For about thirty years, data centre siting was fundamentally a latency problem, and latency meant distance to users. If a transaction had to complete inside a few milliseconds, the machine doing it had to be near the person waiting. That single requirement produced Frankfurt, London, Amsterdam, Paris and Dublin as a coherent cluster, and it produced the interconnection economics that kept them there once they existed.
Large-scale AI training does not have that requirement. A training run of 100 megawatts and above is a batch job. It does not care whether the electrons travelled 46 kilometres or 800, because nobody is waiting on the other end of a request. Remove the latency constraint and the next-most-binding one becomes the constraint, and in Europe that is unambiguously power availability and the queue to connect to it.
Inference is not training. Serving a model to users is latency-bound in exactly the way training is not, which is why the old hubs are still growing rather than emptying. A reader who takes away that data centres are leaving cities has learned the wrong thing.
The FLAP-D markets confirm this. Their combined live capacity has grown from roughly 1.8 gigawatts in 2019 to about 3.8 gigawatts by the first half of 2026, with a further 1.4 gigawatts under construction and 2 gigawatts planned. In the first half of this year Paris led take-up at 72.5 megawatts, ahead of London at 49, Frankfurt at 45, Amsterdam at 16.3 and Dublin at 11.4. This is not a market in retreat. It is a market that stopped being the only market.
What is actually happening is a bifurcation. Latency-bound workloads stay where the users are and keep bidding up the hubs. Capacity-bound workloads leave for rural Spain, the Nordics and wherever else a grid connection can be had this decade. Over half of Europe's expected AI capacity growth now sits in the Nordics and tier-two markets. Describing that as one industry relocating misses that it is one industry separating into two with different physics.
The second number in the JLL report is arguably more consequential than the first. Greenfield sites accounted for 8% of delivered European projects. They account for 39% of the 2026 to 2028 pipeline.
Greenfield means no existing substation, no existing fibre route, no existing planning precedent and frequently no existing road capable of carrying a transformer. Every one of those is a multi-year lead time, and each one is a place where a project can die. Developers avoided greenfield for decades for exactly these reasons. That 39% figure means they are now choosing it, deliberately, at scale.
“Operators have concluded that building a site from nothing is faster than waiting for a grid connection in an established hub. That is a revealed preference, and it is a judgement on the queue rather than on the field.”
Distance does not eliminate cost; it relocates it onto the transmission network. A gigawatt of load in a rural region requires transmission capacity that does not currently exist, and transmission is built by network operators and paid for, in most European markets, through charges socialised across all users. The developer captures cheap land and cheap power. The line to reach it is a shared bill with a ten-year construction horizon.
The local politics are harder than the siting maps suggest. A remote campus brings a very large electrical load, significant water demand for cooling, and a permanent workforce that is small relative to the capital deployed. Communities that were promised industrial regeneration and received a fenced building with forty staff have generally not been persuaded by the capex figure, and several European planning fights now turn on precisely this gap.
The counter-case deserves to be made properly rather than waved through. Remote siting is not obviously worse. Land is cheaper, power is cheaper, and in the Nordic case it is also substantially cleaner, which means the same training run emits less carbon than it would in Frankfurt. Building load where generation already exists is what an efficient system would do. The honest position is that neither the grid-failure framing nor the clever-arbitrage framing survives contact with both halves of the ledger.
JLL expects the four largest hyperscalers to spend around $725 billion in capital expenditure in 2026, up 77% on $410 billion in 2025. That is a brokerage's expectation rather than company guidance, and it should be read as such, but the order of magnitude is the relevant part: this pipeline is not funding-constrained in any way that would slow it down before the transmission question is settled.
The number worth tracking over the next eighteen months is not distance but delivery. A 39% greenfield pipeline is a bet that empty land can be turned into connected capacity faster than an existing queue can be cleared. If those projects land on schedule, the map of European compute is permanently redrawn around generation. If they slip the way greenfield projects historically slip, the hubs will have absorbed the demand anyway, and the 175-kilometre figure will read as a planning-stage aspiration rather than an outcome.
For almost everything you use day to day, no. The facilities moving furthest out are large AI training campuses, which run batch workloads with no user waiting on a response. The inference and cloud workloads that actually serve your requests remain latency-sensitive and are still being built in and around the established hubs, which is why FLAP-D capacity is growing rather than shrinking.
DrafterDaily covers the physical and financial mechanics of AI infrastructure - siting, power, capex and the second-order effects most coverage skips.
Explore our tools coverageCVE-2026-0768 is an unauthenticated root RCE in Langflow. It was disclosed in January, the fix has shipped through seven releases, and attackers are hitting it in September — because low-code AI middleware became critical infrastructure without acquiring a patch owner.
Opposition to data centres is not a referendum on AI. It is a cost-allocation dispute — and PJM's capacity auction is the mechanism that turned an abstract argument into a number on 67 million households' bills.
OpenAI's first published benchmarks for its custom inference chip are real, from a public benchmark, and normalised on nameplate TDP rather than measured power. OpenAI disclosed that itself, in a sentence nobody is quoting.