JLL’s figures describe a change in the geography of planned capacity. They do not mean existing European datacentres have physically moved 129 kilometres away from their users.
Compare the right cohorts
JLL's August 6 announcement contrasts hyperscale campuses planned for 2026-2028, averaging 175 km from hub cities, with a 46 km average for the delivered 2022-2025 cohort. Greenfield projects rise from 8% to 39%. The future group is a pipeline, not a count of completed, powered sites.
The underlying mid-year report, dated August 3, also describes continued growth in the established Frankfurt, London, Amsterdam, Paris and Dublin markets. Expanding elsewhere and expanding existing hubs can happen simultaneously. The earlier article's suggestion of a wholesale departure overstated the finding.
Distance is an input, not a latency measurement
An original calculation shows the scale of propagation alone. Assume a signal speed of 200,000 km/s and an ideal direct route. Round-trip propagation is twice distance divided by speed: 46 km gives 0.46 ms and 175 km gives 1.75 ms, a difference of 1.29 ms.
This deliberately simplified model is not a measured connection between JLL sites. A distance from a hub city is not the actual fibre route, and routing, switching, queueing and endpoints add delays. There is no basis here for a universal multiplier that predicts every operator's RTT.
If an application makes 50 strictly sequential remote round trips, the hypothetical 1.29 ms difference adds 64.5 ms. Parallel or cached requests do not necessarily pay that delay 50 times. That is why request structure can matter as much as the advertised distance.
Place tightly coupled work together
“AI training” does not mean all network latency is irrelevant. Distance to end users is different from latency and bandwidth between workers exchanging training data. A campus far from a city can still host a closely connected training cluster; splitting that cluster across distant locations is a different design decision.
For site evaluation, separate power availability and connection date from network requirements. Measure the actual routes to data sources, dependent services and recovery locations. Keep synchronous dependencies close where the measured delay matters, and assess bulk ingestion and checkpoints separately. The 175 km statistic identifies a market trend; it cannot certify an application's deployment topology.
Distinguish planned and delivered cohorts; correct report date, avoid claimed wholesale relocation, explain hypothetical propagation calculation and intra-cluster training sensitivity.