Someone on a commuter train presses send on a 40-second clip and watches a progress bar move without finishing. That is the failure mode Andrea Donà was pointing at when he compared incoming AI traffic to the signal loss people get in a stadium, and said the difference is that it would be permanent [2]. Stadium loss rarely produces a clean error code. It produces retries and half-uploaded assets, and it shows up in support threads that say "it works at home".
Here's the arithmetic behind the constraint. About 96% of VodafoneThree's upgrades sit on structures that already exist [7], and more than half of those still need planning permission [8], so at least 48% of the whole upgrade programme is waiting on a consent decision rather than an engineer [16]. The same programme cuts the mast count by 30% [6], which means fewer sites each carrying more of the load, and each delayed permit mattering more than it would have in a denser grid.
What's being pitched is procedure. What's being done is £11bn of committed capital [5]. Donà calls planning reform "the lowest possible cost growth policy" that "doesn't require any government money" [4]. The EU wrote that remedy down. Since November, authorities there have four months to decide a permit for a very high capacity network, and the permit is deemed granted if they miss it [9]. The UK is covered by none of that [10], so a British operator's application has no statutory ceiling a product team can plan a launch date against [17].
Coverage checkers show 5G and make median throughput look like the design target. Users don't experience medians, though - they act at the moment they want to, in the place they happen to be, which is disproportionately the slowest decile of sessions. Session counts will not show you this, and neither will time in app. The numbers that move are completion rate on the network-dependent action and time to first useful output, measured on your slowest tenth of sessions rather than your average one.
The forcing function has two axes. First, does the feature need the uplink at a moment the user chose (a deliberate upload) or at a moment they did not (background sync, live capture, streaming model calls). Second, does it degrade into something useful or does it stall. Chosen and degrading is safe to ship. Unchosen and stalling is the cell that generates Friday questions, and it needs a local queue plus an honest "we will finish this later" state before launch, not after the first bad week.
How long this stays true comes down to two details. Operators note that mobile networks count as critical national infrastructure without the business rates and energy reliefs data centres receive [14]. And a Spanish startup has raised €150M to beam 5G to ordinary handsets from orbit, which needs no local consent at all [15]. When the workaround is funded in space, the ground-level queue is not clearing this year.
