UK risks falling behind in AI race unless mobile networks upgrade faster, say telecoms executives
Which? using data from Opensignal found the UK trailing every other G7 country and all 27 EU member states for mobile network performance. The consumer group’s analysis ranks the UK 57th globally for overall network performance, 70th for download speeds and 55th for the reliable quality needed for video calls, streaming and gaming. Those figures matter because the next phase of AI won’t live solely in datacentres, it will increasingly rely on low-latency, high-capacity mobile links for real-time agents, AR experiences and background AI services.
What operators are saying
Senior network executives have been blunt. Andrea Donà, described in reporting as networks director at VodafoneThree, used a sports-stadium image to warn about future demand:
“If with the technology and services we use today [networks] are sub‑optimal, imagine when you add mass AI usage and other associated activities, ”
Andrea Donà
Donà and other telecoms leaders quoted in industry coverage go further, calling the current state embarrassing and urging planning and regulatory reform:
“I would go further. I am embarrassed today at the level of quality of our network[s]. The UK is worthy of what EE, O2 and VodafoneThree can offer, but we need reform.”
Andrea Donà
According to the operator claims cited in reporting, VodafoneThree has committed £11bn to upgrade masts and migrate toward 5G stand‑alone technology, and plans to reduce its nationwide mast count by around 30% as part of densification and modernization. The company says roughly 96% of the upgrades it needs are on existing sites, and that more than half of those will require planning permission.
Executives argue mobile networks are treated as critical national infrastructure in practice, but do not receive comparable reliefs, such as certain business‑rate or energy concessions, that some other infrastructure sectors have obtained. As Donà put it in the remarks reported:
“Our infrastructure is here to help the government and UK economy. Planning reform is the lowest possible cost growth policy, it doesn’t require any government money. It could be a great first 100 days‑win for the new prime minister.”
Andrea Donà
Why 5G stand‑alone matters
5G stand‑alone (5G SA) runs core network functions natively on 5G rather than relying on a 4G core. That reduces round-trip latency, supports greater capacity and enables network slicing, the ability to create virtual networks with dedicated performance characteristics. For distributed AI use cases, real-time inference for agents, AR assistants, or synchronized background model updates, lower latency and guaranteed slices can make the difference between a usable experience and one that collapses under load.
Rollout speed is constrained by practical bottlenecks. Many upgrades occur on existing sites, yet local planning permission and approvals for new small-cell deployments remain slow in numerous areas. Densification strategies that reduce large macro-mast counts typically increase the number of small cells and the demand for fiber backhaul, shifting the planning and energy burden rather than eliminating it.
What this means for business leaders
Network capacity is a strategic dependency for AI adoption. If your roadmap includes AI agents on mobile devices, AR sales tools, real-time personalization in the field, or background AI services pushing continuous updates, treat mobile networks as an infrastructure risk to manage, not a commodity you can assume will keep up.
Practical, measurable steps to reduce risk:
- Set concrete latency targets: measure p50 and p95 round‑trip latency in your target geographies. For many AR and real‑time agent experiences aim for p95 latency under 20-30 ms; if you can’t reach that over public mobile networks, plan edge inference or on‑device models.
- Define SLA metrics with carriers: contract for p95 latency, jitter, and guaranteed throughput, and include clear remedies or penalties if SLOs are missed. Ask carriers about network slicing options where available.
- Design a hybrid AI architecture: push heavy training to datacentres and run latency-sensitive inference nearer users, on edge servers or devices, to reduce reliance on long-haul mobile links.
- Run realistic peak tests: simulate concurrent sessions and aggregated AI traffic in the specific cities or sites where staff and customers are concentrated; include sustained loads, not just short spikes.
- Invest in local compute and caching: micro-data centres, on-prem edge servers and content caching reduce round trips and stabilize performance under load.
- Factor connectivity into vendor selection: when buying AI vendors or platforms, require evidence of performance in your target regions and contractual uptime/latency commitments.
Policy levers industry wants, and the trade‑offs
Executives favour faster planning approvals for upgrades on existing sites, temporary relaxations for small cell deployments, and parity in reliefs for mobile infrastructure relative to datacentres. These are presented as low-cost ways to speed rollout without large direct government spending.
Those proposals have trade-offs. Faster planning approvals can raise local concerns over visual impact, community consultation and electromagnetic exposure. Granting fiscal reliefs narrows government options for managing public finances. Densification also increases backhaul and energy needs, so the net environmental and fiscal impact depends on how deployments are designed and where the energy for backhaul and edge compute comes from.
Industry voices warn that one-off congestion events, the kind you see at Wembley, Silverstone or Twickenham, are a preview of sustained aggregation when billions of AI agents and always-on services become normal:
“What we see now only with a mass aggregation of people is a proxy, an indication of the things that will come. That [one‑off] experience will be constant because of the additional capacity [consumers] will need in an AI world.”
Andrea Donà
Other senior telecoms executives put the timeframe bluntly:
“Things are going to change massively with the next wave of devices, in two or three years when AI is integrated into everything, ”
another senior telecoms executive (unnamed)
“AI companies, and governments, don’t think about networks, but they will when they hold them back.”
another senior telecoms executive (unnamed)
Where to focus budget and attention now
CIOs and line‑of‑business leaders should reallocate a small portion of their AI deployment budgets into connectivity resilience and testing. Key line items to consider:
- Edge compute and micro‑DC placements in priority regions
- Carrier SLAs that include measurable latency and throughput SLOs and remedies
- Subscription to performance monitoring tools that report p50/p95 latency, jitter and packet loss across target geographies
- Funding for peak‑load field tests that mirror expected AI agent concurrency
Key takeaways, questions a curious leader would ask
-
Is the UK actually behind on mobile networks?
Which?, using Opensignal data, ranks the UK 57th globally for network performance, 70th for download speeds and 55th for reliable quality, placing it behind every other G7 country and all 27 EU states in that analysis.
-
Will full‑fibre broadband solve AI‑related network problems?
Full‑fibre rollouts by providers such as BT’s Openreach have improved fixed‑line capacity for homes and offices. But mobile networks introduce separate challenges, coverage, latency and the need for 5G stand‑alone features, that fibre to premises does not directly address for mobile and distributed AI use cases.
-
What are operators doing and how reliable are their figures?
Reporting attributes a £11bn upgrade commitment, a ~30% reduction in mast count and an assertion that ~96% of required upgrades are on existing sites to VodafoneThree. These are operator claims and should be treated as such when assessing rollout timelines and vendor commitments.
-
Why does planning permission matter so much?
Many necessary upgrades sit on existing sites but still trigger local planning processes. Delays in approvals slow deployments for small cells and backhaul, increasing time to deliver the low‑latency, high‑capacity networks AI services will need.
-
What should businesses do right now?
Assume mobile capacity could constrain AI adoption: set measurable latency targets (track p50/p95), negotiate SLAs with carriers (latency, jitter, throughput, penalties), design hybrid on‑device/edge architectures, and perform realistic peak testing in target geographies.
The infrastructure race for AI is not only about GPUs, water and datacentres, it also runs along masts, fibre ducts and planning desks. Companies that bake network realism into their AI strategies now, by testing, contracting, and investing in edge resilience, will avoid being tripped up by a bottleneck that waiting for policy alone may not fix.