For the past two years the telecom industry has been wrestling with an awkward question.

Did the UK build too much fibre?

Too many networks. Too much capital. Too many operators digging up the same streets.

The mood has shifted noticeably from enthusiasm to scepticism. In some corners of the industry the fibre boom is already being described as an overshoot.

But the arrival of the AI economy raises a more interesting – and little-discussed – possibility.

What if the question isn’t whether the UK built too much fibre?

What if the real story is that we may only just have built enough?

To explore that possibility, it helps to ask three simple questions.

First, what happens when the network becomes the bottleneck?

Most of the current debate about AI infrastructure focuses on data centres.

GPUs. Electricity consumption. Cooling systems. The vast buildings required to house them.

Those things matter enormously.

But AI does not just consume computing power.

It consumes data – vast quantities of it – moving continuously across networks.

And data does not move by magic. It moves across networks.

If AI adoption accelerates the way many expect, the real constraint in the AI economy may not be the computers.

It may be the network connecting them.

Training models consume enormous datasets. Inference systems generate continuous streams of requests. Sensors and edge devices constantly feed information back into learning systems.

Machines increasingly talk to other machines.

Unlike traditional internet traffic – which is largely human-driven – AI traffic is continuous, bidirectional and latency-sensitive.

Building ever larger AI data centres without upgrading the networks feeding them is a little like building Formula One engines and asking them to run on country lanes.

The engine isn’t the problem.

The road is.

Second, were broadband networks designed for the traffic AI will create?

Most broadband networks were built around a simple assumption: people downloading things.

Streaming video. Browsing websites. Pulling information from the cloud.

Traffic patterns were overwhelmingly download-heavy, with predictable evening peaks when households pressed play on Netflix.

AI changes that model.

Machines increasingly exchange data with other machines. Training models require enormous datasets. Inference systems generate continuous flows of requests. Edge devices feed telemetry back into AI systems in real time.

The result is traffic that is more symmetrical, more continuous and far more sensitive to latency than the patterns that shaped the design of most broadband networks.

At this point the limitations of some legacy technologies become clear.

Copper is already approaching the end of its useful life.

Cable networks – originally designed for broadcast television – are running into upstream capacity limits that become more visible every year.

Wireless networks remain essential but ultimately depend on fixed infrastructure underneath them.

Which leaves one technology capable of scaling with demand.

Fibre.

Third, if fibre is so critical, have we actually built the right networks?

Over the past decade, the UK has experienced one of the most aggressive fibre build-outs anywhere in the world.

Established operators expanded their networks. A new generation of alternative providers – the altnets – accelerated the rollout even further.

The result has been enormous investment and rapidly expanding coverage.

Recently, however, the narrative has begun to shift.

As capital markets tightened and competition intensified, some commentators began asking whether the country had built too much fibre infrastructure.

That judgement may prove premature.

Because the AI economy is only just beginning.

AI has an extraordinary appetite for data. Moving that data requires the one infrastructure layer that scales almost indefinitely.

Fibre.

That said, fibre alone is not the whole story.

Many networks were designed primarily around the economics of consumer broadband – streaming video and household connectivity.

AI workloads may place rather different demands on networks: greater symmetry between upload and download, lower latency and far more machine-to-machine communication.

In other words, the first generation of fibre networks may only be the foundation layer of the next phase of digital infrastructure.

Finally, what might this mean for the UK?

The internet itself is surprisingly young.

For most practical purposes it is barely 25 years old, and much of the infrastructure we rely on today was designed for a very different digital economy.

AI will test those assumptions.

Today’s infrastructure debate revolves around chips, power and data centres.

Yet none of those systems work without the network connecting them together.

The engines of the AI economy may sit inside data centres.

But the roads those engines run on are networks.

Fibre networks.

If that proves true, the UK’s intense fibre build-out may look rather different in a few years’ time.

What some people currently describe as overbuild may turn out to be something else entirely.

The real question may not be whether the UK built too much fibre.

It may be how long it will take before we discover the network itself has become the next AI bottleneck.

Or put another way:

How long will it be before we realise we didn’t build enough capacity for the AI economy after all?

About the author

David Hilliard is founder of Mentor, specialists in strategic program execution.

You can call him on 0118 359 2444 or email david.hilliard@mentoreurope.com.