Some UK datacentre developers are ordering fibre for sites where the ground has barely been broken. Capacity used to follow tenants in steady increments. Now it’s being bought before anyone has plugged anything in. “We’re talking tens of fibre pairs in the first instance,” explained Lee Myall, CEO of wholesale fibre operator Neos Networks, in an interview with TelcoForge. “And that’s a change in behaviour.”
Clearly, somebody is expecting a lot of traffic. Where that traffic will actually come from is a different matter – and it’s the question the telecoms industry is now trying to answer, with very mixed results.
For Myall, the “X factor” is simple to describe but almost impossible to answer: how close to the user is inference going to get?
Regional fibre, edge GPUs, idle mobile capacity, and new long-haul routes are all being pitched as telecoms’ contribution to the AI boom. In such an embryonic market it’s impossible to tell which approach, if any, is the most viable yet. There are some indicators though.
Granulated
Myall’s starting assumption is that regional networks are about to carry a very different kind of load.
“Regional connectivity is going to have more lifting to do than get your Netflix stream on your TV, or your SME in a business park,” he said. “It’s going to be doing a lot more than that.”
What that load looks like, though, depends on who you ask. Neos’s customers are placing a variety of different bets. “There are people who are trying to mop up little pockets of office space at a very granular level, because they think that is what inference deployment would look like,” Myall said. “With others, it will be urban-population levels of X people and above.”
Neos has tried to put numbers on the demand side. In research it commissioned from Censuswide, 47% of 250 large UK enterprises said their AI workloads need to sit in the same UK region as their users, and a further 22% said the same metro area. Only 22% said anywhere in the UK would do.
Having said that, this is only theoryware right now. A fibre round trip across the whole UK is measured in milliseconds. Having inference localised in the “same region” may say as much about governance and control as it does about any physical constraints.
Myall admits that the picture is still in flux. “We’re catching people right now on this,” he said, adding that a repeat of the survey in 12 months would likely produce different answers “because the landscape would have changed”.
Spilling the Tea
As a wholesaler, Neos sits a step back from the people actually deploying AI. “We’re always trying to read the tea leaves,” Myall said.
It looks like the leaves in Myall’s cup are fibre-shaped. Ethernet customers are moving from 1G to 10G and from 10G to 100G, and 400G deployments are “starting to ramp up” from a low base.
Myall is careful not to overclaim. “I can’t, hand on heart, say I’m going to wholly attribute that to AI,” he said. “But what they don’t want to do is handicap themselves through a lack of connectivity capacity.”
Neoclouds are a different kind of beast: “They go for vast amounts of capacity from day one,” Myall said. “Traditionally, datacentres have built more gradually.”
This isn’t just a UK phenomenon. In the US, Zayo has said its long-haul orders used to average 8 to 12 fibres, but now are ordering an order of magnitude more.
There’s a subtler shift, too. The cloud migration of the last decade mostly pulled data in one direction. “AI is a much more symmetric flow of data, it would appear, than before,” Myall pointed out. That means that upstream capacity will become a much more significant product in its own right, with real implications for how wholesale networks are designed and priced.
‘Edging Bets
So where does the telecoms industry fit? Right now the consensus seems to be “Answer unclear, ask again later”.
Neos’s own research illustrates the point. Asked where AI would drive the most traffic growth on their networks, service providers split almost evenly: edge and access 30%, core and backbone 27%, datacentre interconnect 25%, metro 18%. Neos reads that as demand likely to hit the network end to end, but it’s equally likely to mean everyone’s guessing.
The edge is especially contested. Nvidia is promoting “AI grids” of distributed GPUs inside operator networks, and several carriers, AT&T among them, are trialling them. Yet AT&T’s own CTO has questioned extending compute to the far edge “just to save another millisecond or two milliseconds of latency”. ABI Research’s Dimitris Mavrakis expects inference to start in fewer than ten core locations per country, spreading outward only where latency genuinely matters, for example in autonomous vehicles and other physical AI systems.
There is an alternative argument for supporting Ai in the network, though – utilisation. Mobile networks are dimensioned to support peak traffic and can often run well below capacity outside of peak hours. Some operators, SoftBank among them, are exploring whether that headroom could be sold as inference capacity. This is part of a separate conversation we’ll return to later this week.
Are these approaches competing? Not necessarily. Myall’s granular office-space buyers, ABI’s handful of national cores and an operator sweating its own sites could all coexist. What they share is a single bottleneck: power.
Power Moves
Across Europe, grid capacity is increasingly the reason to decide where compute power is being built.
In the Netherlands, a court has decided the grid near Schiphol is full, and datacentre developers have been told they might not be able to operate properly until around 2035. Ireland’s energy regulator now requires new datacentres to build their own generation capabilities and match 80% of annual demand with new Irish renewables.
None of which makes connectivity irrelevant. It turns out fibre is the cheaper thing to build out, though. Zayo has argued that running fibre to a power supply costs four to ten times less than running energy transmission lines to a site; one reason it’s laying thousands of miles of long-haul routes into US corridors picked for their power supply.
If that cost difference holds true elsewhere, it’s a force for AI regionalisation in its own right. Compute will follow where power generation is, and fibre will follow compute. Which regions gain compute depends on the energy source, therefore: gas can be piped almost anywhere, but large-scale wind, hydro and solar production depend on local factors and are frequently remote – as is nuclear, though for different reasons.
Myall sees exactly this in the UK. “We’re predicting that the north-east of Scotland is becoming more and more interesting,” he said. “More energy generation will come on tap up there.”
Smaller sites change the calculation, to a degree. A swarm of smaller inference sites might draw as much in aggregate as one hyperscale campus. Spread across the grid, though, that load could sit within local distribution capacity and alongside local renewables, rather than queuing for a transmission connection.
Japan’s and communications ministries jointly run a “watt-bit collaboration” programme, planning power and telecoms infrastructure together around datacentres. In Kyushu, the regional power utility is working with 1Finity and fibre operator QTnet to link renewable-powered datacentres across the region, so that processing can shift to wherever generation is strongest. In that model, the network becomes part of how the grid is balanced.
Table Stakes
Japan is the exception, though. Most countries plan and regulate energy and telecoms separately.
For example, in the UK a site must show it can reach 500MW of power by 2030 to qualify as an “AI Growth Zone”. Connectivity is merely “viewed favourably”, evidenced by a letter from a network operator. The energy department, Ofgem and NESO are all working on grid reform, and Growth Zones will get priority access to grid capacity. There’s no significant role for telecoms.
Myall has been making this case loudly. When the Lanarkshire Growth Zone was announced, he argued that “connectivity has to be baked into planning from day one”. That doesn’t sound like a minor point, but disjunction between regulators and silos means it has needed to be made (and still does).
Brussels is moving further. The EU’s proposed Cloud and AI Development Act would have member states designate datacentre acceleration zones with grid capacity and connectivity both accounted for.
Does telecoms need a seat at the table in developing the energy grid, them? Certainly any sizeable sites for data processing or inference will need to have visibility at least into government plans, if only to ensure they aren’t caught out like in Amsterdam.
Meanwhile, if compute starts following renewable power generation in real time, as Japan is testing, the argument gets much stronger. At that point the network is part of the energy system, so it makes little sense not to coordinate.
Running Hot
Does all of this talk about huge growth in fibre and aggressive roll-outs of AI infrastructure seem a bit dotcom-ish? Myall was in the industry then, and can see similarities without being too concerned.
AI companies “Could fail commercially. That would be one thing, of course,” he conceded. “But I think that we’re at this point now where there’s a sort of an unstoppable demand and a different way of doing things that we’ve all woken up to.”
He points to the UK’s altnets, which overbuilt their local fibre infrastructure and are now consolidating. “You don’t need four different fibres outside your house,” he said. Yet the traffic flowing through people’s front doors has only gone one way. If an AI bubble burst it could produce commercial casualties without denting long-term demand.
The lesson, for Myall, is about discipline. He quotes DigitalBridge CEO Marc Ganzi, who pointed out that much of the fibre laid around 2000 is still in use today, but argues that AI-scale demand needs new infrastructure laid where the bet is safe. For Neos, that means 432-fibre cable along Network Rail corridors linking London with Cardiff, Manchester and Newcastle; going “big and deep where you’re placing a solid, solid bet that’s going to basically underpin the heavy lifting of AI”.
In the same breath, he describes his philosophy as “build as little as possible and have it glowing red hot with operational excellence on the top”. Is that a contradiction? Not really. Build heavily, but only where demand is certain.
That points to a particular role for wholesalers. Myall has no interest in moving up the stack. “We need to be very, very vertically integrable, for other people to come and make their money doing what they do really well,” he said.
In practice, that means making self-serve Ethernet, zero-touch APIs, optical services available through a simple customer portal and enough partner reach that a customer can land and expand rapidly anywhere in the country. Neos’s UK-only footprint, once a limitation, is becoming an asset as sovereignty climbs the agenda.
It’s not a glamorous position, but Myall sees it as a positive.
“We have to take glory in being uncelebrated and hidden on a daily basis.”
