Edge inference and the return of distance.

Training centralises. Inference distributes.

The first phase of the AI buildout was about training, and training has simple geography: one enormous campus, wherever power is cheapest, as far from anyone as the fibre allows. That phase is not over, and the gigawatt campuses it demands are a core part of what we build. But the centre of gravity is shifting. Every model trained is a model that then has to run, billions of times a day, for people and machines who will not wait.

Inference is different. It wants to be near the user: for latency, for data residency, for cost. As AI moves out of the laboratory and into daily life, into agents, voice, video, robotics and industrial control, the share of compute spent on inference rises year on year. And inference does not want one campus in the desert. It wants a hundred sites at the edge of a hundred cities.

Distance is back.

For twenty years the cloud taught us that geography did not matter. A server could be anywhere; the network would hide the distance. Inference has undone that. A conversational agent that responds in three hundred milliseconds is usable; the same agent at eight hundred is not. A factory floor running vision models cannot round-trip to another continent. A hospital cannot send patient data across a border to have a sentence completed. Milliseconds matter again, and so does the law of the land the server sits in.

“For twenty years the cloud told us geography was dead. Inference has just put it back on the map, and this time it has a latency budget.”Jamie MacDonald-Murray, Chairman & CEO, Eppur

The consequence is that power near people has become one of the scarcest commodities in the industry. Cities have the demand and none of the headroom. Whoever holds grid positions on the edge of a metropolitan area, on land that can be consented, is holding something that will be very hard to replicate.

The quiet winners.

Edge sites do not get the press releases. A 20 MW facility on the ring road of a regional city is not a gigawatt campus, and nobody writes about it. But a network of them, built on existing grid positions, close to industrial heat users, served from the same platform and financed against the same offtake relationships, is one of the most durable businesses in infrastructure. It sits where the demand is, it is hard to displace, and it compounds.

That is the logic behind our national edge programme in the United Kingdom, and it is a pattern we intend to repeat in every market we anchor. Where our gigawatt campuses sit where power is cheapest, our edge networks sit where intelligence is used.

Two products. One platform.

Both require the same thing at origination: power, secured early, from people who know the grid. A hyperscale campus in Brazil and an edge site in the English Midlands are different buildings with the same first question. Where is the substation, who runs it, and how soon can we connect? Answering that question first, before land, before design, before the customer, is what our platform is for.

“The gigawatt campus gets the press release. The edge network gets the margin.”Jamie MacDonald-Murray, Chairman & CEO, Eppur

Distance is back. The developers who understand that it never really went away will build the networks the AI era runs on.