Picture an electrician from Cork, standing in a half-finished data hall outside Helsinki, torquing busbar connections on a cold Tuesday morning. In about eighteen months, the hall around them will be running AI models that write code, draft contracts and read X-rays.
The question everyone asks about AI is which jobs will it take?
The question almost nobody asks is who builds it?
Every dollar ends up as copper
The largest technology companies in the world will spend more than $600bn on infrastructure this year, up 36% on last year. About three-quarters of that is for AI.
It's easy to picture that money as something abstract: chips, models, the cloud. It isn't. Most of it ends up as concrete, steel, copper, switchgear, chillers and pipework. On AI projects, the mechanical and electrical package now makes up around 75% of the guaranteed maximum price. Electrical work alone accounts for somewhere between 45% and 70% of build cost.
So the most valuable technology of the decade has a very physical bottleneck. We tend to talk about the power constraint as a grid problem, and it is one: Microsoft has disclosed an $80bn backlog of cloud orders it can't fulfil because it can't get power. But substations, switch rooms and the cable between them don't install themselves. Microsoft's president, Brad Smith, has reportedly gone further and called the shortage of electricians the number one problem facing data centre expansion in the US.
The machine that is supposed to automate everything depends on the trades that are hardest to automate.
The case that machines will build it
It's worth taking the other side seriously, because there's a real argument here.
Data centres are becoming products rather than projects. Electrical skids, power rooms and cooling modules are now built in factories and shipped to site. Prefabrication can cut on-site labour by 20–40%, and heavily modular projects are reporting schedules 30–50% shorter, with delivery dropping from 24–36 months to 16–20. Robots are drilling ceilings, marking out floors and handling heavy materials. Each generation of design is more repeatable than the one before.
If you follow that line of thinking, the industry is slowly engineering the people out of its own process. Over a long enough timeline, that's probably partly true.
Why it doesn't add up (yet)
There are three problems with the "machines will build it" story.
First, prefabrication moves the work rather than removing it. A power skid built in a factory still needs fitters, wiremen and testers to build it, and an experienced crew to set it, connect it and integrate it on site. The work shifts from site to factory; it doesn't disappear.
Second, nobody has prefabricated commissioning. Integrated systems testing, the moment you find out whether a hall actually works at full load and under failure conditions, is still done by people. It's skilled and slow, and it's where programmes are won or lost.
Third, and most interesting: the newest work is the least standardised. Average rack density went from around 16kW to 27kW in a single year. AI deployments routinely need 50–70kW per rack, and the newest systems reach 246kW. Only about one operator in five says they're ready for that. Air cooling can't handle those densities, so a large share of the next wave of work is retrofit: taking out air-cooled systems and putting in liquid cooling, inside buildings that are still running.
Every live facility is different. There's no factory version of working around an operational data hall at 2am during a planned shutdown. That's craft work, and it's growing fastest.
So the honest answer is that technology will change how the machine is built and who builds which parts. It isn't going to remove the need for skilled people any time soon. If anything, it's making the best of them more valuable.
Where the builders are going
There's a second part to this story, and it's about geography.
The work is moving away from Europe's traditional hubs. Of the roughly 700MW of capacity added in the second quarter, only 37% went to Frankfurt, London, Amsterdam, Paris and Dublin. New pipeline for 2026–28 sits an average of 175km from the major hubs, up from 46km for projects delivered in 2022–25. Projects now go where the power is: Norway, Finland, secondary German cities, the Belgian and Dutch corridor. The Nordics alone attracted around $15bn of hyperscaler commitments last year.
The people haven't moved with it. In Ireland, the same limited pool of electricians is being pulled three ways at once, between data centres, the national housing programme and home energy retrofits. So Irish and British crews do what they've always done: they travel. Irish M&E contractors now deliver capacity across the Nordics and continental Europe, and their teams go with them on rotation.
There's a historical echo here. In the 1840s, Irish navvies built a large share of Britain's railways, and the railway boom of that decade ended in a famous financial crash. Investors lost fortunes. But the track stayed, and so did the skills of the people who laid it. Whatever happens to AI valuations, the halls going up now will still be there, and the people who know how to build and run them will still be in demand.
What this means on site
- Treat people like long-lead items. You already order transformers and switchgear a year or more ahead. Key people (commissioning leads, MEP managers, experienced supervisors) need the same planning horizon. Arrange them at the procurement stage, not the week before mobilisation.
- Live-environment experience is the new premium skill. The number of people who have delivered work safely inside an operating data hall is small. Find out who on your team has done it, and hold on to them.
- Willingness to travel is now part of the job. Before you think about rate, find out who will actually go to Finland or Norway, on what rotation, and for how long. Data centre electrical roles already pay around 30% more than general commercial work. Rotation terms often decide whether people stay through the whole programme.
Three numbers worth knowing
100%: the share of UK respondents reporting shortages of qualified mechanical, electrical and plumbing workers in Turner & Townsend's August 2026 market report.
1 in 5: the share of data centre operators who say they're ready for the 50–70kW racks that AI now needs.
€5.6–5.8bn: data centre investment stranded in Ireland, by the industry's own estimates. The projects are fully permitted, on land already bought, and they can't connect to the grid.
So, who builds the machine?
People do. Specific people, with specific experience, prepared to go wherever the power is.
The more intelligent our infrastructure becomes, the more it depends on someone who knows what a good termination looks like. It's an odd fact about the age of AI, and a hopeful one if you work in construction.
One question back to you: what's the hardest role to fill on your current project? Reply and tell me. I read every answer, and the best ones will shape a future issue.
Alex The DC Build Network