AI firms are hiring electricians and carpenters by the thousands
The companies building artificial intelligence have run into a shortage money cannot fix quickly: licensed tradespeople. Their answer is to train their own — and the bet those workers are being asked to make is riskier than the wages suggest.
For three years the artificial intelligence industry competed for a very small group of people: researchers who could train large machine-learning models. That competition produced salaries that made headlines. It is not the industry's binding constraint any more. The constraint now is people who can bend electrical conduit, pull cable through it and sign off on the work.
The New York Times reported this week that AI companies and the contractors who build for them are recruiting and training electricians and carpenters in large numbers — funding their own pipelines rather than waiting for the existing ones to produce workers. It is an unusual thing for a software business to do, and it says something clarifying about what these companies actually are.
The product is a building
A data centre is a warehouse full of computers, and from the outside it looks like nothing: a long windowless shed, a fenced yard, a substation. Inside, it is one of the most electrically demanding structures anyone builds. The chips that train and run AI models draw far more power per square metre than the servers that host ordinary websites, and every watt that goes in comes back out as heat that has to be removed before the machines throttle themselves or fail.
That makes the job an electrical and mechanical project first and a computing project second. Someone has to build the switchgear rooms, the backup generators, the uninterruptible power supplies, the transformers, the fire suppression, the miles of busway and cable tray. Someone has to build the substation that connects it all to the regional grid, and often the transmission line that reaches the substation. The carpentry is less glamorous but no less necessary: forms for concrete, framing, and the endless interior fit-out of a structure the size of several football pitches.
This is why the label "AI company" is doing work it shouldn't. The largest of these firms now spend the bulk of their capital on land, steel, turbines, chips and electricity. Their dominant cost is physical infrastructure, and their dominant construction risk is whether enough qualified tradespeople exist within driving distance of a rural county that happens to have spare grid capacity.
Why the shortage cannot be solved with a pay rise
In most labour markets, a shortage is a price signal: raise wages, workers appear. The electrical trade does not work that way in the short run, because the qualification is time-based. Becoming a licensed electrician in the United States typically means an apprenticeship of four to five years — paid work under supervision, combined with classroom hours — before a worker can test for a journeyman licence and work unsupervised. No bonus compresses that.
So the industry is doing the only thing available: paying to widen the pipeline itself, funding and staffing training programmes to pull people in at the bottom. That is genuinely useful. Apprenticeships are among the few remaining routes to a middle-class income in America that do not require university debt, and the trades have been short of entrants for a generation as schools steered students elsewhere.
It also has a cost that falls on people with no stake in AI at all. When a large site absorbs every available electrician within a two-hour radius, ordinary work — a house rewire, a failed heat pump, a small commercial fit-out — either waits or gets quoted at a price designed to be refused. The same crews cannot be in two places.
The pipes are the next bottleneck
The engineering is moving in a direction that widens the demand further. As chips get denser and hotter, blowing cold air across them stops working, and operators are shifting to liquid cooling — running coolant through pipework directly to the processors, roughly the way a car engine is cooled. A rack built this way carries more plumbing than data cabling.
That shifts work toward pipefitters, welders and HVAC technicians, trades already stretched thin, and it introduces a hazard worth naming plainly: very large amounts of electricity and very large amounts of liquid, in the same tightly packed room, installed and maintained by crews assembled in a hurry. The industry's safety record over the next few years deserves closer attention than its energy-efficiency statistics.
Building it and running it are different jobs
Here is the arithmetic anyone weighing this career on the strength of the boom should hold on to. Constructing a large data centre can occupy well over a thousand workers for a couple of years. Operating the finished building takes a small fraction of that — a facilities crew, security, technicians who swap failed hardware. The employment is in the construction, and construction employment ends when the construction does.
Booms of this shape have a consistent history. Alberta's oil sands pulled skilled trades from across Canada with wages two and three times what industrial sites at home could pay; when oil prices collapsed, the camps emptied and those workers were released quickly and in bulk. The same pattern ran through the American shale fields. This is not a criticism of the workers, or even of the employers. It is what happens when demand for a skill is tied to a capital-spending cycle rather than to a stable underlying need.
Whether AI spending is such a cycle is the open question of the moment, and nobody honestly knows. The construction is real, the electricity is real, and the buildings will stand whatever happens to the firms that ordered them. But the spending rests on a bet that demand for AI computing keeps rising fast enough to justify it, and that bet is being placed by a handful of companies with highly correlated views. If it turns, it will turn for all of them at once, and the adjustment will reach the job site before it reaches the earnings call.
The sensible reading is not that the trades are a trap. A licence is portable, the skills carry over to grid work, manufacturing and building maintenance, and demand for electrical work in an electrifying economy will outlast any single boom. The sensible reading is narrower: take the wages, bank the ones that are unusually high, and treat the AI premium as temporary rather than as the new baseline. The people who did well out of the oil sands were the ones who assumed it would end.
Questions
Why would an AI company need carpenters?
Because what it is actually buying is a building. A data centre requires concrete formwork, framing and a vast interior fit-out before any computing hardware arrives, alongside the electrical and mechanical systems that account for most of the cost.
How long does it take to become an electrician?
In the United States, an apprenticeship generally runs four to five years, combining paid on-the-job training with classroom instruction, before a worker can test for a journeyman licence. That timeline is why a sudden shortage cannot be fixed simply by offering higher pay.
Do data centres provide long-term local employment?
Mostly not. Building one can employ over a thousand tradespeople for a couple of years; running the finished facility typically requires a much smaller permanent staff of technicians, facilities workers and security. The jobs are concentrated in the construction phase.
What happens to these workers if AI spending slows?
Construction hiring would contract quickly, as it did after the Alberta oil sands and American shale booms. Electrical licences remain valuable in grid work, manufacturing and building maintenance, but the unusually high wages tied to data centre projects are best treated as temporary.