I own the picks and shovels. Chip equipment, memory, the industrial names that sell into data-centre construction. When someone tells me the build-out is a bubble, I’m not a neutral party, because a bubble would cost me money. So take what follows knowing that.
The Economist’s number is right.
Not approximately right. Not right by accident. Right.
Roughly a million American jobs traceable to artificial intelligence since 2023, set against something like two hundred thousand layoffs blamed on it. The figure is already being flattened into a claim nobody at the magazine made, which is that the economy has netted out a million jobs ahead. It hasn’t, and the article doesn’t say so. But the count itself holds up, and it deserves better than the use both camps are putting it to.
Ask who is actually cashing the cheques
Read what the million is made of. The Economist tracked five industries sitting at the centre of the data-centre build-out, from electrical contracting through to equipment manufacturing, and found employment in them running about 320,000 above where the broader trend would have put it. LinkedIn, looking at its own users, counts close to half a million data-centre jobs created in America between 2023 and 2025, with technicians and engineers among the most common recent hires. Indeed finds installation and maintenance work at data centres advertising something near forty per cent above comparable work elsewhere.
That is a shortage of people who can pull medium-voltage cable, commission chillers, and land a transformer on a pad in Loudoun County. It isn’t a shortage of people who can write prompts.
The genuinely new white-collar layer exists, and Gad Levanon at the Burning Glass Institute has tried to size it: somewhere around one per cent of professional jobs. One per cent. The larger white-collar number, the roughly 730,000 above-trend engineers, developers and data scientists, is downstream of the same capital spending as the electricians. Somebody has to design the thing being built and then keep it running.
The count measures budgets, not capability
Here’s the test I’d apply to any jobs figure attached to a technology. What would have to change for the number to fall?
Suppose the models stop improving tomorrow. Not degrade, just plateau. The hiring keeps climbing anyway, for a year at least, because a shell in New Albany with a signed interconnection agreement and a delivery slot for switchgear gets finished whether or not the next release is any good. Contracts don’t care about benchmark scores.
Now run it backwards. Suppose the models keep getting better and three hyperscalers each trim capital spending by a fifth. The electricians go home. The count falls off a cliff with the technology working better than ever.
A jobs number that moves with budgets and ignores capability is a spending indicator wearing a labour costume. It’s a good indicator, and I watch it, because it tells me the build-out is real in a way that press releases about model capability never will. It tells me concrete is being poured and copper is being ordered. It doesn’t tell me anyone has found a use for the finished product that pays for the copper.
What I still don’t buy
The apocalypse, mostly. People have been predicting mass technological unemployment since the power loom, and they’ve been early every single time, which in markets is the same as wrong.
But I want to be careful about the reverse error, because the crowd celebrating this figure is committing it. The back-office thinning is real. Entry-level hiring in professional services has gone quiet in a way that isn’t explained by the business cycle, and if you graduated in 2025 into a hiring freeze, the fact that a data-centre technician in Ohio is doing well doesn’t help you. Those are different labour markets and different people. Nobody retrains a redundant claims processor into a substation commissioning engineer over a weekend.
What I reject is the confident reading in either direction. The doomers point at a weak 2026 labour market and blame the machine. The boosters point at a million jobs and credit the machine. Both are reading the effects of a spending cycle as evidence about the technology, and they’ll both be surprised by the same thing when capex flattens.
A million jobs is a real number. It’s also a receipt for money spent.
It tells you what got built. It doesn’t tell you what works.
