Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
For three years, investors have been told to brace for the great AI layoff. Analysts modelled it, commentators warned about it, and every quarterly jobs report was scanned for the first sign of white-collar collapse. It never really came. Headline unemployment stayed manageable, corporate profits kept climbing, and executives kept describing their AI rollouts with the word everyone loves to hear on an earnings call: efficiency.
So was the panic simply wrong? Not quite. The adjustment did happen. It just happened in a line of the spreadsheet most investors never look at. Companies did not fire their way into the AI era. They quietly stopped hiring the people at the bottom of the pyramid, and a cost that is never incurred never shows up in a restructuring charge.
That distinction sounds academic. It is not. A layoff is a one-off expense that gets booked, explained and forgotten. A missing cohort of junior staff is a liability that sits off the balance sheet for years, and then arrives all at once.
The Number Everyone Quotes, and What It Really Measures
The figure that launched a thousand headlines came from MIT and Oak Ridge National Laboratory. Their Project Iceberg simulation models 151 million American workers across 923 occupations and more than 32,000 individual skills, then asks how much of that skill base current AI systems could technically perform. The answer was 11.7% of total wage value, roughly $1.2 trillion a year.
Most coverage translated that into “AI can replace 12% of workers today.” The researchers themselves are more careful. The Iceberg Index project page describes it as a measure of technical exposure, explicitly not a prediction of displacement or a timeline for adoption. It is a capability map, not a layoff forecast.
That matters for investors because the scary aggregate number points you in the wrong direction. If you go looking for a broad wave of job losses, you will not find one, and you may conclude the risk has passed. The real damage is narrower, slower and far easier to miss.
Where the Damage Actually Landed
The clearest evidence comes from payroll records rather than surveys. Stanford’s Digital Economy Lab used ADP data covering millions of workers to track what happened after generative AI went mainstream. The latest revision of their Canaries in the Coal Mine working paper finds no sign of widespread, economy-wide displacement. What it does find is that employment among 22 to 25 year olds in AI-exposed occupations sits about 19% below where it would be had it tracked their less-exposed peers. Experienced workers in the same fields show no comparable gap. And the mechanism is mostly hiring that never happened, not people being shown the door.
The reason the young get hit first is almost uncomfortably logical. Junior professional work is the work that has been written down. Drafting a first-pass memo, building a standard model, cleaning a dataset, writing routine code: these tasks live in procedure manuals, templates, textbooks and documentation, which is precisely the material language models were trained on. Senior work rests on something that was never written down: knowing which client will walk if a deadline slips, sensing when a counterparty is bluffing, recognising that a clean-looking number is wrong because you have seen that pattern blow up before.
So the machine absorbs the codified layer, and the people who hold the uncodified layer become more productive, not less. From the boardroom, this looks like a free lunch.
The Plot Twist Hiding in the Data
Here is where the consensus story gets more interesting. Not every company is responding the same way, and the split may be the most investable part of this whole theme.
Research co-authored by MIT Sloan’s Lawrence Schmidt, covering AI exposure across firms from 2010 to 2023, found that when AI can perform most of the tasks in a particular role, that role’s share of a company’s workforce falls by around 14%. Yet the same work, summarised in MIT Sloan’s write-up of the study, found that a large increase in AI use was associated with roughly 6% higher employment growth and 9.5% stronger sales growth over five years. Firms that adopted AI did not necessarily shrink. Many simply grew faster.
More recent data pushes even harder against the doom narrative. A June 2026 working paper from Ramp’s economics team and Revelio Labs, linking actual corporate AI spending to workforce records across more than 21,000 US firms, found that the heaviest AI spenders grew headcount by about 10% over two years, with entry-level headcount up 12%. Light adopters showed no meaningful change at all.
Treat that study with appropriate caution. The heavy adopters were already larger, faster-growing and more often venture-backed before they adopted. It is an early paper, and it comes from a company that sells corporate spending software. But set beside the Stanford findings, a pattern emerges. The junior hiring freeze is not a law of nature imposed by the technology. It is a management choice. Some companies are using AI to grow and are still bringing in young talent. Others are using it to shrink the cost base and are shutting the front door.
For an investor, those are two very different businesses, even when this year’s margins look identical.
Margin Today, Liability Tomorrow
Think of a junior hiring freeze the way you would think of a company that stops spending on maintenance. Profits jump immediately. Nothing breaks in year one or year two. The bill arrives later, larger, and usually at the worst possible moment.
The arithmetic of a talent pipeline is unforgiving. This year’s 23-year-old analyst is the 2031 team lead and the 2035 department head. Skip enough cohorts and, a few years out, a company discovers it has senior people approaching retirement and almost nobody in the middle who has learned the business well enough to replace them. The options at that point are all expensive: bid against rivals for a thin pool of mid-career talent, promote people before they are ready, or accept slower execution. None of that appears in today’s guidance.
History offers a useful warning about getting this wrong. In 2016, AI pioneer Geoffrey Hinton argued that the world should stop training radiologists because machines would soon outperform them. Had medical schools listened, the result would have been a shortage. Instead, the Mayo Clinic built AI into its workflow and grew its radiology staff by 55% to around 400 doctors. The technology made each radiologist more capable, which expanded what the department could take on.
The ATM story runs the same way. Economist James Bessen found that the number of tellers needed to run an average urban bank branch fell from about 20 to 13 between 1988 and 2004, yet teller employment did not fall, because cheaper branches meant banks opened far more of them. Automation lowered the cost of the service and the market for it grew.
The difference this time is that previous waves left the apprenticeship rung intact. Tellers still learned the bank from the counter up. Junior radiologists still read thousands of scans. When the entry-level work itself is what gets automated, the training ground goes with it, unless a company deliberately decides to keep it.
As we argued in our piece on who actually goes broke when the AI bubble pops, the costs of a cycle rarely land where the headlines point. This is another case of that. The people paying first are graduates who never get the call. The companies that froze them out pay later.
The Bottleneck Wears a Hard Hat
There is an old idea in robotics known as Moravec’s paradox. The things humans find intellectually demanding, such as formal logic, calculation and writing code, turn out to be relatively easy for computers. The things we do without thinking, such as balancing, gripping, judging space and adapting to a messy physical environment, turn out to be fiendishly hard. Fifty years of progress have not overturned it.
Nowhere is that clearer than in the AI build-out itself. The trillion-dollar data center boom runs on copper, concrete and switchgear, and someone has to install all of it. Microsoft president Brad Smith has called the electrician shortage the number one problem slowing the company’s US data center expansion, according to Fortune’s reporting on the trades gap. The same report cites estimates that more than 300,000 new electricians will be needed over the next decade, that electrical work accounts for 45% to 70% of data center construction costs, and that the standard apprenticeship takes four to five years. You cannot compress that with a better model.
The education market has noticed. Universal Technical Institute reported new student starts up 13.8% in its fiscal second quarter, with management directly crediting AI for accelerating demand for skilled workers. At the same time, enrollment in four-year computer science programs fell 11.2% in the 2025 to 2026 school year, while engineering rose 7.3%, according to National Student Clearinghouse figures reported by Built In. Even at MIT, undergraduate enrollment in its core computer science major is down about 18% from its 2022 peak, according to the student newspaper The Tech.
The contrarian caution is that “buy the trades” is no longer a secret. Training providers are spending heavily on new campuses, which has squeezed their margins even as enrollment grows, and the market will not pay for growth forever if profits lag. There is also a less obvious angle: a labour bottleneck in the physical build-out is a cost pressure on every data center project, which means it is a quiet headwind for the returns on AI capital spending, not just a tailwind for vocational schools.
Karp’s Other Bet
Palantir chief executive Alex Karp offered the bluntest version of this thesis on the TBPN podcast. As Fortune reported, he said there are basically two ways to know you have a future in the AI era: vocational training, or being neurodivergent. His reasoning is that conventional education rewards learning a framework and applying it correctly, which is exactly what language models do best. Minds that struggle to follow the standard playbook, Karp, who is dyslexic, argues, are forced to find other routes to an answer, and those routes become valuable once the playbook is automated.
Palantir has backed the talk with hiring programmes. Its Meritocracy Fellowship recruits high school graduates who skip university, requires an SAT score of at least 1460 or an ACT of 33, and pays $5,400 a month for a four-month stint in New York. Its Neurodivergent Fellowship offers full-time roles paying between $110,000 and $200,000, with no formal diagnosis required.
Not everyone buys the framing. Daniel Wendler, a psychologist who is himself autistic, told Fortune that Karp has a point about different thinkers being valuable when organisations are changing, but rejected the idea that every neurodivergent person is a hidden genius ready for the AI economy.
For investors, the label is less important than the behaviour. One of the most AI-intensive software companies in the world is not freezing its junior intake. It is building unconventional new pipelines to feed it. That is a tell worth noticing.
What to Watch Instead of the Unemployment Rate
If the real story is hiring that does not happen, investors need different signals. A few worth tracking:
Revenue per employee, and why it is rising. Growth driven by higher revenue is healthy. Growth driven mainly by a shrinking denominator deserves a lower multiple, because part of that margin is borrowed from the future.
Graduate programmes and headcount mix. Cuts to campus recruiting, disappearing analyst classes and a workforce that skews steadily older are early warnings, even when the totals look stable.
Mid-career wage inflation. If the pipeline thesis is right, the first visible symptom will be bidding wars for experienced staff in fields like software, law and finance a few years from now. Businesses that kept training will be able to promote from within. Those that did not will pay up.
The physical constraint on AI spending. Data center timelines are increasingly set by labour, not chips. Delays and cost overruns on the build-out side are a risk to the AI capex trade that few models capture.
Paying for the obvious. Trades-exposed businesses have a genuine structural tailwind, but crowded themes get expensive. Separate the companies converting demand into profit from those simply spending to chase it.
The AI job shock was supposed to be loud. Instead it has been almost silent, a cohort of young professionals who were never brought in, and a set of corporate margins that look better than they should. Silence is not the same as safety. The companies pulling up the ladder are making a bet that they will never need the people they declined to train. Investors who assume that bet is free may find the invoice arrives later, and larger, than the market expects.
