What Happened to Workers Who Joined the Last Tech Boom?
September 29, 2026
Most of the debate about artificial intelligence and the labor market is about displacement: which workers AI will replace, and how fast. In his July 2026 testimony to Congress, Federal Reserve chair Kevin Warsh announced a task force to study what the scale of AI investment means for "America's productive capacity and for American workers," adding that "we might be seeing changes of a different order" than past technological advances. He has compared the current wave to the productivity boom of the late 1990s. If that is the right reference point, the last boom has a question to ask about a risk that gets much less attention than displacement: what happens to the people who join the booming sector while its technology is still being worked out?
In a recent paper, Johan Hombert (HEC Paris) and I follow the skilled workers who started their careers in the French information and communication technology (ICT) sector during the dot-com boom. Picture two engineers leaving the same school in 1999, the same age, the same degree. One takes a job at a software company that is hiring robustly and paying handsomely for talent, the other at an aircraft manufacturer. The software job pays 8 percent more to start. Fifteen years later, averaging across everyone who made that choice, the 8 percent lead has become a 7.7 percent deficit. Of these two workers, the one who went into aircraft is the better paid one.
The reason we use French data is that they follow the same individual from one employer to the next, year after year, so someone who took a first job at a start-up in 1999 can still be located in 2015 with their wage on file. No comparable public US source follows individual careers over that span, which is why the US evidence on this question stops at how many young workers entered the sector rather than what happened to them afterward. The rest of this post explains where the 7.7 percent figure comes from, why the evidence points to skills losing their value rather than to the bust, and which features of that episode are visible in US data today.
A boom that absorbed a quarter of skilled young workers
The dotcom boom combined rapid technological change with abundant financing. In France, the stock market value of ICT firms ran up sharply between 1997 and 2000, and equity issuance by ICT firms peaked in 2000 and 2001. Cheap capital let technology firms bid aggressively for talent.
Skilled young workers moved toward the sector in large numbers. Figure 1 shows, for each year, the share of skilled workers taking their first job in France who started in ICT. The share was about 11 percent in 1994. At the peak of the boom, in 1999 and 2001, the ICT sector hired about one-quarter of all skilled workers entering the labor market, and its share of total skilled employment rose from 10.6 percent in 1996 to 14.0 percent in 2001.
The inflow ran almost entirely through people starting their careers, and figure 2 shows this on both sides of the Atlantic. In France (panel A), the ICT share of skilled employment among workers in their first four years on the job rose from about 13 percent in 1994 to nearly 22 percent in 2001, while among workers with five or more years of experience it moved from about 10 percent to about 13 percent. In the United States (panel B), the ICT share of skilled workers under 30 rose from about 12 percent in 1995 to nearly 17 percent in 2001 before falling back, while the share of workers over 30 barely moved.
The concentration of the boom among new entrants matters for what follows. What you learn in your first years on the job is built on whatever technology your employer is using at the time. The people who joined ICT between 1998 and 2001 learned the boom's technologies, and those technologies were unusually experimental.
Paid more at entry, yet earning less 15 years later
Our approach compares people who entered the labor market in the same years but started in different sectors. These data come from a mandatory employer filing that covers every private-sector job in France, and for a subsample of workers an identifier stays with the person as they change employers, so we observe a career rather than a snapshot. We split skilled entrants between 1994 and 2015 by the year they took their first job: those who started before the boom (1994 to 1996), those who started during it (1998 to 2001), and those who started after it (2003 to 2005). Inside each group, we compare the ones whose first job was in ICT to people of the same sex, age, starting year, and occupation who started somewhere else.
Figure 3 traces this comparison for the whole sample. Workers who joined ICT during the boom started with a wage premium of about 8 percent, consistent with a sector competing hard for scarce talent. The premium then eroded, turned negative in the mid-2000s, and kept widening: by 2015 these workers earned about 9 percent less than comparable peers who had started outside ICT. Averaged over the final years of our data, 2011 to 2015, the shortfall is 7.7 percent. Because a year of early-career experience raises wages by 3.0 to 4.4 percent in our sample, that shortfall is equivalent to losing about two years of experience. The loss also deepens with exposure: among those who started during the boom, each additional year spent in it adds 2.3 percent to the long-run shortfall.
One might think the software job's higher starting wage was the price of taking a known risk, so that over a whole career the two engineers come out even. We can add it up. Take everything each engineer earned from 1999 to 2015, including short and part-time jobs, and discount later years at 5 percent a year so that the early premium counts for more than a late shortfall. Two years in, the software engineer's cumulative earnings were 5 percent ahead. By 2005 the two were even, and by 2015 the software engineer had earned 5.5 percent less in total, about 24,000 euros. The premium at entry was too small, and lasted too briefly, to pay for what came after. If anything, this understates the gap: a risk-averse worker would have needed a premium large enough to come out ahead on average, not merely even, so the entry wage fell short of compensating for the risk by more than this 5.5 percent. Whether the workers who took those jobs saw the risk coming, the data cannot say; what they say is that the entry wage did not cover it.
The obvious culprit behind the eventual shortfall is the crash itself: our engineer was in tech when the market faltered in 2001. But two things rule it out.
The first is the gold line in figure 3, representing the workers who entered ICT in 2003 to 2005, after the crash. If the bust had left a lasting mark on ICT pay, it would show up in their earnings too. It does not, however: their wages track those of comparable workers who started elsewhere all the way to 2015, and this observation holds when we compare the two groups at similar firms, including firms whose fortunes after the boom were similar. So the losses of those who started during the boom cannot come from a sector that stayed depressed.
The second is the shape of those losses. A crash hits everyone in a sector at once, on the same date, while these losses grow with how long a worker had been in the boom before it ended, by the 2.3 percent per year noted above. The shortfall lines up with how much of a career was built on the boom's technologies, not with the timing of the collapse.
A different objection is that the boom attracted weaker workers to begin with. Those who entered ICT in 1994 to 1996 chose the sector before it turned hot, so nothing about the boom could have drawn them in, and they show the same long-run decline. Their skills were formed during the boom even though their choice was not made then. Job loss does not explain the pattern either; the ICT workers who started during the boom were somewhat more likely to lose their jobs after 2001, but the effect is an order of magnitude too small to account for the wage shortfall.
Skills tied to experimental technologies depreciated fastest
What our engineer learned in 1999 stopped being worth much once the technology settled down. During the boom, ICT firms churned through approaches at an unusual pace. In US patent data, ICT patents filed during the boom cited younger patents, showed more dispersed citation outcomes, and used more novel language than before or after. Hand-coded static websites, the thing to know in 1999, were displaced within a few years by style sheets and database-driven frameworks.
Now add a third person to the pair we started with: someone who joined the same software company in 1999, but in sales rather than engineering. Fifteen years later that person is doing fine. The wage shortfall falls on workers in STEM occupations, at 9.7 percent over 2011 to 2015, while the people who started at the same kinds of ICT firms in sales, administration, or general management show nothing comparable. Across industries, the shortfall appears only where the STEM share of the workforce is high. So the engineer in our example is carrying a shortfall of 9.7 percent rather than the 7.7 percent average, and the colleague who sat down the corridor in sales is carrying nothing.
The money went where the skills would face obsolescence fastest. Capital flowed disproportionately to the ICT industries with the highest STEM intensity and the fastest technology turnover afterward. As figure 4 shows, workers who started in industries receiving above-median capital inflows during the boom carry a long-run wage shortfall of roughly 14 percent, while workers who started in industries that attracted less of the capital came out roughly even. (The estimate is about 4 percent, too small for us to call it different from zero.)
This is the part of the episode most relevant for thinking about AI. Beyond inflating valuations, the financing boom determined how many workers were exposed to fast-depreciating skills, and it amplified the aggregate loss by sending the largest groups of entrants to the firms where skills would become obsolete fastest.
How tight is the parallel with AI?
The French episode is one boom in one country, and nothing guarantees that AI will repeat it. What our research offers is a set of observable conditions: a financing boom concentrated in the new technology sector, a large reallocation of young skilled workers into that sector, high entry wages, and rapid technology turnover. Each is measurable in US data today.
The financing condition is already met, on at least one measure. Figure 5 plots private fixed investment in information processing equipment and software as a share of GDP, a series that covers both eras. The share rose from 3.24 percent in early 1995 to a dotcom peak of 4.46 percent at the end of 2000, then fell back to 3.43 percent by early 2003. In the current boom it has climbed from 3.91 percent at the end of 2019 to 4.88 percent in the first quarter of 2026, above the dotcom peak.
The labor market conditions are partially visible. The share of US online job postings requiring AI skills rose from 0.84 percent in 2019 to 1.62 percent during the first eight months of 2024, as documented in an earlier Policy Hub: Macroblog post, and Atlanta Fed survey work finds firms planning substantial AI investment. What we cannot yet see is whether AI is absorbing skilled workers entering the labor market on the scale the ICT sector did in the late 1990s. The technology-turnover measures we construct from patent data for the ICT boom could also be computed for AI in real time.
That leaves the question of what an AI bust would do. What we measure is not a bust: the losses appear in workers whose skills were formed during the boom and not in those who joined the same sector two years later, and they scale with time spent inside the boom rather than with the collapse. The risk we document therefore does not require a crash. Skills built on a technology that is still moving lose value once it settles, whether or not valuations fall on the way. A collapse in AI valuations would be a second and separate blow, one that works through job loss and a smaller sector rather than through obsolescence. Workers who have specialized in today's AI tools would then carry both, and only the first of the two is what our estimates capture.
Lessons from the last boom
The dotcom episode suggests that the labor market risks of a technology boom are not limited to the workers a new technology might replace. The workers who build and deploy the technology face their own risk: that the skills they acquire while the technology is unstable can lose value quickly once it matures, and the financing that fuels the boom decides how many workers are exposed. In the French data, the talented individuals who joined the boom are not the ones who ultimately benefited from it.
None of this discussion suggests that the AI boom will end the same way; our evidence concerns the workers who joined the last boom, not whether the boom's investments paid off in aggregate. But it says an AI boom can damage careers in two ways, not just one. The familiar way is losing a job to the technology. The other is spending your late twenties becoming expert in a version of it that has become obsolete by the time you turn forty, which is what happened to our engineer. How many people are walking into that second position right now is something we can already measure.