Boom, bust, and a tech market that never fully reopened

Between early 2020 and early 2022, software jobs were advertised at a pace that more than doubled. On Indeed, software-development postings peaked in Feb 2022 at 229, more than twice their February 2020 level of 100. Postings across all occupations peaked later and lower: 160 in May 2022.

By Sep 2026, software postings had fallen to 77, about 34% of the peak and below where they started. Postings overall sat at 103, close to their pre-pandemic level. So the drop is not simply the whole job market cooling. Software fell much further than the market it sits in.

How to read this. The index counts postings on Indeed only and is relative to February 2020, so it shows direction and proportion, not a count of jobs. A posting is also not a hire; chapter 4 looks at the difference.

Three dates cluster in early 2022. Software postings peaked in February. The Federal Reserve began raising interest rates from near zero in March. And a tax rule took effect that, for tax years beginning after December 31, 2021, required companies to spread the cost of software development over several years instead of deducting it at once (the IRS's Section 174 guidance is in the sources).

The chart below puts them on one time axis. In July 2025 Congress restored immediate deduction of domestic research costs, including software development, for tax years beginning in 2025. Software postings stood at 65 that month and 77 in Sep 2026: a small rise from a low base, not a rebound.

Employment tells a different story from postings. Jobs in computer systems design kept growing after postings fell: they peaked in Mar 2023 at +11.4% versus January 2020, and by Sep 2026 stood at +5.7%, compared with +4.6% for all nonfarm jobs. Hiring slowed far more than headcount shrank.

Lining up in time is not cause and effect. The same period saw the end of pandemic stimulus, high inflation and, later, widely available AI tools. Chapter 6 lays out each candidate cause and what it would predict. Computer systems design is one slice of tech employment (it excludes software publishers and large platform companies), so treat it as a partial view.

Announced layoffs ran in waves. The outplacement firm Challenger, Gray & Christmas counted 2.30 million announced U.S. job cuts in 2020, almost all of it pandemic-driven. After that: 722,000 in 2023, 761,000 in 2024, 1.21 million in 2025, and 573,000 in the first nine months of 2026.

These counts cover every industry. Challenger's own reports attribute the early-2025 surge largely to federal government cuts, so the headline total says little about tech by itself. Challenger does break out technology, but its published text states a monthly tech figure only for some months, so we do not chart tech cuts on their own.

A second measure, from the Bureau of Labor Statistics, covers the information sector, which includes software publishers, data processing and hosting, telecommunications and media. Layoffs and discharges there averaged 38,000 a month in 2019 and 48,000 a month in 2026 so far (January to August).

Employers have also begun naming AI as the reason. Challenger recorded 4,247 cuts citing AI in 2023, 54,836 in 2025 (5% of all announced cuts) and 120,136 in the first nine months of 2026 (21% of all announced cuts).

This is a stated reason, not a measured cause. Challenger began tracking AI as its own reason in May 2023, and through 2025 its monthly reports often grouped it under "Technological Updates" (20,219 cuts through September 2025). The 2023, 2025 and 2026 totals are therefore not like-for-like. The jump means more employers now name AI. It could reflect substitution, relabeling of other cost-cutting, or both, and these numbers cannot say which. Announced cuts are also plans, not confirmed separations.

If employers post jobs they never fill, openings pile up relative to hires. The BLS's JOLTS survey counts openings on the last day of each month and hires during the month, so the ratio between them is a rough gauge of that gap.

In the information sector there were about 1.49 openings per monthly hire in 2019. At the Apr 2022 peak that rose to 2.66. That is consistent with employers advertising far more roles than they filled, which is what "ghost jobs" would look like, but it is also what a tight market where roles take longer to fill would look like. The data cannot separate the two.

Hires in the sector averaged 95,000 a month in 2019 and 106,000 in 2022. Over the 12 months to August 2026 they averaged 75,000. Openings fell back faster than hires did, which narrowed the gap.

From March to June 2026 the information sector's ratio ran between 0.96 and 1.06, close to one opening per hire. Those are the lowest readings in this series, apart from a few months in 2020 and 2021, and they sit below the whole economy's roughly 1.4. By this measure, the boom-era gap between advertised roles and hires has narrowed sharply.

A few quiet months are not a trend. The ratio swings a lot: it was near 2.0 as recently as September 2025. And August 2026 shows 123,000 openings against 44,000 hires, which lifts the 3-month ratio to 1.6. JOLTS industry estimates carry wide margins of error and the newest month is preliminary, so we are not reading that jump as a trend either; it will be revised. Also, "information" is only a rough stand-in for tech, because it includes telecom and media.

The market is not equally hard on everyone. The sharpest strain is at the entry point. Over the 12 months to Jun 2026, recent college graduates (ages 22 to 27) were unemployed at an average 5.6%, up from 3.9% in 2019. For all workers the rate rose from 3.6% to 4.2%. The gap between graduates and everyone else widened from 0.3 to 1.4 percentage points.

The share of graduates in jobs that typically do not require a degree stayed near 42% (42% in 2019). So the squeeze shows up as unemployment, not as a rise in underemployment.

Computer majors sit near the top of the list. In 2024 data from the American Community Survey, recent computer science graduates were unemployed at 7.0%, the 4th highest of 74 majors. Computer engineering ranked 2nd. The median major was 4.2%. Underemployment among computer science graduates, though, was low (19%, against about 42% for all recent graduates), which suggests those who find work mostly find degree-level jobs.

Read the rankings loosely. The by-major figures cover a single year (2024) from a survey, and samples for individual majors are smaller, so neighboring majors are not precisely ranked. The unemployment series is smoothed over three months and excludes students, and October 2025 is estimated because data were missing.

Now the people the headline rate leaves out. The unemployment rate counts only people actively looking for work. Comparing 2019 averages with the 12 months to Sep 2026, the headline rate (U-3) rose from 3.7% to 4.3%, and the broader U-6 measure, which adds discouraged workers, other marginally attached workers and people stuck in part-time jobs, rose from 7.1% to 8.0%. The number of people outside the labor force who say they want a job rose 19%, from 5.0 million to 6.0 million. Within that, marginally attached workers rose from 1,403,000 to 1,729,000 and discouraged workers from 382,000 to 473,000.

The group outside the count grew by about 19%, close to the 16% rise in the headline rate. So by these measures the hidden group has grown about as fast as the visible one, not faster. And for people in the middle of their working lives, work is not scarcer than before the pandemic: participation among workers aged 25 to 54 averaged 83.7% over the last 12 months against 82.5% in 2019, and their employment rate was 80.6% against 80.0%. The weakness is concentrated among people trying to enter, and in particular fields, more than across the whole labor force.

What these measures cannot show. People outside the labor force may want a job for reasons unrelated to the market, and the series that counts them is not seasonally adjusted, so we compare full-year averages. An aging population also lowers overall participation, which is why we use the prime-age (25 to 54) figures. None of this says why any individual stopped looking.

Chapters 1 to 5 describe what happened. This one asks why, and says plainly how far the evidence goes. The short version: several plausible causes overlap in time, and the data here cannot rank them.

The table lists each candidate, what we would expect to see if it were the main driver, and what our data shows. The status column is about our evidence, not about whether a cause is real.

Sources: Indeed, JOLTS, FRED, Challenger, Section 174 sources.

Two things the table does support. First, no single-cause story fits: the decline started in early 2022, before AI cuts were being reported, and it overlaps with the end of the boom, the rate rises and the tax change. Second, AI-attributed cuts appear mostly in 2025 and 2026, and the strain shows up at the entry level more than in headcount: computer-systems-design employment is still above its January 2020 level.

Markets can also feed themselves. One loop is measurable in part. When layoffs and uncertainty make staying put safer, fewer people quit. Quits create openings, because employers backfill, and openings become hires. In the information sector, quits averaged 46,000 a month in 2019 and 34,000 over the last 12 months, and the whole-economy quits rate fell from 2.32% to 1.95%.

The diagram shows that loop and a second one, about the entry-level pipeline. We measured the levels in the solid boxes. The arrows are proposed mechanisms: we did not test whether one step causes the next, and the dashed boxes are steps we have no data for.

The second loop could break. If the bench of mid-level engineers gets thin enough, the shortage raises hiring pressure again, but the delay could be years. Nothing here says it will happen or when.

What would settle this. Ranking these causes needs firm-level data: which employers cut, in what roles, and what they reported about taxes, costs and AI at the time. Layoff notices filed under the WARN Act and company filings are the obvious places to look, and we set both aside for this draft. Until then, treat each cause here as a hypothesis with a stated status.