AI Did Not Kill Engineering Jobs. It Repriced Them.
Engineering roles are down 11% since 2019, while tech jobs overall fell 25%. That is the headline from SignalFire’s State of Talent 2026, reported by TechCrunch on June 24. Engineers were 55% of all 2025 new hires, up from 46% in 2019. At early-stage startups, engineering headcount grew 7%. The category that the “AI replaces coders” thesis marked for deletion turned out to be the most resilient one in tech.
The forecast that called for this collapse got the mechanism wrong, not just the timing. The work itself kept existing, in a different shape and at a different price. It repriced.
The numbers point the other way
The cleanest way to read the SignalFire data is to compare two lines. One line is total tech employment, down a quarter from its 2019 baseline. The other is engineering employment, down a fraction of that. When a category shrinks far slower than its parent, it is gaining share, not losing it. Engineers went from 46% of new hires to 55% in six years. That is a sector reallocating toward engineering, not away from it.
The executive commentary lines up with the data instead of fighting it. Nvidia’s CEO told the press that software engineers are “busier than ever.” An Anthropic economist, studying labor exposure to AI, found no “material difference in unemployment rates” between AI-exposed and non-exposed workers. Statements like these track rising demand and shifting work, the opposite of what you would hear during a profession’s collapse.
A caution on the source: SignalFire is a venture firm with a portfolio interest in startup hiring staying healthy, and the survey reflects the companies it tracks. Treat the direction as well-supported and the exact percentages as indicative. The direction is the part that matters here, and the direction is unambiguous.
Jevons, not the apocalypse
There is a 160-year-old economic result that explains the resilience better than any vendor deck. In 1865, William Stanley Jevons observed that more efficient steam engines did not reduce England’s coal consumption. They increased it. When a resource gets cheaper to use, people find vastly more uses for it, and total demand rises. Cheaper light gave us cities that never go dark. Cheaper compute gave us a phone that does what a 1990s data center could not.
Software engineering is now living through its Jevons moment. LLMs made one specific slice of the work, writing a first-draft implementation from a clear spec, dramatically cheaper. The naive forecast assumed the total amount of engineering would stay fixed, so cheaper production meant fewer engineers. The forecast had the elasticity backwards. When shipping a feature gets cheaper, companies ship more features, start more products, and attempt systems they would not have staffed before. The early-stage +7% number is Jevons in miniature: lower the cost of building, and more building gets attempted.
Commoditization and value move to different layers. The implementation-heavy, generalist middle of the work, the part an LLM can draft from a prompt, is exactly the part being commodified. The value migrates to the parts the model cannot do alone: making a system reliable under real production load, scaling it without it falling over, securing it against adversaries who also have LLMs, instrumenting it so you can see what it does at 3am, and exercising the operational judgment that decides which of a dozen plausible designs survives contact with reality.
The moat is one hard thing, known exceptionally well
The blogger grandimam put the second half of the argument cleanly: “the biggest returns come not from knowing a little about everything, but from knowing one hard thing exceptionally well.” That sentence is the practical guide to where engineering labor reprices upward.
The market is not paying a premium for the “AI engineer” job title. Slapping “AI” on a resume is itself a commodity move now, available to everyone in an afternoon. The premium goes to demonstrable depth in something that stays hard even when implementation is free. Distributed systems consistency. Database internals under agent-generated query load. Security architecture. Latency engineering at scale. The on-call judgment that comes only from having broken production and fixed it. These are not skills you draft from a prompt, because the model has no production scars and no accountability when the pager goes off.
This is the thread that connects to a position we have argued before: the humans are the baseline, not the agents. We measure engineering teams as humans plus AI on one scoreboard, because the durable value is the human depth that decides whether the cheap implementation actually ships, holds, and survives. We made the spend-side version of this case in Enterprise AI Repricing, where AI budgets are reallocating rather than collapsing. The labor market is running the same play with people. The generalist implementation layer gets cheaper; the specialist judgment layer gets dearer. Both are repricing, in opposite directions, at the same time.
What this means for hiring and for careers
For engineering leaders, the SignalFire numbers are an argument against the headcount-freeze reflex. If your competitors read “AI replaces engineers” and froze hiring, the data says they misread the market and you can take share, especially at the senior and specialist end where the repricing concentrates value. The bottleneck in an AI-accelerated org is whoever can take a first draft to reliable production, a scarcer skill than ever before.
For individual engineers, the strategy that the data rewards is depth over breadth, and depth in something that resists automation. A generalist who can prompt an LLM to produce a CRUD app is competing with everyone who can do the same, which is now most of the field. An engineer who owns database performance, or distributed consensus, or the security model, owns a thing the model cannot replace and the business cannot ship without. The job market for engineering talent is not contracting. It is sorting, by depth.
Do this now
Run one calculation on your own team this quarter. List every engineer and ask, for each: what is the one hard thing this person knows exceptionally well, that an LLM cannot do from a prompt? If you can name it, that is a repriced-upward asset, and you should be protecting and growing it. If the honest answer for someone is “generalist implementation,” that is not a firing signal; it is a development plan. The market is repricing the work toward depth. Your hiring, your reviews, and your training budget should reprice with it, starting now.
This analysis synthesizes AI was supposed to kill engineering jobs, but new data suggests they’re the most resilient (TechCrunch / SignalFire State of Talent 2026, June 2026) and Repricing of software engineering labor (grandimam, June 2026).
Victorino Group helps engineering organizations measure teams of humans plus AI on one scoreboard and find the depth that reprices upward. Let’s talk.
All articles on The Thinking Wire are written with the assistance of Anthropic's Opus LLM. Each piece goes through multi-agent research to verify facts and surface contradictions, followed by human review and approval before publication. If you find any inaccurate information or wish to contact our editorial team, please reach out at editorial@victorinollc.com . About The Thinking Wire →
If this resonates, let's talk
We help companies implement AI without losing control.
Schedule a Conversation