The Jevons Paradox: Why AI Is Creating More AI Engineering Jobs, Not Fewer
Last updated: July 7, 2026
Direct Answer
AI is creating more AI engineering jobs, not fewer. Engineers made up 55% of new hires at major tech companies in 2025, up from 46% in 2019, according to SignalFire data reported by TechCrunch. The mechanism is the Jevons paradox: when AI makes code cheaper to produce, companies build more software, which requires more engineering judgment to architect, integrate, and maintain.
Overview
- What the SignalFire/TechCrunch hiring data actually shows
- The Jevons paradox, from 1865 coal to 2026 code
- Why cheaper code explodes demand (the Elastic Backlog)
- Breaking down the 46% → 55% shift
- What the displacement narrative gets wrong
- FAQ
What the Data Says About AI Engineering Jobs in 2026
The prevailing 2023–2024 narrative was a displacement story: AI would write the code, and software engineering headcount would collapse. The data now says otherwise.
SignalFire's State of Tech Talent Report 2026, covered by TechCrunch in June 2026, tracked hiring across 12 "Tech Majors": Alphabet, Meta, Apple, Amazon, Microsoft, Netflix, Nvidia, Tesla, Uber, Airbnb, Block, and Stripe. Three numbers stand out:
- Engineers were 55% of all new hires at the Tech Majors in 2025, up from 46% in 2019.
- Overall hiring at those companies is down 25% from 2019, but engineering hiring is down only 11%. Engineering is the most resilient function, not the least.
- Early-stage startups hired 7% more engineers in 2025 than in 2019.
So during the exact window when AI coding assistants went from novelty to default tooling, engineering didn't just survive the hiring contraction; it took a larger share of it. That pattern has a name, and it's 161 years old.
What Is the Jevons Paradox?
The Jevons paradox is the observation that making a resource more efficient to use increases, rather than decreases, total consumption of it. William Stanley Jevons documented it in his 1865 book The Coal Question: more efficient steam engines made coal cheaper per unit of work, making coal-powered machinery economically viable in far more industries, thereby increasing total coal consumption.
The condition that makes the paradox fire is elastic demand: demand that expands faster than efficiency reduces per-unit cost. Software demand is about as elastic as demand gets. When building software gets 30% cheaper and faster, companies don't pocket the savings and go home early. They build more things, often far more than the efficiency gain "freed up."
The table above maps Jevons' 1865 coal observation directly onto AI-assisted software development: in both cases, per-unit efficiency gains triggered a demand expansion large enough to increase total consumption of the underlying resource-coal then, engineering capacity now.
The Elastic Backlog: Why Cheaper Code Explodes Demand
The Elastic Backlog: every engineering org's graveyard of ideas that were never technically impossible, just too expensive to justify: internal tools, custom integrations, experimental features, migrations deferred for the third year running. AI didn't create this backlog. It repriced it.
,
When the cost per feature drops, items cross the worth-building threshold in bulk. The internal dashboard that couldn't justify two engineer-months at 2022 prices clears the bar at 2026 prices. Multiply that across every team in every company and you get a demand explosion that outruns the efficiency gain.
The GitHub Octoverse 2025 report shows the output side of this repricing: 630 million repositories, just under a billion commits in 2025 alone, 36 million new developers in a single year, and 4.3 million AI-related repositories — nearly double the 2023 count. GitHub's Copilot coding agent alone authored over 1 million pull requests in five months.
But code doesn't live in a vacuum, and this is where the second half of the paradox kicks in. More code means more integration debt, more security surface area, more edge cases, and a heavier maintenance burden. AI generates the initial blocks fast; it does not absorb the systemic complexity those blocks create. That complexity is exactly what senior engineering judgment exists to manage.
The Jevons feedback loop: AI lowers cost per unit of code → backlogged ideas cross the worth-building threshold → more software gets built → complexity compounds (integration debt, security surface, maintenance load) → demand for engineering judgment rises → more engineering hires → loop repeats.
Breaking Down the 46% → 55% Shift
Why did engineering's share of hiring accelerate during peak AI adoption (2023–2025)? Because the task profile shifted, the headcount needed did not.
Time spent on rote syntax writing has plummeted — Octoverse data shows 80% of new GitHub developers use Copilot in their first week. But demand for system architecture, data engineering, AI orchestration, and security verification has grown in its place. Companies aren't hiring less. They're hiring differently: fewer people to type code, more people to decide what code should exist, verify what AI produced, and keep the growing system coherent.
The macro forecast agrees. The U.S. Bureau of Labor Statistics projects software developer employment to grow 15% from 2024 to 2034 — "much faster than average" against 4% for all occupations — with roughly 129,200 openings per year. BLS explicitly names continued AI development as a growth driver, not a headwind.
Takeaway: the 46% → 55% shift is not engineering surviving AI. It's engineering absorbing the demand AI created.
What the Displacement Narrative Gets Wrong
The displacement story wasn't baseless-it just misread three things:
- It conflated pandemic-correction layoffs with AI displacement. The 2022–2024 layoff waves were largely a reversal of pandemic-era overhiring. Overall, Tech Major hiring is down 25% from 2019, but engineering is the function that contracted the least (−11%). If AI were displacing engineers, engineering would have contracted most.
- It assumed the demand for software is fixed. The whole displacement argument rests on a fixed pile of work: if AI does 30% of it, 30% of the people go. Jevons showed in 1865 why that assumption fails for any elastic resource. Software demand is not a fixed pile — it's a backlog that reprices.
- It ignored the skill distribution of the impact. The friction is real but concentrated: entry-level, rote-coding roles face genuine pressure, and SignalFire's own earlier data showed entry-level hiring declining sharply. Meanwhile, the premium on system-level engineers with domain judgment has never been higher. "AI replaces engineers" and "AI restructures the engineering ladder" are very different claims — the data supports the second.
FAQ
Will AI replace software engineers?
No, current hiring data shows the opposite. Engineers rose from 46% of Big Tech new hires in 2019 to 55% in 2025 (SignalFire, via TechCrunch), and the BLS projects 15% growth in software developer employment through 2034. AI is changing what engineers do, not eliminating the role.
What is the Jevons paradox in AI?
The Jevons paradox in AI is the pattern in which AI makes software development more efficient, lowering its cost and increasing total demand for software and, therefore, for engineers. It mirrors Jevons' 1865 finding that efficient steam engines increased total coal consumption.
Which AI engineering jobs are growing fastest?
Roles centered on judgment over syntax: system architecture, data engineering, AI orchestration, and security verification. These are the functions absorbing the complexity created by AI-accelerated code production.
Are entry-level engineering jobs safe?
Entry-level roles face the most real pressure. Rote-coding tasks are the most automatable, and early-career hiring has contracted. The path increasingly requires demonstrating system-level judgment earlier, the skill AI can't yet substitute.
Conclusion: AI Is a Lever, Not a Replacement
AI acts as a lever on engineering output, and levers increase what gets attempted. The binding constraint on software is no longer how fast anyone can type code: it's how much complex software an organization can safely architect, govern, and maintain. That constraint is made of engineering judgment, and the hiring data shows companies buying more of it, not less.
Related reading: How to Upskill Your Engineering Team for AI in 2026: A 6-Step Playbook
A Methodology Note: Hiring figures come from SignalFire's State of Tech Talent Report 2026 (based on tracking of 650M+ employees and 80M+ organizations), as reported by TechCrunch on June 24, 2026. Code-volume figures come from GitHub's Octoverse 2025 report (platform-wide activity data, Aug 2024–Aug 2025). Employment projections come from the U.S. BLS Occupational Outlook Handbook, 2024–2034 cycle. No figures in this piece are estimated or extrapolated by the author.
