Jevons Paradox Came for Software: Why AI Made Engineers More Valuable, Not Less

For three years the dominant narrative in tech has been substitution: AI writes code, engineers lose jobs, the math is simple. The layoff headlines, the campus recruiting freezes, the Y Combinator founders bragging about shipping with zero engineers — all of it pointed in one direction. What the aggregate hiring data says is more complicated, and more interesting. Engineering was the most resilient job function of 2025. The paradox has a name, and it came for software right on schedule.

The Fear Is Understandable — and Some of the Data Supports It

Let's acknowledge the fear honestly first, because dismissing it outright would be dishonest. This blog has written about the missing rung in junior developer hiring and the structural disruption facing new graduates. The same dynamics that show up in the optimistic aggregate data also produced a separate, grimmer statistic: CS grads now face higher unemployment than fine arts majors, a figure that would have seemed absurd five years ago. Recruitment freezes at large firms are real. Campus hiring programs that absorbed entire graduating cohorts have been quietly wound down or eliminated. The era of hiring junior engineers to brute-force code volume is over, and a generation that structured its career expectations around that model is absorbing the impact.

But the aggregate hiring data — the totality of what happened across the industry in 2025, not just the visible layoffs — tells a more complicated story. And when you run that story through the lens of economic history, the shape of it becomes recognizable. This is not the first time a productivity technology was expected to destroy the profession it made more productive. It happened with steam power. It happened with spreadsheets. It happened with CAD tools in architecture and engineering. The pattern has a name: the Jevons paradox, and it landed in software on something close to schedule.

What the Aggregate Hiring Data Actually Shows

SignalFire's 2026 State of Talent report is one of the few data sources that cuts across hundreds of companies and tracks what actually happened at the hiring level in 2025 — not survey responses about intent, but real hiring flows. The finding that stands out is the resilience of engineering relative to every other function. In a year when large tech companies collectively cut total headcount by 25%, engineering hiring fell only 11%. More strikingly, engineers still made up 55% of all new hires across the "Tech Majors" category, a share that increased from the prior year even as the overall headcount shrank.

The Bureau of Labor Statistics' 17% growth projection for software developers covers the decade through 2033 and was published after AI coding tools were already widely deployed and adopted across the industry. The BLS methodology accounts for technology substitution; it does not project growth for professions it expects to shrink. The fact that the projection held — and in some revisions increased — is not an accident. The economists modeling the long-run labor market are not seeing the same story the tech layoff discourse is generating.

None of this means individual engineers face no risk. It means the risk is not uniformly distributed across seniority levels and skill sets, and the direction of aggregate demand is not what the substitution narrative claims. The question worth asking is why the demand held, and the answer is a 170-year-old economic principle that most engineers have never heard of — but that economists have watched play out across every major productivity technology in history.

Jevons Paradox: Why Cheaper Production Expands Demand

In 1865, the British economist William Stanley Jevons observed something counterintuitive about coal. The invention of the steam engine — which made coal consumption dramatically more efficient per unit of work — did not reduce total coal use. It increased it. Because each unit of energy was now cheaper to extract and deploy, people found more uses for energy than they had before. Total consumption rose even as efficiency per unit fell. Jevons called this the paradox at the heart of efficiency gains: the more productive a resource becomes, the more consumption of that resource grows to fill the space the productivity opened.

Apply this directly to software engineering. AI coding tools made writing code dramatically cheaper in 2024 and 2025. If the substitution model were correct, companies would need fewer engineers to produce the same amount of software, and total demand for engineers would fall. But the Jevons model predicts something different: companies will produce more software — because the marginal cost of a feature fell — and will therefore need more engineers to design, specify, evaluate, and own a larger portfolio of software than they had before. The hiring data from SignalFire is consistent with the Jevons model. It is not consistent with the simple substitution model.

The expansion of demand is visible in adjacent indicators too. Enterprise software backlogs — the internal lists of tools companies want to build but have not yet funded — have lengthened as AI coding became available, not shortened. More boards now believe it is feasible to automate more internal processes, which means more software projects get greenlit than before. Firms with AI-native development workflows are not doing the same work with fewer people; they are attempting more ambitious software than they ever tried before, and they need engineers to architect and own the outcomes. That is the Jevons dynamic operating exactly as Jevons described it.

The Role Is Transforming: What Engineers Actually Do Now

BCG's analysis of the engineering labor market through 2025 makes an important distinction that the substitution narrative misses: the skills being displaced are not the same as the profession being displaced. What AI is eliminating is the most rote, most repetitive, highest-volume component of the job — writing boilerplate, translating a specification into an obvious implementation, searching documentation for an API signature. What it has not displaced, and what it has arguably made more valuable, is the judgment layer: system architecture, performance and cost tradeoffs, security design, and the translation of ambiguous business needs into precise technical requirements that a system can actually satisfy.

The analogy to the architect is instructive. The invention of computer-aided design did not reduce demand for architects — it expanded what architects could design and how fast they could design it. A firm with CAD could take on more projects, attempt more complex structures, and iterate faster with clients. The core skill of spatial judgment, structural understanding, and client translation remained entirely human. CAD removed the drafting table as a bottleneck — it did not eliminate the need for someone who understood why a building should be designed a particular way. AI coding tools are doing the same thing to software engineering: removing the keystroke as the bottleneck, leaving the architectural judgment fully in the hands of the engineer who can exercise it.

The New Titles Being Created Around the New Work

If the profession were genuinely contracting, you would expect job title diversity to shrink — fewer variants, fewer specializations, a market converging on a smaller and simpler definition of what an engineer does. The opposite is happening. Stack Overflow's 2026 Developer Survey tracked new job title categories that barely existed in 2024, and the fastest-growing of them is an explicit acknowledgment that a new kind of engineering work has appeared and is being actively hired for.

"AI Integration Engineer" grew 156% year-over-year in postings tracked by Stack Overflow's 2026 survey — the fastest-growing developer title in the dataset. The role sits at the intersection of engineering and applied AI: wiring models into production systems, designing evaluation harnesses, building the orchestration layers that turn a capable model into a reliable workflow. Context engineering and agent orchestration roles are growing on a similar trajectory. These are not replacements for the old job description; they are what the old job expands into when the old bottleneck — writing code manually, one function at a time — gets automated away.

The engineers landing those roles are the same engineers who, five years ago, might have spent their days in JIRA tickets writing CRUD endpoints. The skill that made them valuable — understanding how systems fit together, how data flows, where things break under load — is the same skill. The surface on which that skill gets applied has changed. The 10-agent engineer is the 10x engineer's successor: the leverage has multiplied, not disappeared. The engineers who understand that early are the ones collecting the raises that show up in the senior-engineer salary data while the junior entry-level pipeline remains disrupted.

The Honest Caveat: The Missing Rung Is Real and Structural

The Jevons story does not hold evenly across seniority levels, and any analysis that pretends otherwise is doing the junior developer cohort a disservice. A Stanford study published in 2026 found that employment for software developers aged 22–25 fell nearly 20% from 2022 to 2025. IEEE Spectrum reported that the graduate-hiring outlook for CS majors was the most pessimistic since 2020. The entry-level pipeline that relied on new graduates absorbing volume work while they learned — the apprenticeship model that was never formally called an apprenticeship — has been disrupted at exactly the point where AI can do that volume work instead.

The missing rung is not a temporary dip that will self-correct when the business cycle turns. It is a structural problem: the on-ramp to senior engineering that ran through years of volume-coding work has been shortened, and it is not obvious what replaces it. Companies that once hired five junior developers to ship a feature now hire one senior developer and give them an AI coding assistant. The junior head count that funded the learning-while-doing model is gone. This is the legitimate critique of unconditional Jevons optimism — the expansion of aggregate demand is real, but the distribution of that demand has shifted sharply toward experience and judgment, leaving new entrants without the path to acquire either.

The honest answer to a junior developer in 2026 is not "relax, Jevons will save you." It is: the path into the profession has changed, the tools you need to learn first have changed, and the skills that differentiate you from an AI coding assistant are different from the skills that differentiated your predecessors from each other. That is genuinely harder for a 22-year-old navigating the entry ramp for the first time, and it deserves acknowledgment rather than dismissal.

What Engineers Should Actually Do With This Information

The strategic implication of the Jevons model is not passivity. The profession is not safe because some economic principle protects it — it is safe for engineers who move toward the work that the paradox creates demand for. Concretely, that means the direction of investment is clear: agent orchestration, AI evaluation, system design for AI-first products, and the judgment skills that distinguish good AI output from plausible but wrong AI output. These are not speculative future skills; they are the skills the 156% growth in AI Integration Engineer postings is already paying for today.

For engineers mid-career, the transition is manageable and the window for it is open. The firms building AI-native systems in 2026 are actively looking for people who understand both sides of the stack — model behavior and production constraints — and that combination is rare enough to command a meaningful premium. The engineers who made themselves the connective tissue between AI capability and business outcome in 2025 are the ones collecting the raises that SignalFire and BCG are tracking in their data.

AI Deleted the Parts of the Job That Were Never the Point

The most clarifying frame for the Jevons story is also the most uncomfortable one: AI did not delete software engineering, it deleted the parts of software engineering that engineers were never really paid to do. Writing boilerplate CRUD endpoints was never what made a codebase better. Translating a Jira ticket into a function call was never the core skill the company was actually hiring for. These were necessary costs of turning thought into working software, and they absorbed a large fraction of engineering time because there was no alternative. Now there is an alternative, and the costs have shifted accordingly.

What remains after you remove the volume coding is what companies actually needed when they hired engineers: the ability to understand a complex problem domain, decompose it into a system that will behave correctly under adversarial conditions, make tradeoffs between performance and cost and maintainability, and take ownership of outcomes rather than outputs. Those skills did not become less valuable when AI got better at generating code. They became more visible, because now the engineer's leverage over the outcome is clearer — they are steering the system, not typing it into existence.

Jevons would recognize the pattern immediately. Every time a core production bottleneck has been mechanized in history — from steam power to spreadsheets to CAD — the humans involved did not disappear. They moved up the value chain to the judgment work that the machine could not replicate, and demand for that judgment expanded to fill the space the machine's productivity opened. Software engineering in 2026 is in exactly that transition. The engineers who move first are building the most interesting and ambitious systems in the world right now. The data shows they are also the ones still getting hired, still getting raises, and — if the Jevons paradox holds as firmly for software as it held for coal — the ones whose long-run demand will expand in proportion to how much cheaper the tools get.