Most predictions about AI and employment run along a simple line: AI gets cheaper, so companies need fewer people, so jobs disappear. It sounds obvious. It is also at odds with how cheaper technology has usually behaved. There is an economic principle, Jevons Paradox, that explains why falling AI prices have so far driven more demand, more spending and more work rather than less. It is worth understanding, because it changes what falling costs actually mean for a business.
The core idea is counterintuitive. When something becomes cheaper and more efficient to use, total consumption of it tends to rise rather than fall, because the lower price opens up uses that were never worth it before.
What Jevons Paradox actually is
In 1865, the English economist William Stanley Jevons published The Coal Question and made an argument that surprised people. As steam engines became more fuel-efficient, Britain did not burn less coal. It burned far more. Cheaper, more efficient engines made coal-powered work viable in more industries, those industries adopted steam, and total coal use climbed even as each engine used less per unit of output. Jevons put it bluntly: "It is a confusion of ideas to suppose that the economical use of fuel is equivalent to diminished consumption. The very contrary is the truth." (Jevons paradox, overview)
The pattern has recurred ever since. More fuel-efficient cars led to more driving. Cheaper air travel led to more flights. More efficient lighting led to more light, not less electricity spent on it. Cheaper computing turned software into a multi-trillion-dollar industry and created tens of millions of jobs. In each case, efficiency widened the range of things worth doing, and demand grew to fill it.
How far AI costs have fallen
To see why the same logic applies to AI, start with the price of running a model. The fall has been steep. For a given level of quality, the cost of inference has been dropping by roughly ten times a year, and for models at GPT-4's quality level the price has come down about sixty-fold since GPT-4 arrived in March 2023, according to the venture firm a16z's "LLMflation" analysis. In rough terms, a task that cost tens of dollars per million output tokens at the start of 2023 cost on the order of a dollar two years later.
This is not a gentle decline. It is a collapse, and it has not stopped. Better hardware, software optimisations such as quantisation and batching, and hard competition between providers keep pushing prices down. Open-weight models push them lower still for teams willing to host their own.
The obvious conclusion would be that AI is getting cheaper to run, so total AI spending should fall. The opposite is happening.
The demand that follows falling prices
When the price of running a model drops by that much, companies do not spend proportionally less. They build far more. Worldwide spending on AI infrastructure roughly doubled in a single year, from about 153 billion dollars in 2024 to 318 billion in 2025, and is forecast to pass a trillion dollars by 2029, according to IDC. Those are not the spending patterns of an industry whose demand is shrinking.
The reason is that each price drop opens new categories of use. When inference was expensive, it made sense only for high-value tasks: complex document analysis, critical customer interactions, demanding coding work. You would be selective. Once it is cheap, the calculation changes. It becomes reasonable to run AI over every support ticket rather than only escalated ones, to draft every internal document rather than only the external ones, to build tools for every department rather than only the technical team, and to process every record in a database rather than only the flagged ones. The cheaper it gets, the more places it goes, and every new place needs people to design, deploy and maintain it.
Why this tends to create work, not remove it
The "AI takes jobs" view assumes a fixed pile of tasks being fought over. Jevons Paradox suggests that is the wrong frame, because the pile itself grows.
New roles appear to meet new demand
As AI became cheap enough to put inside products, companies needed people who understood what it could and could not do. As agents became workable for real workflows, demand for people to design those workflows followed. The World Economic Forum's Future of Jobs Report 2025 projects that structural change, including AI and automation, will create around 170 million roles and displace about 92 million by 2030, a net gain of roughly 78 million. Roles such as AI product managers, workflow designers and operations engineers barely existed a few years ago and now appear in large numbers of job adverts.
Cheaper AI grows the market, it does not only shift tasks
Think about spreadsheets. When VisiCalc and then Excel made calculation quick and cheap, firms did not sack their accountants. They asked them to analyse more, model more scenarios and take on broader planning. The role grew. The same is happening with AI. A lawyer who can draft a contract in minutes does not stop billing; they take on more matters and serve smaller clients who could not previously afford them. A marketer who can produce ten variants does not get cut; they run more experiments and enter more markets. Productivity gains do not only help the company using them. They make goods and services cheaper, which expands the market itself.
The infrastructure behind AI is its own employer
Building the capacity to deliver cheap inference is a large employer in its own right. The data-centre boom has created real demand for electricians, construction workers, civil engineers and facilities staff, and the rush to make AI chips has pulled in billions of investment in new fabrication. This indirect effect is easy to miss: the resource whose cost is falling, AI compute, needs an enormous physical base to deliver it.
Where the inflation shows up
Jevons Paradox also explains something that caught economists out. In parts of the labour market, AI is pushing wages up rather than down. When demand for AI-capable staff runs ahead of supply, pay rises, which is what has happened for engineers with machine-learning experience, for data scientists and for people who can lead AI work. Those salaries have held up or climbed even while other technology roles were cut. The compute market tells a similar story: per-unit inference is cheaper, yet total demand for graphics processors has driven prices and waiting times for cloud capacity upward. Efficiency in one part of the system creates pressure in the parts next to it.
Industries being reshaped now
The effect is uneven. Some sectors show it more plainly than others.
In software, coding assistants make individual developers faster, but the result has not been a smaller workforce. It has been more software: more applications, more features, faster iteration, and projects that are now worth building for smaller markets. In legal and financial services, AI document review has cut the hours that routine work takes, yet firms have largely responded by handling more matters and moving into markets that were previously too expensive to serve. In customer support, AI handles more routine queries, which has shifted human effort towards complex cases, escalations and relationships. In content and media, cheaper production has led to more content across more formats and channels, and continued demand for editors and directors who can point AI output at something useful.
The counterargument worth taking seriously
None of this means AI never reduces employment in a given role. Jevons Paradox is about aggregate demand, not individual outcomes. Some categories will genuinely shrink, particularly high-volume, low-judgement data processing, routine template-based document work, and simple pattern-matching in tasks such as moderation and classification.
So the question is not whether AI displaces any work, because it clearly does. It is whether the net effect is destruction or transformation. Historically, Jevons dynamics have tended towards net expansion even as the nature of the work changes. The transition is still real for the people caught in it. The jobs created often need different skills from the ones removed, and those distributional effects matter even when the overall picture is positive.
What it means for strategy
If you accept that Jevons Paradox applies to AI, a few practical points follow.
Do not budget for AI as a cost-cutting exercise alone. The firms that get the most from it tend to treat it as a way to do more things, not the same things for less. Savings are real, but they are usually smaller than the value of what becomes newly possible.
Expect your total AI spend to rise, not fall. Even as the price per query drops, using AI seriously tends to increase the bill, because you keep finding more worth doing. That is a sign of value found, not waste.
Invest in the people who can direct these systems. The scarce resource in an AI-rich environment is not the AI. It is the judgement, domain knowledge and design sense that turn raw output into something useful, and that becomes more valuable as the tools spread, not less.
Prepare for faster cycles. When a workflow can be built and changed quickly, the advantage moves to teams that can experiment, measure and improve faster than their competitors.
Common questions
What is Jevons Paradox in plain terms?
It is the observation that when a technology becomes cheaper or more efficient to use, total consumption of it tends to rise rather than fall. The lower cost per unit makes the technology worth using in more situations, which lifts total demand. It is named after William Stanley Jevons, who saw the effect in steam-engine efficiency and coal use in the 1860s.
Does it mean AI will never cut any jobs?
No. It describes the aggregate, not individual roles. Specific tasks and jobs will be automated and reduced. The argument is that, across the whole economy, the expansion of AI uses tends to create more activity and work than it removes, while changing how that work is distributed. People in highly routine roles face real risk; people who can direct and improve AI systems tend to see their work expand.
Why is AI spending rising if models are getting cheaper?
Because cheaper inference makes more applications worth building. When the cost falls sharply, companies do not bank the saving; they pursue uses that were not worth it at the old price. Lower marginal cost leads to more use cases, which leads to higher total spend. The pattern is visible in the infrastructure investment figures above.
Is this the same as the rebound effect?
They are closely related. The rebound effect is the general term for efficiency gains being partly or wholly offset by higher consumption. Jevons Paradox is the strong form, where the rebound is more than complete, so total consumption rises rather than just partly recovering.
Which AI-related jobs are growing?
Demand is strong for AI product managers, people who tune and adapt models, workflow designers, machine-learning operations engineers, and AI governance and safety specialists. Roles in the physical supply chain, data-centre operations and chip work, are growing too. Most of these reward a mix of domain knowledge and AI fluency rather than deep research skills.
Will AI-driven wage and compute pressure ease?
Probably, in time. Today's pressure reflects demand outrunning the supply of skilled people and compute. As training routes widen, hardware production scales and open-weight models reduce pricing power, the pressure should ease, though the lag means it is likely to persist for a few more years.
Key takeaways
- Jevons Paradox predicts that cheaper AI drives more demand for AI, not less, and the spending data fits.
- For a fixed level of quality, inference costs have fallen by roughly ten times a year, while total AI spending has risen sharply.
- Cheaper AI makes more uses worth pursuing, which expands the overall market rather than just redistributing existing work.
- New categories of AI-related jobs are growing faster than simple automation fears assume.
- The distributional effects are real, since some roles and tasks will be automated, but the net employment effect looks expansive rather than contractive.
- The infrastructure behind AI, from data centres to chips to power, is generating significant employment beyond the technology sector itself.