Every major technology cycle produces the same misreading. Observers measure what a new tool eliminates and conclude that less work remains. They are almost always wrong. Efficiency does not destroy demand. It relocates it, scales it, and frequently amplifies it beyond what the prior system could have absorbed. Artificial intelligence is not an exception to this pattern. It is its clearest expression yet.
The argument for displacement rests on a static model of demand. If a task previously required ten hours of human labor and AI completes it in one, the assumption is that nine hours of economic activity have been removed from the ledger. This is accountant reasoning applied to a dynamic system. It ignores that the cost reduction changes the calculus of who commissions work, at what volume, and for what purposes.
Jevons in the Data Center
William Stanley Jevons observed in 1865 that improvements in steam engine efficiency did not reduce coal consumption in Britain. They increased it. Cheaper energy per unit of output meant that more outputs became economically viable. The efficiency gain expanded the market rather than capping it. The same logic applies with uncomfortable precision to AI.
When language models reduced the cost of producing a first-draft document from several hours of professional time to minutes, the prediction was that fewer writers would be needed. The operational reality has been more complex. Content production has accelerated across every sector that adopted the tools. Legal teams that once produced fifty contract summaries a quarter now produce five hundred. Marketing departments that maintained three active campaigns now maintain thirty. The unit cost fell; the unit volume rose faster than the cost fell.
This is not incidental. It reflects a fundamental property of markets: suppressed demand exists at every price point above the clearing price. AI is a price-reduction mechanism operating across an enormous range of cognitive tasks simultaneously. When prices fall, latent demand surfaces.
New Categories, Not Just More of the Same
The efficiency-generates-demand dynamic operates through two channels. The first is volume expansion in existing categories, as described above. The second is category creation, which is both harder to measure and more consequential over time.
Tasks that were previously uneconomical to perform at all become routine once AI reduces their cost below the threshold of practical consideration. Personalized medical summaries for every patient encounter. Real-time translation of every support interaction. Automated monitoring of every regulatory filing. None of these were standard practice not because organizations lacked interest, but because they lacked the budget to staff for them. AI does not merely make existing work cheaper. It makes formerly impossible work possible.
The history of general-purpose technologies supports this framing. Electrification did not simply make factories faster. It restructured which products could be manufactured, where plants could be located, and what working hours looked like. The internet did not simply accelerate communication. It created categories of commerce, media, and social coordination that had no prior analog. AI is a general-purpose technology. Its category-creation effects will dwarf its labor-substitution effects over any horizon longer than a business cycle.
Where the Pressure Actually Falls
This does not mean displacement is fictional. It means displacement is narrower and more specific than aggregate claims suggest. The workers most exposed are those performing high-volume, low-judgment execution of defined tasks within large organizations that have both the capital to adopt AI and the scale to realize efficiency gains quickly. The pressure is real at that intersection. It is not evenly distributed across the labor market.
Meanwhile, demand is expanding for the skills that sit adjacent to AI output: judgment, verification, contextual interpretation, client-facing translation of machine-generated work into accountable decisions. These are not new skills. They are old skills that become more valuable, not less, as the volume of AI-generated material requiring human sign-off increases. A lawyer who reviews AI-produced contracts is not being displaced by the AI. She is being multiplied by it, assuming she can maintain quality control across a higher volume of work than was previously possible.
The organizations that will mismanage this transition are those that treat headcount reduction as the primary metric of AI success. They will extract efficiency gains in the short term and forfeit competitive position to rivals who reinvest those gains into higher output volume and new capability development. This is what it means to optimize for the wrong variable.
Infrastructure as the Tell
The best empirical evidence for the efficiency-generates-demand thesis is not in labor statistics. It is in capital expenditure. Global investment in AI infrastructure — data centers, power generation, networking, chip fabrication — has accelerated through every period of proclaimed AI saturation. Companies spending tens of billions of dollars on capacity are not anticipating a world in which AI reduces overall activity. They are anticipating a world in which AI-enabled activity expands faster than current infrastructure can support.
Hyperscalers do not build for displacement cycles. They build for demand cycles. The infrastructure buildout is a forward-looking demand signal, and it is pointing in one direction.
The Frame That Holds
Efficiency and abundance are not opposites. In mature technology transitions, efficiency is the mechanism by which abundance is produced. AI is making cognitive work cheaper. Cheaper cognitive work means more cognitive work gets done, across more domains, by more actors, than the prior cost structure permitted. The displacement narrative mistakes a transition in the composition of demand for a reduction in demand itself.
The Jevons paradox has been correct about every major efficiency technology in the industrial era. There is no principled reason to expect AI to be the first exception. Efficiency generates demand. The question worth asking is not whether this will happen, but who will be positioned to capture the resulting expansion.