Jevons Paradox and AI: Why Efficiency Keeps Making Compute More Expensive

01 A Pattern Older Than the Transistor

In 1865 the economist William Stanley Jevons studied English coal consumption and found something that still unsettles planners today. As steam engines became more efficient at burning coal, England did not use less coal. It used dramatically more, because every improvement in efficiency lowered the effective cost of doing work, and cheaper work found new uses faster than the old uses shrank. The pattern has since been observed in lighting, in shipping, in computing, and now, with unusual force, in artificial intelligence. The Jevons paradox is not a law of physics. It is a statement about human response to falling costs, and in 2026 it has become the single most important framework for understanding why AI keeps getting more expensive to run even as each individual computation gets cheaper.

The paradox matters now because the industry has spent three years delivering exactly the kind of efficiency gains that trigger it. Newer accelerator generations deliver multiples of the compute-per-dollar of their predecessors, model compression and distillation slash the cost of serving a given capability, and inference pricing from the major labs has fallen repeatedly. By the textbook logic, the compute bill should be shrinking. Instead, the buildout of data centers, substations, and transmission capacity is the largest infrastructure program the private sector has ever undertaken.

02 Why Inference Consumes Its Own Savings

The mechanism is demand elasticity. When the cost of generating a token falls by an order of magnitude, developers do not bank the savings and run the same workload. They expand the workload until the marginal value of the next token roughly equals the new, lower marginal cost. A customer-support assistant that summarized tickets becomes an agent that reads every ticket, drafts every response, verifies every order status, and loops through tools a dozen times per interaction. The per-query cost collapsed. The number of queries per customer interaction rose faster.

Two structural features of AI workloads amplify the effect beyond what ordinary software exhibited. First, model quality is compounding: a cheaper token from a stronger model unlocks use cases that were economically impossible at the old price, so each price drop does not just expand existing demand, it creates new categories. Second, the agentic pattern multiplies consumption per task by factors of ten to a hundred compared with a single prompt-and-response, and agents only became practical once per-token costs crossed a threshold. The efficiency gains did not merely accommodate the agents. They caused the agents.

03 The Buildout in Numbers

The scale of current spending has no clean precedent in the technology industry. The largest cloud operators collectively reported capital expenditures in the low hundreds of billions of dollars across 2025, and public guidance for 2026 points substantially higher, with the incremental spend overwhelmingly directed at AI-ready data centers rather than general-purpose cloud capacity. Individual campuses are being announced in the gigawatt class, which means each site draws power on the scale of a mid-sized city.

04 Where the Grid Meets the Limit

The binding constraint has shifted from chips to electricity. Interconnection queues for new data-center loads stretch for years in several major markets, utilities are proposing dedicated tariffs for AI customers, and the share of national electricity consumption attributable to data centers has been revised upward repeatedly. In the United States, data centers represented a small but stable slice of demand for two decades, roughly one to two percent through the late 2010s. Industry analyses now place the share around four percent and climbing, with projections that approach high single digits by the end of the decade if current buildout plans are realized.

05 The Saturation Counterargument

The Jevons frame has critics, and their argument deserves a fair statement. Every demand curve eventually saturates. There are finitely many support tickets, finitely many documents to summarize, and finitely many hours of human attention for AI output to complement or replace. If model-level efficiency continues improving faster than usage grows, total compute demand could plateau even while capability rises. Distillation, in particular, points in this direction: the frontier model does the expensive training run once, and the resulting capability is then compressed into small models served at trivial cost. Some analysts argue the endgame is precisely that, an expensive frontier and a cheap long tail, with aggregate demand far below what current buildout implies.

The historical record pushes back. Lighting, computing, and communications all passed through phases where informed observers predicted saturation, and in each case new use categories absorbed the efficiency gains. Whether AI is different depends on an empirical question that is still open: whether the value of machine intelligence has a natural ceiling below the cost curve. Nothing in the 2026 demand data has yet suggested that it does.

06 What Breakthrough Claims Do to the Curve

Every week of 2026 has delivered another claimed breakthrough, and each one lands on an already strained system as a demand event rather than a cost saving. The pattern is visible in the news cycle itself: this article's source video is a briefing on one such claimed breakthrough, and whatever the ultimate verdict on that system turns out to be, the announcement measurably increased the amount of compute the market expects to buy. A genuine capability jump expands the addressable workload. A claimed jump expands expected demand almost as reliably, because planners cannot afford to dismiss it.

This is the reflexive loop that makes the current cycle historically unusual. Efficiency gains recruit new demand directly, and the announcements of progress recruit capital before the demand even materializes. The result is infrastructure built ahead of validated workloads, which is either prudent preparation or overbuild depending on demand elasticity that nobody can measure in advance.

07 The Practical Verdict

For operators and investors, the actionable reading is that efficiency improvements are a demand signal, not a cost saving. A lab announcing that its newest model serves the same quality at a tenth of the cost has announced, in the same breath, an expectation of tenfold usage growth or more. Watch tokens consumed per user per day, interconnection queue lengths, and utility tariff filings; those three numbers describe the real economy of AI better than any benchmark. For everyone else, the paradox explains a genuine puzzle: why the technology in your pocket keeps getting cheaper per task while the industry building it keeps raising ever larger sums. Both facts are the same fact. The cheaper it gets, the more of it the world buys.