In 1865, the English economist William Stanley Jevons published "The Coal Question," a book that disturbed the prevailing assumption of Victorian policymakers. Britain's coal reserves were finite. Steam engines were becoming more efficient. The intuition seemed straightforward: better engines would stretch coal supplies further, buying the empire time. Jevons demonstrated the opposite. Every improvement in steam engine efficiency — from Newcomen's 1% thermal efficiency to Watt's 4% to the compound engines approaching 12% — had been followed not by a reduction in coal consumption but by a dramatic expansion. Britain burned 16 million tons of coal in 1829. By 1865, it burned 83 million tons. The engines consumed less coal per unit of work. The economy consumed vastly more coal in total.
The mechanism is deceptively simple. When a resource becomes more efficient to use, the effective price of the service it provides drops. Cheaper energy per unit of work means more work gets done. Cheaper lighting per lumen means more lumens get deployed. Cheaper compute per operation means more operations get run. The demand increase overwhelms the per-unit savings. Total consumption rises, often by multiples of the original level. The paradox is not that efficiency fails — it succeeds spectacularly at its stated goal. The paradox is that success at the unit level produces the opposite of the expected outcome at the aggregate level.
This is not a theoretical curiosity. It is one of the most consistently observed dynamics in economic history, and it operates across every category of resource and technology.
The Bessemer process made steel 80% cheaper per ton between 1856 and 1890. Global steel production didn't fall by 80%. It rose from 500,000 tons to 28 million tons. Cheap steel enabled skyscrapers, transcontinental railroads, naval fleets, and mass-produced machinery — applications that were structurally impossible when steel cost ten times more. The green revolution increased crop yields per acre by 200–300% starting in the 1960s. Global agricultural land use didn't shrink proportionally — food became cheaper, populations grew, diets shifted toward resource-intensive meat, and total agricultural resource consumption expanded. LED lighting uses roughly 75% less electricity per lumen than incandescent bulbs. McKinsey estimated in 2012 that LED adoption would reduce global lighting electricity by 40%. By 2023, the actual reduction was closer to 15%, because the world deployed dramatically more lumens — illuminating buildings, stadiums, digital signage, and decorative installations that were economically unfeasible under incandescent pricing.
The bandwidth market tells the same story. The cost of transmitting one gigabyte of data fell from approximately $150 in 1998 to under $0.01 by 2023 — a reduction of over 99.99%. Global internet traffic didn't stay flat. It grew from roughly 100 petabytes per month in 2002 to over 400 exabytes per month by 2023 — a four-million-fold increase. Cheaper bandwidth created video streaming, cloud computing, social media, remote work infrastructure, and the entire mobile application economy. None of these existed at 1998 bandwidth prices. All of them are now among the largest sectors in the global economy.
The paradox operates most powerfully when three conditions converge. First, the resource must serve a function with elastic demand — meaning that a price reduction significantly increases the quantity demanded. Coal-powered mechanical work, automobile transportation, and cloud computing all meet this criterion. Second, the efficiency improvement must be large enough to materially lower the effective cost to the end user. A 5% improvement rarely triggers a Jevons effect. A 50% improvement often does. Third, the economy must be sufficiently flexible to absorb the increased demand — through new applications, new users, new industries, or geographic expansion. When all three conditions hold, the rebound effect exceeds 100% of the efficiency gain, and total resource consumption increases. Economists call this "backfire." Jevons simply called it obvious.
The error that Jevons identified has been repeated by policymakers and strategists in every century since. In the 1970s, the US government projected that improved automobile fuel efficiency would reduce national gasoline consumption. It didn't — Americans drove more. In the 2000s, analysts projected that server virtualization would reduce total data center energy consumption. It didn't — virtualization made compute so cheap that workloads proliferated. In the 2020s, AI companies project that more efficient inference will reduce the cost of deploying language models. It will reduce the per-query cost. It will not reduce total spending on inference. The pattern repeats because the underlying economics are invariant: when efficiency reduces the effective price of a service, and demand for that service is elastic, consumption expands.
The concept matters for founders because the technology industry is the most fertile environment for Jevons effects in modern economic history. Software reduces the marginal cost of distribution to near zero. Moore's Law reduces the cost of compute by roughly 50% every eighteen months. Cloud infrastructure makes server capacity a variable expense instead of a capital investment. Each of these efficiency gains has consistently produced not conservation but explosion — in usage, in spending, in the creation of entirely new markets that didn't exist before the efficiency improvement made them economically viable. Understanding Jevons Paradox doesn't just explain history. It predicts where demand will grow next.
Section 2
How to See It
Jevons Paradox leaves a distinctive signature: per-unit metrics improve while aggregate metrics move in the opposite direction from what those per-unit improvements would suggest. The gap between expected and actual aggregate consumption is the paradox at work. Look for the pattern anywhere a technology makes a resource meaningfully cheaper to use.
The signals are often hidden in plain sight because analysts report per-unit metrics and aggregate metrics separately without connecting them. A company announces that its new chip is 3x more energy-efficient — presented as a sustainability win. The same company's data center energy consumption grows 40% year over year — presented as a separate operational challenge. These are the same phenomenon. The efficiency and the consumption growth are causally linked, not independent data points.
Technology
You're seeing Jevons Paradox when the cost per unit of cloud compute drops by 90% over a decade and total cloud spending grows from $6 billion in 2008 to over $600 billion by 2024. AWS has reduced the price of its standard compute instances more than seventy times since 2006. Each reduction made it rational for more workloads to migrate from on-premises servers to the cloud, for startups to build applications that would have been prohibitively expensive before, and for enterprises to run analytics, machine learning, and batch processing at scales they never previously attempted. The per-instance price fell continuously. Total spend on instances rose continuously. Amazon, Microsoft, and Google collectively invested over $150 billion in data center infrastructure in 2024 alone — infrastructure driven not by waste but by demand that cheaper compute created.
Markets
You're seeing Jevons Paradox when fuel efficiency standards improve and total gasoline consumption stays flat or rises. The US Corporate Average Fuel Economy standard pushed average fleet fuel efficiency from 13 miles per gallon in 1975 to 25 mpg by 2020. In that period, total US vehicle miles traveled nearly tripled — from 1.03 trillion miles in 1975 to 3.2 trillion in 2019. More efficient cars made each mile cheaper to drive. Americans responded by driving more miles, buying larger vehicles, commuting from farther suburbs, and taking more road trips. Per-mile fuel consumption fell. Total fuel consumption held roughly steady at 140 billion gallons per year through the 2010s, despite the fleet being nearly twice as efficient.
Business
You're seeing Jevons Paradox when a SaaS company reduces the cost of its API calls by 70% and its total infrastructure bill increases. OpenAI reduced the cost of GPT-3.5 Turbo API calls by approximately 90% between March 2023 and early 2025. The price drop didn't reduce the company's compute costs. It triggered an explosion in API usage as developers integrated language models into applications — customer support bots, code assistants, content generators, search engines — that were uneconomical at the prior price point. Each price cut expanded the addressable market for inference, and the usage growth more than offset the per-call savings.
Investing
You're seeing Jevons Paradox when an industry's total spending on a resource increases precisely because that resource became cheaper per unit. Global spending on semiconductor chips grew from $200 billion in 2000 to $574 billion in 2022. The cost per transistor fell by a factor of roughly 10,000 over that period. Cheaper transistors didn't mean cheaper chip bills — they meant chips went into phones, cars, refrigerators, thermostats, watches, industrial sensors, and data centers. The number of devices containing semiconductors expanded from hundreds of millions to tens of billions. The total transistor count in use grew faster than the cost per transistor fell, producing a net increase in total semiconductor spending.
Section 3
How to Use It
Decision filter
"If we make this process significantly more efficient or cheaper, will the total demand for the underlying service expand enough to offset the per-unit savings? If the service has elastic demand and no hard ceiling on usage, assume total consumption will increase. Budget and plan for growth, not savings."
As a founder
Jevons Paradox reframes efficiency improvements from cost-saving measures into market-expansion strategies. When Jeff Bezos reduced AWS compute prices, he wasn't sacrificing margin — he was manufacturing demand. Each price cut made previously uneconomical use cases viable: machine learning experiments that would have cost $50,000 on dedicated hardware became $500 on spot instances. Startups that couldn't justify server infrastructure found they could launch products for under $100 per month. The volume increase more than compensated for the margin compression on each unit. Founders building infrastructure or platform products should model efficiency gains as demand accelerators. If your per-unit cost drops 80%, don't forecast an 80% reduction in customer bills. Forecast a 5x increase in customer usage. The customers who stay within their old budget are the minority. The customers who couldn't afford you before — and the use cases that didn't exist before — are the growth.
As an investor
The Jevons effect identifies where total addressable markets are being systematically underestimated. When analysts project the AI inference market, they typically model current use cases at declining per-query prices and arrive at modest growth figures. What they miss is the demand creation: at sufficiently low inference costs, language models get embedded into every software application, every search query, every customer interaction, every document workflow. The total number of inference operations doesn't grow linearly — it grows by orders of magnitude as each cost threshold unlocks a new tier of applications.
The investor's edge is recognizing that the denominator in a "cost per X" metric that is falling rapidly signals a numerator (total X) that is about to rise faster. Look for the same pattern that Jevons observed in coal: total spending on the resource increasing in the same period that per-unit cost is falling. That divergence — falling unit costs and rising total expenditure — is the financial signature of the paradox at work. It means the market is pricing in yesterday's demand while tomorrow's demand is already compounding. The semiconductor industry has exhibited this signature continuously since the 1960s. The cloud computing industry has exhibited it since 2006. The AI inference market is exhibiting it now.
As a decision-maker
Inside an organization, Jevons Paradox means that efficiency initiatives will generate demand, not just savings. When a company migrates to a more efficient cloud architecture, the initial cost reduction is real — for about one quarter. Then engineering teams discover they can run experiments they previously couldn't justify, data teams spin up analytics pipelines at the new marginal cost, and product teams launch features that require compute they'd been denied under the old cost structure. The infrastructure bill returns to its prior level within a year, then exceeds it.
This is not waste. It is rational behavior by teams responding to lower marginal costs. Every organization that has adopted FinOps practices to manage cloud spending has encountered this pattern: optimization reduces per-workload costs, which frees budget, which gets reallocated to new workloads, which consume the savings. The strategic response is to plan for it explicitly: set hard consumption budgets with governance if cost savings are the genuine goal, or embrace the demand creation and measure success by capability expansion rather than cost reduction. The worst outcome is presenting an efficiency initiative to the CFO as a cost-saving measure and then watching the bill grow. That outcome is not a failure of the efficiency initiative. It is Jevons Paradox operating exactly as predicted.
Common misapplication: Assuming Jevons Paradox applies to every efficiency improvement regardless of demand elasticity. The paradox requires elastic demand — meaning that a price reduction produces a proportionally larger increase in quantity demanded. For resources with inelastic demand, efficiency improvements do reduce total consumption. Per-capita water usage in most developed countries has declined as appliances became more efficient, because there is a natural ceiling on how much water a household uses regardless of price. The test: if making something 50% cheaper would not roughly double or more than double the total usage, the Jevons effect will be partial at best. The paradox is strongest where latent demand is vast and price is the primary constraint.
Second misapplication: Treating Jevons Paradox as an argument against efficiency improvements. This is a policy error, not a strategic one, but it corrupts analysis. The paradox describes what happens to total resource consumption. It does not mean efficiency improvements are pointless. A steam engine that produces the same output with half the coal generates twice the economic value per ton of coal — even if total coal consumption rises. The efficiency gain creates wealth. The consumption increase is a consequence of that wealth creation. Conflating the two leads to paralysis: refusing to improve efficiency because total consumption might rise. Jevons himself did not advocate against efficiency. He advocated for realistic planning about what efficiency would produce.
Third misapplication: Invoking Jevons Paradox when the demand increase is driven by factors unrelated to efficiency. Global energy consumption has risen continuously for two centuries. Not all of that increase is attributable to efficiency improvements — population growth, industrialization, rising living standards, and urbanization all contribute independently. The paradox specifically describes the causal channel from efficiency improvement to demand expansion. If consumption rises for reasons unrelated to a cost reduction, that is not a Jevons effect. The analytical discipline is to isolate the efficiency-driven demand response from the baseline growth that would have occurred regardless.
Section 4
The Mechanism
Section 5
Founders & Leaders in Action
Jevons Paradox is not merely an economist's observation about coal. It is the operating logic behind some of the most valuable companies ever built. The founders below didn't just benefit from the paradox accidentally — they recognized, implicitly or explicitly, that making something dramatically cheaper per unit would expand total demand far beyond what linear forecasting predicted. Each built business models calibrated to capture the demand that their own efficiency improvements created.
The common thread is a willingness to sacrifice per-unit margin in exchange for volume growth that more than compensates — and a conviction that the demand curve extends far below the current price point. Rockefeller cut kerosene prices to levels that competitors considered suicidal. Ford priced the Model T below what the existing automobile market suggested was rational. Bezos offered AWS compute at prices that made Wall Street question the company's profitability. Huang invested in GPU architectures for workloads that didn't yet exist. Altman reduced inference costs before the applications that would consume the cheaper inference had been built.
In each case, the founder bet on the Jevons mechanism: that drastically lower per-unit costs would create demand, not merely serve it. The analytical community modeled current demand at declining prices and projected moderate markets. The founders modeled latent demand at dramatically lower prices and projected massive markets. The founders were right. Not because they were optimists, but because the historical evidence for elastic-demand resources consistently supports the Jevons outcome over the conservation outcome.
Rockefeller's strategic obsession was reducing the cost of refining kerosene. When he entered the oil business in 1863, kerosene cost roughly $0.58 per gallon at wholesale. By the mid-1880s, Standard Oil had driven the price below $0.08 — a reduction of over 85%. The efficiency came from every link in the chain: better refining processes that extracted more kerosene per barrel of crude, consolidated transportation that reduced shipping costs by 60%, and barrel manufacturing at scale that cut packaging costs in half.
The intuitive prediction would be that cheaper kerosene meant less spending on kerosene. The actual outcome was the opposite. At $0.58, kerosene was a luxury — rural American families rationed their lighting, burning candles or whale oil instead. At $0.08, kerosene became universal. Households that had gone dark after sunset lit their homes for hours. Factories added evening shifts. Kerosene-powered machinery appeared on farms. The product went from a luxury good to an essential commodity, and total consumption expanded by orders of magnitude. Standard Oil's annual kerosene output grew from 4.8 million barrels in 1872 to over 25 million barrels by 1890.
Rockefeller understood this dynamic. His letters and memos repeatedly emphasized that price reductions would "enlarge the market" more than enough to compensate for the margin compression. He priced aggressively not despite the volume consequences but because of them. Lower cost per gallon produced a larger market, which produced more volume, which enabled further cost reductions. The flywheel was Jevons Paradox operating as a business strategy.
The irony is that Rockefeller's efficiency improvements in kerosene refining also contributed to the conditions that made petroleum the dominant energy source of the twentieth century. By making petroleum products cheap and ubiquitous, Standard Oil created the infrastructure — refineries, distribution networks, consumer habits — that the automobile industry later inherited. Ford's cars ran on gasoline refined by processes Rockefeller had perfected. The Jevons effect didn't just increase kerosene consumption. It built the foundation for a century of petroleum-based transportation.
Ford's Model T is the canonical industrial case of Jevons Paradox applied to a manufactured product. The car's price fell from $850 in 1908 to $260 in 1925 — a 69% reduction driven by assembly line efficiencies, vertical integration, and relentless process optimization. The resource in question was personal transportation, and its effective cost per mile dropped proportionally.
The conventional expectation might have been that Americans would drive roughly the same amount and spend less on transportation. The reality: registered motor vehicles in the US grew from 458,000 in 1910 to 20 million by 1925. Total gasoline consumption grew from effectively zero to 15 billion gallons per year. Total spending on roads, bridges, fuel, tires, insurance, and vehicle maintenance created an entirely new sector of the economy that by the 1920s represented over 10% of GDP. Ford's efficiency in automobile production didn't conserve transportation resources. It triggered the largest expansion of personal mobility in human history and created resource consumption on a scale that horse-drawn transportation could never have produced.
Ford grasped this intuitively. His entire pricing philosophy — reduce the price first, expand the market second, let volume cover the fixed costs third — was a practical application of the Jevons mechanism. He described it in "My Life and Work" (1922): the goal was to make the car so cheap that every working family could own one. The efficiency gain was the instrument. The demand explosion was the objective.
The secondary effects compounded the paradox. Cheap cars required roads. The Federal Aid Highway Act of 1916 and its successors poured billions into road construction, which made driving more convenient, which increased car purchases, which increased gasoline consumption, which funded more road construction. Suburbs — impossible without affordable automobiles — emerged as the dominant residential pattern in America by the 1950s, locking in transportation resource consumption at levels that would have been inconceivable when a car cost twelve months of a factory worker's wages. Ford's manufacturing efficiency didn't conserve transportation resources. It restructured American geography around the assumption of cheap, abundant personal mobility.
AWS is the purest modern expression of Jevons Paradox as deliberate corporate strategy. When Amazon launched Elastic Compute Cloud in 2006, renting a virtual server cost $0.10 per hour. By 2024, the equivalent compute power cost less than $0.01 per hour in many configurations — a 90-plus percent decline. Bezos reduced prices more than seventy times, not because margins were too high, but because he understood that each price reduction manufactured new demand.
At $0.10 per compute-hour, cloud was cost-effective for web hosting and basic applications. At $0.01, it became viable for machine learning training, genomics processing, real-time video transcoding, IoT data pipelines, and hundreds of other workloads that no one ran in the cloud in 2006 because the economics didn't justify it. The per-unit price fell by 90%. AWS revenue grew from $3.1 billion in 2013 to $90.8 billion in 2023 — a nearly 30x increase during a period of continuous price reductions. Total global cloud infrastructure spending exceeded $600 billion by 2024, up from effectively zero in 2005.
Bezos articulated the logic in his 2012 letter to shareholders: "We believe in customer centricity, not competitor centricity. Lower prices lead to more customers, more customers lead to more usage, more usage allows us to lower costs, lower costs allow us to lower prices." He was describing a flywheel, but the fuel for the flywheel was the Jevons effect — each efficiency improvement expanding total demand rather than reducing total spending.
The scale of the demand creation is visible in the infrastructure investment. In 2024, Amazon, Microsoft, and Google collectively spent over $150 billion on data center capital expenditure — infrastructure driven by demand that didn't exist when cloud compute cost ten times more per hour. Every dollar of that investment is a consequence of efficiency improvements that lowered the per-unit cost of compute enough to make millions of new workloads economically rational. Bezos didn't predict the specific workloads. He predicted the demand response to lower prices. Jevons could have told him the outcome in 1865.
Jensen HuangCo-founder and CEO, NVIDIA, 1993–present
NVIDIA's GPU architecture has improved computational efficiency by roughly 1,000x per dollar over the past decade for AI training workloads. Under naive forecasting, this should have meant that training an AI model in 2024 cost 0.1% of what it cost in 2014. Instead, total spending on AI training compute grew from negligible to an estimated $50–100 billion annually by 2024. The models got bigger faster than the chips got cheaper.
Huang recognized this early. In NVIDIA's investor communications starting around 2016, he began framing GPU efficiency improvements not as cost reducers but as capability enablers. Each generation of GPUs — Pascal, Volta, Ampere, Hopper — didn't just make existing workloads cheaper. It made new classes of models possible: larger language models, higher-resolution image generators, real-time inference at scale. The demand wasn't fixed. It expanded to absorb every unit of compute that efficiency gains made available, then demanded more.
By 2024, NVIDIA's data center revenue exceeded $47 billion — up from $3 billion just three years earlier. The customers weren't spending less because GPUs were more efficient. They were spending more because efficient GPUs made previously impossible AI applications viable. Huang's strategic positioning of NVIDIA as the platform for AI compute was a direct bet on the Jevons mechanism: that better hardware would create demand faster than it satisfied it.
Sam AltmanCo-founder and CEO, OpenAI, 2015–present
Altman has presided over the fastest cost deflation in the history of computing — and the fastest demand expansion. When OpenAI released the GPT-3 API in June 2020, pricing was approximately $0.06 per 1,000 tokens. By early 2025, GPT-4-class models were available at under $0.002 per 1,000 tokens through various providers, and GPT-3.5-equivalent capability cost a fraction of a cent. A 95-plus percent reduction in cost per unit of intelligence.
Under a conservation model, this should have produced a shrinking market. OpenAI would charge less per query, developers would use roughly the same number of queries, and the market would contract. Instead, total API usage grew by orders of magnitude. At $0.06 per call, developers built simple chatbots and text classifiers. At $0.002, they embedded inference into code editors, email clients, customer support platforms, legal document review tools, financial analysis pipelines, and educational software. Use cases that couldn't justify a $0.06 API call at volume became default features at $0.002.
Altman's public statements reflect an understanding of this dynamic. In a 2023 interview, he described the goal as making intelligence "too cheap to meter" — a conscious echo of Lewis Strauss's famous (and wrong) prediction about nuclear energy. The difference is that software intelligence has effectively unbounded demand. There is no natural saturation point for "how much thinking should software do?" Every reduction in inference cost opens a new tier of applications. OpenAI's investment in custom silicon, model distillation, and inference optimization is not about protecting margins. It is about manufacturing the next wave of demand by pushing the cost per unit of intelligence low enough to unlock the next tier of applications that don't yet exist.
Section 6
Visual Explanation
Jevons Paradox — How efficiency improvements increase total resource consumption by expanding demand beyond the per-unit savings
Section 7
Connected Models
Jevons Paradox gains its full explanatory power when placed alongside the frameworks that describe how efficiency, demand, and market structure interact. The paradox is not a standalone phenomenon — it is embedded in feedback loops, disruption dynamics, and compounding processes that amplify or constrain its effects. The strongest strategic analysis accounts for these interactions rather than treating efficiency and demand as independent variables.
The connections below reveal a consistent pattern: Jevons Paradox is the demand-side engine that powers many of the most studied phenomena in business strategy. It reinforces the models that describe self-sustaining growth. It creates tension with models that assume static demand. And it leads to the structural market transformations that define entire economic eras.
Understanding these connections prevents the most common analytical error when using the paradox: treating demand expansion as an isolated phenomenon rather than as the trigger for cascading strategic effects. When Jevons effects activate in a market, they don't just increase consumption — they create the conditions for flywheel dynamics, network effects, and creative destruction that reshape the competitive landscape. The founders who built the most durable advantages understood these cascading effects intuitively. Their strategies were calibrated not just for the initial demand expansion but for the secondary and tertiary effects that followed.
Reinforces
Economies of [Scale](/mental-models/scale)
Economies of scale and Jevons Paradox form a mutually reinforcing loop that has driven industrial expansion for two centuries. Scale economies reduce per-unit costs. Per Jevons, lower per-unit costs expand demand. Expanded demand increases production volume. Higher volume deepens the scale economies. Carnegie experienced this with steel: every efficiency improvement at the Edgar Thomson Works lowered the per-ton cost, which expanded the market for steel (railroads, bridges, construction), which increased Carnegie's production volume, which financed the next round of efficiency improvements. AWS operates an identical loop today. The reinforcement runs in both directions — Jevons creates the demand growth that feeds scale economies, and scale economies create the cost reductions that trigger Jevons effects. Companies positioned at the intersection of both dynamics capture compounding advantages that widen with each cycle.
Reinforces
[Flywheel](/mental-models/flywheel) Effect
The Jevons mechanism is the hidden energy source inside many of the most powerful business flywheels. Bezos's famous napkin sketch — lower prices attract more customers, more customers enable more volume, more volume enables lower costs, lower costs enable lower prices — is a flywheel whose kinetic energy comes from the Jevons effect at each rotation. Without the demand expansion that lower prices create, the flywheel has no momentum. Costco's membership model operates on the same physics: lower prices attract members, more members fund lower prices, and total spending per member actually increases because cheaper goods unlock purchases that members previously skipped. The flywheel concept describes the self-reinforcing structure. Jevons Paradox explains why the demand side of the structure generates more energy with each rotation rather than reaching equilibrium.
Tension
Second-Order Thinking
Section 8
One Key Quote
"It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth."
— William Stanley Jevons, The Coal Question (1865)
Section 9
Analyst's Take
Faster Than Normal — Editorial View
Jevons Paradox is the most underpriced mental model in technology strategy. Every efficiency-driven TAM analysis I encounter makes the same error: it models current use cases at declining prices and arrives at a shrinking market. The actual pattern — documented across 160 years of industrial history — is that efficiency creates markets. The coal didn't get conserved. The steel didn't get conserved. The bandwidth didn't get conserved. The compute isn't getting conserved. Total consumption rises because cheaper resources unlock latent demand that was invisible at the higher price.
This is not a fringe observation. It is the central mechanism by which technology-driven economic growth operates. Every major technology wave — steam, electricity, internal combustion, semiconductors, the internet, mobile computing, cloud infrastructure, AI — followed the Jevons pattern. The technology made something dramatically cheaper per unit. That cost reduction created demand that dwarfed the prior market. Total spending on the resource — and the economic activity built on top of it — expanded by orders of magnitude. The founders and investors who grasped this dynamic early captured disproportionate value. Those who modeled the future based on current demand at declining prices consistently underestimated the opportunity.
The AI inference market is the most consequential current example. When GPT-3 launched in 2020, a single API call cost roughly $0.06 per 1,000 tokens. By early 2025, equivalent capability cost under $0.001 per 1,000 tokens — a 98% reduction. Every cost model built in 2021 predicted that AI inference would be a modest market because per-query costs would fall to near zero. What those models missed was the demand side. At $0.06 per call, you build a chatbot. At $0.001 per call, you embed inference into every search result, every email draft, every code completion, every customer service interaction, every document summary. The number of inference calls doesn't grow 60x to offset the 98% price decline. It grows 1,000x. Total spending on AI inference is rising, not falling, precisely because the per-unit cost is collapsing.
The strategic error this produces is chronic underinvestment. If you believe that efficiency gains will reduce total demand, you build conservatively. You forecast declining revenue per unit and modest volume growth. You staff for a stable market. You underinvest in capacity. And then you watch a competitor who sized the Jevons effect correctly capture the demand you didn't build for.
NVIDIA's competitors made exactly this mistake in 2020–2022: they modeled GPU demand based on existing AI workloads at declining per-chip costs and concluded the market was approaching saturation. NVIDIA, operating on an implicit Jevons thesis, invested aggressively in next-generation architectures designed for workloads that didn't yet exist at scale — training runs that cost $100 million, inference deployments serving billions of daily queries. The workloads materialized because the hardware efficiency made them economically viable. Jevons rewarded the company that anticipated the demand explosion and punished those that forecast based on current usage. By 2024, NVIDIA's market capitalization exceeded $3 trillion — a valuation predicated almost entirely on the thesis that AI compute demand will grow faster than per-unit costs decline.
Section 10
Test Yourself
Jevons Paradox is frequently invoked but inconsistently applied. These scenarios test whether you can distinguish genuine Jevons effects — where efficiency improvements increase total consumption — from simple demand growth, partial rebound, and situations where the paradox doesn't apply.
The critical distinction: Jevons Paradox requires an efficiency improvement that lowers the effective cost of a service, demand that is elastic enough to expand significantly in response to that cost reduction, and total consumption that rises above the pre-efficiency baseline. A company growing revenue because it entered new markets is not a Jevons effect. A resource becoming more expensive is not a Jevons effect. And a modest demand response that partially offsets but doesn't exceed the efficiency gain is a rebound, not the full paradox.
The most diagnostic question for each scenario: did an efficiency improvement cause the demand expansion, or would the demand have grown regardless? Jevons Paradox is a causal claim — that efficiency is the mechanism driving consumption growth. Correlation between efficiency improvements and consumption growth is common. Causation through the price-elasticity channel is the specific claim being tested.
Is this mental model at work here?
Scenario 1
A data center operator deploys new servers that deliver 3x the compute performance per watt. The operator's energy bill for the facility drops by 15% in the first quarter. Over the following two years, the operator adds workloads — AI training jobs, real-time analytics, video encoding — that were previously uneconomical at the old energy cost per operation. By year three, the facility's total energy consumption is 40% higher than before the server upgrade.
Scenario 2
A city replaces all 200,000 of its streetlights with LED fixtures that use 60% less electricity per lumen. Five years later, the city's total electricity consumption for street lighting has declined by 52%. The city did not add new streets or increase lighting density during this period.
Scenario 3
A streaming platform reduces its video encoding costs by 80% through a new compression algorithm. Within eighteen months, the platform adds 4K streaming, introduces a free ad-supported tier, expands to 40 new countries, and begins offering live sports. Total bandwidth consumption across the platform increases by 6x. Total spending on content delivery infrastructure rises by 3x despite the per-gigabyte cost being 80% lower.
Section 11
Top Resources
The literature on Jevons Paradox spans Victorian political economy, twentieth-century energy economics, and contemporary technology strategy. The concept sits at the intersection of microeconomics (price elasticity of demand), macroeconomics (structural transformation), and business strategy (demand creation through cost reduction). Start with Jevons himself for the original argument — it remains more readable and rigorous than most modern summaries. Advance to the empirical work for quantitative estimates of rebound effects, which are essential for distinguishing markets where the paradox produces full backfire from those where it produces only partial rebound.
The original text. Jevons's prose is Victorian but his logic is sharp and empirical. Chapters VII through X contain the core argument about efficiency and consumption, supported by sixty years of coal production data. Chapter VII — "Of the Economy of Fuel" — contains the paradox's clearest articulation, arguing from iron production, steam power, and textile manufacturing data that every efficiency improvement was followed by consumption growth. The book is freely available online and shorter than its reputation suggests. Essential reading for anyone who wants the primary source rather than a secondhand summary.
The most comprehensive modern academic treatment of the paradox and its implications for resource policy. Polimeni and colleagues survey the empirical evidence for Jevons effects across energy, transportation, agriculture, and manufacturing. The key contribution is demonstrating that micro-level rebound studies (which typically find 20–40% rebound) systematically miss macro-level structural transformations that push actual rebound well above 100% in high-elasticity sectors. The chapters on why sector-specific estimates fail to capture economy-wide effects are particularly valuable for anyone using partial-rebound figures in strategic or investment analysis.
The energy economics literature published in Energy Policy has produced the most rigorous quantitative estimates of rebound effects across sectors and geographies. Key papers by Harry Saunders (the Khazzoom-Brookes postulate, showing macroeconomic backfire under neoclassical growth conditions), Lorna Greening (meta-analysis of household energy rebound at 10–40%), and Steve Sorrell (taxonomy of direct, indirect, and economy-wide effects) provide estimates ranging from 10% to over 100% depending on the sector, timeframe, and level of analysis. Essential for calibrating how large the Jevons effect is likely to be in specific contexts rather than treating it as binary — the difference between 20% rebound and 200% backfire has massive implications for infrastructure investment and market sizing.
Saunders's 1992 paper formalized the conditions under which improved energy efficiency leads to increased total energy consumption at the macroeconomic level — the "backfire" scenario that Jevons described qualitatively but never modeled mathematically. His neoclassical growth model shows that when energy is a complement to capital and labor (rather than a substitute), efficiency improvements increase the productivity of all factors simultaneously, driving economic growth that increases total energy demand. The critical insight: the complementarity assumption holds for most real economies, because energy is an input to virtually all production processes. The theoretical rigor and mathematical precision make this the key bridge between Jevons's 1865 historical observation and modern quantitative economic modeling of rebound effects.
The UK Energy Research Centre's definitive review of rebound evidence, and the single best synthesis of the empirical literature. Sorrell surveys over 500 studies and concludes that direct rebound effects are typically 10–30%, but economy-wide effects are substantially larger and poorly measured. The report's three-level taxonomy of rebound mechanisms — direct (same-service substitution), indirect (income-effect reallocation), and economy-wide (structural transformation) — provides the clearest framework for analyzing where Jevons effects will be strong versus weak in any given sector. The finding that economy-wide effects are the largest but least studied channel has direct implications for technology strategy: the biggest demand creation from efficiency improvements comes through structural channels that sector-specific analysis systematically misses.
Second-order thinking and Jevons Paradox exist in productive tension. Second-order thinking asks: "And then what?" Applied to an efficiency improvement, the second-order thinker should predict the demand rebound. But in practice, even sophisticated second-order analyses consistently underestimate the magnitude of the Jevons effect. Energy policy models in the 1990s acknowledged the rebound effect but estimated it at 10–30% — meaning that a 100% efficiency improvement would still reduce total consumption by 70–90%. Actual rebounds in computing, transportation, and lighting have regularly exceeded 100%, producing net increases in consumption. The tension is that second-order thinking is necessary to see the paradox but insufficient to size it. The demand response to efficiency gains operates through structural channels — new industries, new use cases, behavioral changes — that are difficult to forecast even with disciplined second-order analysis.
Tension
[Inversion](/mental-models/inversion)
Inversion asks: "What if the opposite is true?" Jevons Paradox is the answer to that question applied to resource efficiency. The naive model says efficiency reduces consumption. Inversion suggests checking whether it increases consumption. The tension is that Jevons Paradox doesn't merely invert the expected outcome — it reveals that the expected outcome was based on a static model of demand, while the actual outcome operates in a dynamic system where prices, preferences, and possibilities change simultaneously. Inversion is a useful heuristic for spotting the paradox, but the paradox goes deeper than simple inversion because the mechanism involves structural economic transformation, not just a reversal of direction. The steam engine didn't just increase coal use — it created entirely new categories of coal-consuming activity that didn't exist before. Inversion identifies the directional surprise. Jevons explains the structural mechanism.
Leads-to
Creative Destruction
Jevons Paradox is one of the primary engines of Schumpeterian creative destruction. When efficiency improvements make a resource dramatically cheaper, the demand explosion creates new industries — and those new industries displace the old ones that depended on the resource being scarce or expensive. Cheap steel didn't just mean more of the same bridges and railroads. It enabled skyscrapers, automobiles, and mass-produced machinery — each of which destroyed the industries they replaced (wooden buildings, horse-drawn transport, artisan manufacturing). Cheap compute didn't just accelerate existing software. It created cloud computing, mobile applications, social media, and AI — industries that destroyed traditional software distribution, print media, brick-and-mortar retail, and are now disrupting knowledge work. The Jevons effect creates the surplus demand that funds the new entrants. Creative destruction is the market's mechanism for allocating that surplus toward the most productive new applications.
Leads-to
Network Effects
When Jevons effects drive a product to mass adoption, network effects frequently emerge as a secondary dynamic that amplifies the consumption growth beyond what price elasticity alone would predict. Cheap telephony (driven by switching and transmission efficiency) reached enough households to trigger network effects — each new phone making every existing phone more valuable — which accelerated adoption beyond what price alone explained. Cheap smartphones (driven by semiconductor efficiency) reached enough users to make app development economically viable, which made smartphones more valuable, which drove more adoption. Cheap cloud compute enabled SaaS platforms to reach enough users to benefit from data network effects and marketplace dynamics. In each case, the Jevons-driven demand expansion was the necessary precondition for network effects to activate. Without the price decline to critical mass, the network effects remain latent. Jevons provides the ignition. Network effects provide the acceleration.
The pattern extends to energy, where the stakes are measured in trillions of dollars and planetary-scale infrastructure. Solar electricity costs have fallen roughly 90% per kilowatt-hour since 2010. Analysts in 2010 modeled total solar generation growing modestly as costs declined — the International Energy Agency's 2010 projection for global solar capacity in 2024 was off by a factor of six. Actual solar generation grew more than 20x, and total global electricity consumption continued rising because cheaper electricity made new applications viable — data centers, electric vehicles, heat pumps, desalination, indoor agriculture, cryptocurrency mining.
The per-kilowatt-hour cost fell dramatically. Total kilowatt-hours consumed rose relentlessly. Every country that built energy policy around "efficiency will reduce total consumption" has been surprised by the Jevons effect. Germany's Energiewende invested hundreds of billions in renewable energy and efficiency standards; total German electricity consumption declined only modestly, and the efficiency gains were largely absorbed by new demand categories. China, which built solar manufacturing capacity on the thesis that cheap solar would expand global electricity demand rather than merely displace fossil generation, captured the largest share of the resulting market. Planning for demand growth is not optimistic forecasting. It is what the historical evidence supports.
Where most analysts get it wrong is in confusing partial rebound with no rebound. Empirical studies in energy economics typically estimate rebound effects of 20–60% for specific efficiency improvements. That is, a 100% efficiency improvement produces a net reduction in consumption, but smaller than expected — usage rises 20–60% to offset some of the savings. These estimates are correct at the micro level for individual technologies with constrained use cases. They miss the macro-level structural transformations that produce full backfire.
The car didn't just make horse-based transport more efficient. It created suburbs, supply chains, tourism industries, and a geopolitical order organized around petroleum. Cloud computing didn't just make existing IT workloads cheaper. It created SaaS, mobile apps, social media analytics, real-time recommendation engines, and the entire data science profession. The structural transformation is where the Jevons effect operates at multiples, not percentages — and it is precisely the structural transformation that sector-specific rebound studies fail to capture. By the time the new industries are visible, the Jevons effect has already operated for years. Modeling it retroactively is easy. Forecasting it prospectively requires the same conviction that Rockefeller, Ford, and Bezos demonstrated: that lowering the cost of a broadly useful resource will create demand you cannot yet specify.
The founder-level insight is this: if you are building something that makes a service dramatically cheaper per unit, you are not in the cost-reduction business. You are in the demand-creation business. Price your product to accelerate adoption, not to preserve margin. Invest in infrastructure for volume levels that current demand doesn't justify. Build for the market that will exist after your efficiency improvement reshapes behavior — not the market that exists today. Rockefeller, Ford, Bezos, Huang, and Altman all operated on this principle. They understood that making each unit cheaper wouldn't shrink the market. It would expand it beyond what anyone staring at current demand could imagine.
The contrarian position this creates is powerful: when everyone else in the market is projecting modest growth because per-unit prices are falling, the Jevons-informed investor or founder projects explosive growth precisely because per-unit prices are falling. The asymmetry of this bet — modest downside if the demand response is partial, enormous upside if it's full Jevons backfire — has been one of the most reliable sources of outsized returns in technology investing over the past three decades.
The clearest tell for a Jevons-susceptible market: look at what happens immediately after a step-function cost reduction. If usage spikes within months, latent demand was enormous and the paradox will compound from there. AWS's price cuts produced usage spikes within weeks. GPT-3.5 Turbo's price cut in mid-2023 produced a visible surge in API call volume within days. NVIDIA's A100 GPU launch in 2020 was followed by an immediate backlog of orders from organizations that had been waiting for the performance-per-dollar threshold to justify their training runs. The speed of the demand response after a cost reduction is the leading indicator for the magnitude of the Jevons effect. Slow demand response suggests inelastic demand and modest rebound. Immediate demand response suggests vast latent demand and full backfire.
The one place to apply skepticism: Jevons Paradox requires elastic demand. Not every efficiency improvement triggers the effect. Making a niche industrial chemical 50% cheaper to produce won't 10x the market if the chemical's applications are limited. The paradox operates most powerfully where latent demand is massive and price is the binding constraint — energy, compute, transportation, communication, information processing. In those domains, treat every efficiency breakthrough as a demand signal, not a conservation signal.
Scenario 4
A pharmaceutical company develops a new synthesis process that reduces the manufacturing cost of a specialty cancer drug by 45%. The drug treats a rare cancer affecting approximately 8,000 patients per year in the US. The company reduces the wholesale price by 20% and increases its gross margin. Total unit sales grow by 12% over three years as slightly more patients gain insurance coverage for the treatment. Total spending on the drug declines by approximately 10%.
