Jevons' Paradox Applied to Intelligence

1. The Theoretical Framework: Mapping Jevons Onto Intelligence

Defining the Variables Precisely

Jevons' Paradox requires precise variable mapping to be analytically useful rather than metaphorical:

• The resource: Cognitive labor hours — the fundamental input to any knowledge-work task. A legal review, a medical diagnosis, a financial analysis, a code review — each consumes a measurable quantity of cognitive labor.
• Efficiency: Cost per unit of cognitive output. Measurable as $/million tokens, $/task-completion, $/correct-diagnosis, etc. This has declined 10x annually, faster than PC compute costs or dot-com bandwidth. [7]
• Consumption: Total cognitive tasks performed across the economy. Not just the tasks previously done by humans, but the sum of human + AI cognitive output.
• The paradox condition: Total consumption rises when efficiency gains outpace substitution effects — i.e., when demand elasticity |ε| > 1.

The Zhang & Zhang "Structural Jevons Paradox"

The critical contribution of Zhang & Zhang (2026) is demonstrating that AI demand isn't just elastic — it's architecturally elastic. [3] Their model introduces the "token multiplier": as inference costs fall, developers don't merely run the same prompts more often. They fundamentally redesign their systems:

This is why standard demand analysis fails. A traditional economist would model AI demand as a downward-sloping curve along a fixed production function: lower price → more units consumed, ceteris paribus. Zhang & Zhang show the production function itself changes. The "ceteris" are not "paribus." The entire technology stack reorganizes around cheaper intelligence, just as the entire economy reorganized around cheaper energy after the steam engine.

The Narayanan-Pace Five-Phase Model

Narayanan & Pace (2025) formalize the transition dynamics using a Variable Elasticity of Substitution (VES) framework. [4] Their model identifies five phases of AI market penetration, governed by the compound parameter:

The five phases:

1. Novelty (σ ≪ 1): AI is a curiosity. Human and AI are not substitutes. No displacement.
2. Complement (σ < 1): AI augments human workers. Productivity rises. Demand for both human and AI labor increases. This is the Jevons sweet spot — total cognitive output expands.
3. Substitution threshold (σ → 1): AI begins displacing human labor in specific tasks. Wages for affected tasks decline. The political economy of AI heats up.
4. Substitution (σ > 1): AI is cheaper and often better than human labor for a growing set of tasks. But Jevons operates: the total market for cognitive work expands as new applications become viable. Who captures the surplus becomes the critical question.
5. Transformation (σ ≫ 1): Human labor and AI are highly substitutable in most cognitive tasks. New institutional arrangements determine whether Jevons expansion or Chen-style demand deficiency prevails.

2. Where Demand Is Super-Elastic vs. Inelastic

The Jevons paradox does not apply uniformly. The key question for investors: in which sectors is demand for intelligence super-elastic (|ε| > 1, meaning Jevons applies), and in which is it inelastic (|ε| < 1, meaning simple cost savings)?

The Elasticity Test: Three Conditions for Super-Elastic Demand

A sector exhibits super-elastic demand for intelligence when:

1. Large unserved demand exists at lower price points. There are people who need the service but can't afford it at current intelligence costs. (Legal services: 80% of civil legal needs in the US go unmet. [10])
2. Quality improvements unlock new use cases. It's not just cheaper — it's better enough to enable applications that weren't possible before. (Drug discovery: AI can screen 10⁹ molecular candidates vs. 10³ in traditional wet labs.)
3. The service was previously rationed by human attention, not by demand. There was always demand for personalized tutoring; the binding constraint was the supply of tutors, not the desire for tutoring.

3. Mechanisms of New Market Creation

Saying "new markets emerge" is insufficient for LP-grade analysis. The question is how — through what specific economic mechanism does cheaper intelligence create markets that didn't exist?

Mechanism 1: Threshold Economics (The "Below the Floor" Effect)

Every cognitive service has a minimum viable cost. Below that cost, the service cannot be profitably delivered. When intelligence costs drop by 100x, activities that were below the economic floor become viable.

Historical parallel: Long-distance phone calls. In 1970, a coast-to-coast call cost ~$12/minute (2024 dollars). Calls were rationed — "keep it short." When VoIP dropped the cost to essentially $0, people didn't just make cheaper calls. They restructured their social and business relationships around the assumption of free communication. Long-distance friendships, remote work, global teams — all "new markets" that were inconceivable when communication was expensive.

Mechanism 2: Baumol's Cost Disease in Reverse

William Baumol observed in 1967 that labor-intensive services (education, healthcare, performing arts) rise in cost relative to goods because they can't achieve productivity gains — it still takes the same number of musicians to play a string quartet as in Beethoven's time. This "cost disease" has made services increasingly expensive relative to manufactured goods for 60 years.

AI is the first technology that may reverse Baumol's cost disease. Legal services, education, medical diagnosis, financial advisory, therapy — all have been trapped in Baumol's disease because they require irreducibly human cognitive labor. If AI provides a substitute for that cognitive labor, these services can follow the cost curve of manufactured goods: falling in real price over time. [9]

The Jevons implication: When Baumol-afflicted services suddenly become cheap, demand doesn't just grow linearly — it erupts. We have 60 years of pent-up demand for services that got more expensive every year. Education, legal services, healthcare, and financial advisory together represent >$15T in global spending constrained by Baumol's disease. If AI breaks the disease, even a fraction of that pent-up demand becoming addressable creates multi-trillion-dollar new markets.

Mechanism 3: The Intermediation Collapse (Chen's Warning)

Chen (2026) identifies a mechanism that both creates and destroys value. [5] When AI reduces information frictions, intermediary margins compress toward pure logistics costs. SaaS, consulting, insurance brokerage, financial advisory, real estate agents — all derive value from information asymmetry. AI collapses that asymmetry.

This is simultaneously Jevons (positive) and displacement (negative):

• Jevons side: End consumers access services that were previously gatekept by intermediaries. More people get legal help, investment advice, and insurance optimization → market expands.
• Chen side: The intermediaries employed significant cognitive labor. That labor is displaced. The income they earned circulated through the economy as consumption. When it disappears, demand contracts — "Ghost GDP."

Investor implication: Jevons and Ghost GDP can be simultaneously true in different sectors. The investment strategy isn't to bet on one thesis — it's to bet on sectors where Jevons dominates (unserved demand > displacement) and avoid sectors where Ghost GDP dominates (displacement > new creation). The elasticity estimates in the table above are the guide.

4. Historical GPT Adoption Chains

General Purpose Technologies (GPTs) in the economic sense — technologies that permeate across sectors and create cascading downstream effects — follow remarkably consistent patterns. The academic literature on GPT diffusion (Bresnahan & Trajtenberg, 1995) identifies three properties: pervasiveness, improvement over time, and innovation spawning. AI exhibits all three. [3]

Chain 1: Steam Power (1765-1900)

Non-obvious link: Steam → department stores. The department store required (a) urbanization (concentrated consumers), (b) railroads (supply chain logistics), and (c) mass production (inventory). None could exist without cheap energy. But nobody building steam engines in 1765 would have predicted browsing-based retail 93 years later. [11] [12]

Chain 2: Electrification (1882-1960)

Non-obvious link: Electricity → women's workforce participation. The washing machine, refrigerator, and vacuum cleaner reduced household labor by ~40 hours/week, enabling women to enter the paid workforce at scale. This was a more consequential economic effect than the factory productivity gains that justified electrification — but it took 50+ years to manifest.

The productivity paradox note: Economists measured essentially zero productivity gain from electrification for the first 20 years (1882-1900s) because factories simply replaced steam with electric motors without redesigning production. Productivity exploded only when factories were redesigned around the unique properties of electricity (individual motor drive, flexible layout). David (1990) argues AI may follow the same pattern — the gains come from institutional redesign, not drop-in replacement. [13]

Chain 3: Internet (1995-2025)

Non-obvious link: Cheap information → gig economy. The internet didn't just make buying things cheaper — it made coordinating strangers cheap. Uber's insight wasn't "cheaper taxis"; it was "you can match supply and demand for rides in real-time at zero coordination cost." The creator economy ($250B+ by 2025) was inconceivable in 1995. The internet didn't make existing media cheaper — it made everyone a potential media company.

Chain 4: Intelligence (2022-?)

5. The Counter-Thesis: When Jevons Fails

Chen's "Abundant Intelligence and Deficient Demand" (2026) is the most rigorous challenge to Jevons optimism. [5] It deserves serious engagement, not dismissal.

Three Mechanisms of Demand Destruction

1. The Displacement Spiral. Each firm's rational decision to substitute AI for labor reduces aggregate labor income, which reduces aggregate demand, which accelerates further AI adoption. Chen derives conditions on the AI capability growth rate, diffusion speed, and reinstatement rate under which this feedback is self-limiting versus explosive. If new job creation (Marguerit's "augmentation AI" effect [9]) lags displacement, the spiral feeds itself.

2. Ghost GDP. When AI-generated output substitutes for labor-generated output, monetary velocity declines. The economy produces more "stuff" but the income to buy it shrinks. Measured GDP might rise while consumption-relevant income falls. This is a demand-side crisis masquerading as supply-side abundance.

3. Intermediation Collapse. AI agents that reduce information frictions compress intermediary margins toward pure logistics costs. SaaS (30-40% of revenue is information arbitrage), consulting ($500B globally), insurance brokerage, real estate — all face margin compression as AI eliminates the information asymmetry that justified their margins. The income these sectors generate (high salaries for knowledge workers) has outsized consumption impact because top-quintile earners drive 47-65% of US consumption.

Can Jevons and Chen Both Be True?

Yes — in different sectors simultaneously. The resolution:

6. Investable Implications

The Quantitative Framework

Using Zhang & Zhang's calibrations: if the token multiplier for agentic workflows is ~10x (conservative — they observe up to 100x), and demand elasticity θ = 2.5, we can project:

Where to Invest: The Jevons Filter

Apply three filters simultaneously:

Filter 1: Is demand super-elastic? (|ε| > 1 — large unserved market below current cost floor)

Filter 2: Is it Jevons, not Chen? (Creating new demand, not compressing existing intermediaries)

Filter 3: Is the timing right? (Is the threshold price being crossed now, not in 5 years?)

Sectors that pass all three filters today:

1. Sub-$500 legal services: Unserved demand is 80%+. AI provides a genuinely new service (nobody was getting legal help for $200 disputes before). Threshold is being crossed now. This is the canonical Jevons opportunity.
2. Personalized education at the $10-100/month price point: 1:1 tutoring was $50-100/hr. AI tutoring at $10/month creates an entirely new market — not displacing Kumon or tutors, but serving the 95% who couldn't afford them.
3. Financial advisory for sub-$100K households: 90% of US households below typical AUM minimums. AI opens this market at $5-50/month. The advisor market doesn't shrink — it expands 10x by adding the unserved.
4. Continuous scientific experimentation: The number of hypotheses worth testing is infinite. Human attention was the bottleneck. AI removes it. R&D spending expands into "long-tail research" — questions too small for traditional funding but valuable in aggregate.
5. Software creation for non-developers: Every business has custom software needs. At current dev costs ($150K+/engineer), only the top 5% of businesses can afford custom software. At AI-assisted development costs ($5-50K), the bottom 95% becomes addressable. [14]

The Scarcity Inversion

In every Jevons cycle, making one resource abundant creates scarcity in an adjacent resource. For investors, the newly scarce resource is often where the most durable value accrues: