Moore's Law vs Jevons' Paradox

The Supply-Demand Feedback Loop

Moore's Law and Jevons' Paradox are not competing theories — they are two halves of a single flywheel. Moore's Law (and its successors like Wright's Law and Huang's Law) drives down the unit cost of computation. Jevons' Paradox describes what happens next: cheaper compute unlocks new use cases, new users, and new architectures that consume far more total compute than the efficiency gain saved. The semiconductor industry has traced this loop for decades — the cost per transistor fell by a factor of roughly 10 million between 1970 and 2020, yet global semiconductor revenue grew from under $1 billion to over $600 billion, because each generation of cheaper chips created entirely new markets that consumed orders of magnitude more silicon.

From Coal to Tokens: The AI Inflection

The AI era offers the most dramatic demonstration of this interaction in history. Between November 2022 and October 2024, the cost to run inference at GPT-3.5-level performance dropped from $20.00 to $0.07 per million tokens — a 280× decline. By August 2025, the cost-efficiency frontier had fallen 99.7% from GPT-4's March 2023 pricing. Rather than shrinking the AI market, this price collapse triggered an explosion of demand. A January 2026 paper by Zhang and Zhang formalized this as the Structural Jevons Paradox: as the unit price of intelligence falls, firms don't just use the same amount more cheaply — they fundamentally redesign their agent architectures to consume dramatically more compute, adopting deeper reasoning loops, larger context windows, and multi-agent orchestration patterns that would have been economically impossible at prior price points.

Why Dennard's Breakdown Matters for Both

The breakdown of Dennard Scaling around 2006 — when smaller transistors stopped using proportionally less power — is a critical shared inflection point. For Moore's Law, it meant that raw clock-speed improvements stalled and the industry pivoted to parallelism (multi-core CPUs, then GPUs, then AI accelerators). For Jevons' Paradox, the pivot to parallelism was what made GPUs viable for machine learning, which unlocked the deep-learning revolution, which created the super-elastic demand for AI compute we see today. In other words, a constraint on Moore's Law triggered the architectural shift that produced the most powerful Jevons effect in computing history. NVIDIA's Rubin GPU — 336 billion transistors on TSMC 3nm, delivering 50 petaFLOPS of FP4 inference — embodies both dynamics: Moore's Law scaling packed into Huang's Law architectural gains, feeding Jevons-driven demand that has NVIDIA's order backlog exceeding $1 trillion.

The Limits of Each Framework

Neither framework is universal. Moore's Law faces hard physical constraints: quantum tunneling effects at sub-5nm geometries, the astronomical cost of leading-edge fabs (TSMC's Arizona fab costs over $40 billion), and the transition from FinFET to gate-all-around nanosheet transistors that buys perhaps one or two more generations of classical scaling. Jevons' Paradox, meanwhile, requires that demand be sufficiently elastic — if a market is saturated or regulated, efficiency gains can genuinely reduce total consumption. Energy markets sometimes show this: LED lighting efficiency gains have, in some mature markets, actually reduced total electricity consumption for illumination. The question for AI is whether compute demand will remain super-elastic as the technology matures, or whether saturation effects will eventually dampen the rebound. Current evidence — with AI CAPEX still accelerating in 2026 and AMD warning at Semicon China 2026 that efficiency-fueled demand growth shows no signs of plateauing — favors continued Jevons dynamics for the foreseeable future.

Implications for the Agentic Economy

For builders and investors in the agentic economy, the interaction between these two frameworks produces a clear strategic signal: the cost of intelligence is falling on a Moore's-Law-like curve, but total spending on intelligence will grow on a Jevons' Paradox curve. This means the market for AI infrastructure, tooling, and orchestration is expanding faster than efficiency gains would naively suggest. Software engineer job postings — after dipping below baseline in early 2025 as companies absorbed initial AI productivity gains — sharply accelerated by year-end, growing significantly faster than overall postings. The pattern mirrors every prior Jevons cycle: a brief contraction as the efficiency gain is absorbed, followed by demand expansion that overwhelms the original savings. Companies building for the agentic economy should plan for compute budgets that grow even as unit costs fall — because Jevons' Paradox and flywheel economics ensure that cheaper intelligence creates more demand for intelligence, not less.

Historical Parallels Beyond Computing

The Moore's Law / Jevons' Paradox interaction is not unique to semiconductors — it recurs wherever deflationary technology meets elastic demand. Watt's steam engine made coal usage per unit of work plummet, then total coal consumption soared as steam power became economical for factories, railways, and ships. Henry Ford's assembly line (a manufacturing analog of Moore's Law) cut the cost of a car from $850 to $260, and Jevons' Paradox ensured that American car registrations went from 8,000 in 1900 to 23 million by 1930. The internet followed the same arc: bandwidth costs per megabit fell roughly 99% between 1998 and 2015, while total internet traffic grew over 1,000×. Each case illustrates the same principle: Wright's Law cost curves drive down unit prices, and Jevons' Paradox converts those savings into market expansion that dwarfs the original efficiency gain.