Wright's Law (Experience / Learning Curve)
Wright's Law (Experience / Learning Curve)
One-line summary: Unit cost falls a roughly constant percentage per doubling of cumulative production (the "learning rate") — the cost engine that drives the adoption S-curve, and empirically the best single predictor of technological progress.
The insight
First observed by Theodore Wright in 1936 for aircraft, Wright's Law holds that cost declines as a power law of cumulative output. Across an expanded dataset of 150 technologies the median learning rate is ~21% per doubling. It is the cause sitting beneath adoption: as cost falls and crosses a price/performance tipping point, the new technology beats the incumbent and adoption goes non-linear. The live exemplar is AI inference ("LLMflation"): cost per fixed capability is falling ~10×/year — an MMLU-42 capability fell from $60/M-tokens (Nov 2021, GPT-3) to $0.06 (Nov 2024, Llama 3.2 3B), ~1,000× in three years, "even faster than … compute cost during the PC revolution or bandwidth during the dotcom boom."
Two important caveats: the learning rate is not constant (it changes for ~66% of technologies, and "the observed learning rate is not a good predictor of future learning"), and a falling per-capability cost can coexist with rising frontier spend (AI frontier-model run-cost rose ~3-18×/yr even as commodity inference collapsed).
Evidence
- From 2026-06-16-academic-research-technology-adoption-s-curves: "unit costs decline as a power law of cumulative production"; median learning rate ~21%; GPU compute the highest measured; nuclear shows persistent anti-learning (Gogerty 2026, preprint — figures provisional).
- From 2026-06-16-academic-research-technology-adoption-s-curves: "the observed learning rate is not a good predictor of future learning"; learning rates change for ~66% of technologies (Carlino et al. 2025).
- From 2026-06-16-autoresearch-tony-seba-technology-disruption-s-curves: LLMflation — "$60/million tokens in Nov 2021 … $0.06 by Nov 2024 — a ~1,000× decline in three years"; experience-curve dynamic (a16z); utility solar under ~4¢/kWh and ~$100/kWh lithium packs.
Design implications
- The cost curve is the forecastable part of a disruption (see forecastability-of-technological-progress); lean on it for direction, treat adoption timing as a distribution.
- ARK's stated method: cost falls per Wright's Law → "positive deflationary forces drive down prices and drive up demand," sometimes expanding the addressable market (claim from search snippet; ARK page unfetched — flagged unverified in the source).
Contradictions / tensions
- Constant-learning-rate models are systematically optimistic/pessimistic in different regimes; stepwise learning rates fit better for most technologies (Carlino et al. 2025).
- Some technologies show negative learning (nuclear) — Wright's Law is not a law of nature, it's an empirical regularity with exceptions.