Wright's Law reveals a deep regularity: for every doubling of cumulative production, technology costs fall by a remarkably consistent percentage. This isn't hindsight — it's a forecasting tool. 155 technologies, 168 years of data, one simple equation.
In 1936, Theodore Wright observed that every time cumulative aircraft production doubled, the cost per unit dropped by a fixed percentage. This "learning curve" turned out to be one of the most powerful prediction tools in economics.
A landmark study by the Santa Fe Institute (Nagy et al., 2013) tested six competing forecasting models against 62 technologies and found Wright's Law to be the best simple prediction model — outperforming Moore's Law, Goddard's, and others. The result is striking: a single-parameter power law, fit on past data, reliably forecasts future costs.
Gogerty (2025) explains why this works: learning curves emerge from the interaction of selfish symbols — ideas, techniques, and capital competing to reduce cost. The learning rate measures how efficiently a technology's ecosystem converts experience into savings. This dataset extends the SFI database from 62 to 155 technologies, adding AI, carbon removal, and 2025 price updates.
The learning rate is the percentage cost reduction per doubling of cumulative production. A 20% learning rate means costs drop 20% each time total output doubles — and this rate tends to persist, making it a forecasting input.
Wright's Law isn't just descriptive — it's predictive. Here's the science behind why a simple equation forecasts technology costs better than expert consensus.
The Santa Fe Institute's Performance Curve Database is the gold standard for empirical technology forecasting. In their landmark study, Nagy, Farmer, Bui & Trancik (2013) tested six competing models against 62 technologies and found Wright's Law produced the most accurate out-of-sample forecasts.
The result is remarkable: a single-parameter power law, fit on historical data, reliably predicts future costs — often decades ahead. Solar PV costs were forecast in 2010 to reach $0.50/W by 2030. They hit $0.10/W by 2024.
🏛 Explore the Santa Fe Institute PCDB →Why does a single equation outperform complex models and expert panels? Gogerty (2025) argues that learning curves are an emergent property of selfish symbols — ideas, techniques, and innovations competing within technological ecosystems to reduce cost and increase efficiency.
This evolutionary view, drawn from The Symbology, explains why learning rates are so stable: they measure the rate of evolution in a technology's knowledge ecosystem. Technologies with rich, competitive idea-spaces (like solar PV and GPUs) learn fast. Technologies trapped in regulatory complexity (like nuclear) learn in reverse.
This dataset extends the Santa Fe Institute's original 62-technology database to 155 technologies, adding AI/machine learning (GPU compute, LLM inference, algorithmic efficiency), carbon removal (DAC, BECCS, biochar, enhanced weathering), and updated 2024–2025 prices across energy and hardware.
The full methodology, regression results, and confidence intervals are available in the research paper: Experience Curves Extended: Wright's Law Across 155 Technologies.
📄 Read the Paper on SSRN →Every learning rate in this database is a forecast input. If solar PV has a 20.5% learning rate, and you can estimate future cumulative production, you can project future cost. This is how IRENA, BloombergNEF, and energy modelers build their scenarios.
Carbon removal technologies are early in their curves — tracking where solar was in the 1980s. If DAC follows a 15–20% learning rate (as BECCS and biochar suggest), costs could fall from $600/tonne to under $100/tonne within two decades of scaling.
Prediction is not certainty. Learning rates can shift if policy, regulation, or material constraints change. See methodology notes below.
Each bubble is a technology. The x-axis shows its learning rate — the cost decline per doubling of cumulative production. Click any bubble to explore its data.
Distribution of learning rates across 83 technologies with Wright's Law fits. Each rate is a forecast input.
If individual technologies learn, does civilization itself learn?
And if so — fast enough?
Energy intensity vs. cumulative GDP on a log-log scale, 1970–2023. The slope gives an 18.3% learning rate (R² = 0.96) — remarkably close to the median across 155 individual technologies.
Source: Maddison Project Database, BP/Energy Institute Statistical Review. OLS regression on log-transformed data.
What the world thinks it's learning, what it's actually learning, and what it must learn.
Under the current learning rate, annual CO₂ emissions do not decline — they rise continuously. The Jevons paradox at planetary scale: GDP growth overwhelms intensity improvement.
Projection: 3% real GDP growth, 22.2% carbon learning rate held constant. Paris pathways per IPCC AR6.
Individual technologies learn fast. The global economy, weighed down by materials, institutions, and externalities, learns much more slowly. Solar PV at 49% is the closest precedent for the 52% required for 1.5°C.
Sources: SFI Performance Curve Database, Farmer & Lafond (2016), IRENA (2023), Ziegler & Trancik (2021).
Direct Air Capture, BECCS, and Enhanced Weathering follow their own experience curves. Current costs are high — but so were solar panels in 2005.
Projected via Wright's Law: C(x) = C0 × (x/x0)−b where b = −log2(1 − LR). Cumulative removal targets per IPCC AR6 & Global CCS Institute.
With 10 Gt/yr of carbon removal by 2050, the required learning rate drops from 52% (unprecedented) to ~35% (demonstrated by LED lighting and wind).
Required learning rate recalculated assuming CDR offsets gross emissions. Reference rates from SFI Performance Curve Database.
At 37.4 Gt/year, the 1.5°C budget (50% probability) is exhausted by approximately 2030–2031.
The budget may already be functionally exhausted once carbon cycle feedbacks and non-CO₂ forcing are included.
“The learning curve is a tool. What we learn on it is up to us.”
Every other technology in this database gets cheaper with scale. Nuclear got more expensive. Costs rose 5× while cumulative deployment grew 400× — the most dramatic failure of learning-curve economics in industrial history.
Had nuclear followed even a modest 15% learning rate, costs would be under $500/kWe today. The 8× gap is the defining chart in energy economics.
Source: EIA, Lovering et al. (2016), IAEA PRIS. 2023 real dollars. CarbonSig Research, Feb 2026.
Solar PV has maintained a stable 20% learning rate for four decades. Nuclear’s negative rate means building more reactors has been associated with higher costs.
Learning rates from Grubler (2010), IRENA (2024), NREL ATB (2024), and this database (curve_161).
Small Modular Reactors are the first credible structural attempt to satisfy Wright’s Law preconditions for nuclear: factory fabrication, standardized designs, and shorter build times. But the evidence so far is sobering.
FOAK estimates from NREL ATB 2024, NuScale CFPP filings. NOAK projections assume 9.5% (optimistic) and 5% (base) learning rates.
Factory fabrication moves 35% of capital cost into controlled environments (vs 5% for large reactors). Standardized designs enable worker continuity. Target build time: 3–5 years vs 10+ today.
NuScale’s CFPP reached $20,139/kW before cancellation — comparable to Vogtle. Zero operational Western SMRs. No factory exists. The 9.5% learning rate assumption has never been demonstrated for nuclear.
At 9.5% learning rate, SMRs reach $3,600/kW target after ~50 serial builds. At 5%, it takes 200+. At the historical −23%, they never get there. The first 10–20 units reveal which trajectory prevails.
Speed may matter more than overnight cost. At 8% WACC, a 10-year build doubles the overnight cost via financing. A 3-year SMR build at $10,000/kW beats a 15-year large reactor at $6,600/kW on total installed cost.
From record-breaking learners to anti-learners — each learning rate is a window into the future of that technology.
Distribution of Wright's Law learning rates by sector. Each point is a technology. The box shows the interquartile range; the line marks the median.
83 technologies with Wright's Law fits. Search, sort, and click to explore.
| Name | Category | Learning Rate | R² | Year Range | Points | Trend |
|---|
The complete Wright's Law dataset — 155 technologies, 168 years of cost data, with 2025 updates and full source attribution. All data freely available under Creative Commons CC BY 4.0.
All 155 technologies with learning rates, R², confidence intervals, latest 2025 prices, cumulative production, and source URLs.
Latest cost and production data points from 2024-2025, ready to extend existing experience curves.
Every data source cited — 100+ URLs, publishers, years, and coverage. Full provenance chain.
Time series for each technology with annual cost, production, and cumulative totals. 155 JSON files.
Gogerty (2025): "Experience Curves Extended: Wright's Law Across 155 Technologies." Full methodology.
Gogerty (2025): "The Global Learning Rate: The World Economy as a Learning System." Energy, carbon, and material learning rates.
Gogerty (2026): "Why Nuclear Energy Costs Keep Rising — And Whether SMRs Can Reverse the Curve." 20 expert analyses, 20 charts.
Copy the iframe code below to embed the interactive scatter plot on your own site.
Every data point is traceable. Below are the 100+ sources underpinning this dataset, organized by domain. All data was verified as of February 2025.