Jorge Quitério Search
Technology

Wright's law and Moore's law: experience-based and time-based models of progress

Moore's law relates cost decline to time; Wright's law relates it to cumulative production. When production grows exponentially the two laws are equivalent, which is why historical data rarely distinguish them. A simulation shows how they diverge when production stalls.

In 1936 the aeronautical engineer Theodore Wright observed that, in aircraft manufacturing, every time the cumulative number of aircraft built doubled, the labour needed per aircraft fell by a roughly constant percentage. In 1965 Gordon Moore observed that the number of components on a chip doubled at a regular pace, year after year.

One law relates progress to experience, the other to time. They are often presented as competing theories of technological progress, and the difference matters for strategy. Under Moore's law, progress arrives with time regardless of what is done. Under Wright's law, progress is obtained through production, so supporting deployment lowers costs. The question is whether data can distinguish the two.

Consumers see the outcome in the prices of solar panels and electric-car batteries. Solar modules cost more than 99% less per watt than in the mid-1970s, and lithium-ion battery packs cost roughly 90% less per kilowatt-hour in 2020 than in 2010. Whether such declines continue, and what drives them, depends on which law describes them.

The question has a long history. Wright's law, also called the experience or learning curve, became a staple of manufacturing and of strategy consulting. Moore's law, generalised to "performance improves exponentially with time", became the default assumption for everything from batteries to genome sequencing; Magee and colleagues (2016) found exponential trends with time in all 28 technological domains they studied, at very different rates. In 1979 Devendra Sahal showed that if production grows exponentially, the two laws are mathematically the same. Nagy, Farmer, Bui and Trancik (2013) tested the candidates on historical data for dozens of technologies by hindcasting. Wright's law forecast best, Moore's law was close behind, and the reason was Sahal's: production does tend to grow exponentially.

The two laws

Write ctc_t for unit cost, QtQ_t for cumulative production and tt for time.

Moore:ct=c0 e−mt,Wright:ct=c0(QtQ0)−b. \text{Moore:}\quad c_t = c_0\, e^{-m t}, \qquad\qquad \text{Wright:}\quad c_t = c_0 \left(\frac{Q_t}{Q_0}\right)^{-b}.

Wright's exponent is usually reported as a learning rate: the cost falls by 1−2−b1 - 2^{-b} every time cumulative production doubles. A learning rate of 20% means b=−log⁡20.8≈0.32b = -\log_2 0.8 \approx 0.32.

Sahal's equivalence

Suppose production grows at a constant rate gg, so that output in year tt is q0egtq_0 e^{g t}. Cumulative production is then

Qt≈q0g egt⟹ln⁡ct=const−b g t. Q_t \approx \frac{q_0}{g}\, e^{g t} \quad\Longrightarrow\quad \ln c_t = \text{const} - b\,g\,t .

That is Moore's law with m=b gm = b\,g. A technology learning at 20% per doubling while output grows 30% a year will show costs falling at about e−0.32×0.3−1≈−9%e^{-0.32 \times 0.3} - 1 \approx -9\% a year, and nothing in the cost history alone says whether the calendar or the factory did it.

Practical implications

The two laws agree as long as production keeps growing exponentially. They part ways when it does not: a market saturates, a subsidy ends, a war doubles demand. Wright predicts that a stall in production stalls the cost decline; Moore predicts the decline carries on. Policy that aims to bring a technology down its learning curve (solar panels, batteries, electrolysers) is a bet on Wright.

The difference also affects a household deciding whether to buy a heat pump now or wait for prices to fall. Under Moore's law, waiting always brings a lower price. Under Wright's law, the price falls only as fast as installations accumulate, so if subsidies end and sales slow, prices stop falling and waiting brings little.

The figure below fits both laws to the same simulated data. With exponential production, the fitted Moore rate equals bgb g and both laws forecast equally well. When production stalls, Moore's forecast fails and Wright's does not.

  1. Sahal's identity. With the defaults (production +30% a year, learning rate 20%), compare "Moore: cost change per year" with "Sahal's prediction bgb g". Both sit near −9%. With new data, the fitted value varies with the noise while the prediction does not.
  2. Equal forecast accuracy. Look at the forecast errors for years 15-29. Moore and Wright miss by similar amounts, typically 10-20% with the default noise. Hindcasting cannot distinguish the two laws here, which is consistent with what Nagy and colleagues found in real data.
  3. Production stall. Switch on "Production stalls after year 15". Both laws are still fitted to the first 15 years only, so their parameters barely change. But Moore's forecast now misses by 50-80%, while Wright's stays around 10%, because the experience curve uses the actual production and therefore reflects the slower learning.
  4. Fast learning, slow growth. Set the learning rate to 30% and growth to 10%. The yearly decline follows the product bg≈0.05b g \approx 0.05, about −5%. A technology with a high learning rate in a slowly growing market improves slowly over time.

Why the two laws coincide. With qs=q0egsq_s = q_0 e^{g s}, cumulative production is Qt=Q−∞+∫0tq0egs dsQ_t = Q_{-\infty} + \int_0^t q_0 e^{g s}\,ds. Starting the history from the exponential past gives Qt=(q0/g) egtQ_t = (q_0/g)\,e^{g t} exactly, so ln⁡Qt=ln⁡(q0/g)+gt\ln Q_t = \ln(q_0/g) + g t. Substituting into ln⁡ct=ln⁡c0−bln⁡(Qt/Q0)\ln c_t = \ln c_0 - b \ln(Q_t/Q_0) gives ln⁡ct=const−bgt\ln c_t = \text{const} - b g t. The figure uses discrete years and noise, so the match is close rather than exact.

When a technology's cost curve is a straight line on a log scale, check what production did over the same period. If it grew exponentially, the data support both laws equally and cannot choose between them. The difference shows up when demand changes, which is exactly when planners need a forecast, and in that situation the experience curve is the more reliable model.

References

  1. T. P. Wright, "Factors affecting the cost of airplanes", Journal of the Aeronautical Sciences, 3(4), 122-128, 1936.
  2. G. E. Moore, "Cramming more components onto integrated circuits", Electronics, 38(8), 114-117, 1965.
  3. D. Sahal, "A theory of progress functions", AIIE Transactions, 11(1), 23-29, 1979.
  4. B. Nagy, J. D. Farmer, Q. M. Bui and J. E. Trancik, "Statistical basis for predicting technological progress", PLoS ONE, 8(2), e52669, 2013.
  5. C. L. Magee, S. Basnet, J. L. Funk and C. L. Benson, "Quantitative empirical trends in technical performance", Technological Forecasting and Social Change, 104, 237-246, 2016.