Wright's Law

How do we get better at things? From mastering a craft to manufacturing a jet, the principle of "learning-by-doing" is fundamental to human progress. This intuitive idea is not just a platitude; it can be described by a powerful mathematical relationship known as Wright's Law. This law moves beyond the simple notion that things get cheaper over time, addressing a more fundamental question: what is the true engine of progress? It posits that cost is a function not of the calendar, but of cumulative experience, revealing why proactive investment and production are critical for innovation. This article unpacks the power of this concept in two main parts. The first part, "Principles and Mechanisms," will explore the mathematical foundation of Wright's Law, contrasting it with other models of progress and examining its nuances, such as the role of forgetting and the dynamics of complex systems. The second part, "Applications and Interdisciplinary Connections," will demonstrate the law's profound impact on strategic decision-making in fields ranging from medicine and climate policy to competitive business environments. We will see how this simple rule guides our ability to forecast the future and, more importantly, to actively shape it.

At the heart of every great achievement, from the crafting of a Stradivarius violin to the mass production of the silicon chip, lies a universal, almost musical, truth: practice makes perfect. More than just a folk wisdom, this principle of "learning-by-doing" can be described with surprising mathematical elegance. The first time you try to bake a complex cake, you follow the recipe with painstaking slowness, ingredients are measured nervously, and the result might be... educational. The hundredth time, your hands move with an ingrained rhythm, your intuition for temperature and texture is sharp, and the outcome is reliably delicious. The cost—in time, effort, and wasted ingredients—plummets. This is the essence of ​​Wright's Law​​. It tells us that to understand the cost of production, we shouldn't be looking at the calendar, but at the counter. The key variable isn't time; it's ​​cumulative experience​​.

In 1936, Theodore Wright, an engineer studying aircraft manufacturing, noticed a remarkably consistent pattern. He found that the cost to produce an airplane wasn't random, nor did it simply decrease steadily with time. Instead, it decreased by a predictable percentage each time the total number of aircraft ever produced doubled. This observation is the bedrock of Wright's Law, and it is captured in a beautifully simple power-law relationship:

Let’s unpack this. Here, $C(Q)$ is the cost to produce the $Q$-th unit. The term $Q$ isn't the number of units made in a day or a year, but the cumulative total produced since the very beginning. $C_0$ is a starting constant, representing the cost of the very first unit. The star of the show is the exponent, $-\alpha$. The positive number $\alpha$ is called the ​​experience index​​ or ​​learning elasticity​​. It controls how quickly costs fall as experience is gained.

The magic of this power-law form lies in its scale-free nature. It implies a constant "doubling" rule that is incredibly intuitive. The cost reduction you get from going from your 100th unit to your 200th is the same percentage reduction as going from your 1,000,000th unit to your 2,000,000th. This constant factor is called the ​​Progress Ratio (PR)​​. It is the ratio of the new cost to the old cost after a doubling of experience:

This ratio is a constant! It depends only on the experience index $\alpha$. If a technology has a PR of $0.80$, it means its cost drops to 80% of its previous value every time cumulative production doubles. This leads us to the more commonly cited metric, the ​​Learning Rate (LR)​​, which is simply the percentage cost reduction:

So, a Progress Ratio of $0.80$ corresponds to a Learning Rate of $0.20$, or 20%. Let's make this concrete, inspired by the dramatic scale-up of penicillin production during World War II. Imagine a new antibiotic has a Learning Rate of 20% ($PR = 0.80$) and the first unit costs a hypothetical $100. After the first doubling of production, the cost falls to$100 \times 0.80 = $80$. After a second doubling, it falls again to$80 \times 0.80 = $64$. In just two doublings, the cost has been reduced by$36. This relentless, predictable march downwards is what makes Wright's Law a powerful engine of technological progress. And this law is symmetric: if for some reason a technology's effective experience base were to be halved—perhaps through the loss of institutional memory—the cost would increase by a factor of $2^\alpha$, or $\frac{1}{PR}$.

It is tempting to confuse learning-by-doing with the general passage of time. After all, don't things just get better over the years? This brings up a crucial distinction between Wright's Law and what we might call a "Moore's Law" type of progress. Moore's Law, in its original form, observed that the number of transistors on a chip doubled roughly every two years. Its analogy in cost modeling is an exponential decay with time: $C(t) = C_0 \exp(-\lambda t)$. Here, cost falls at a constant rate $\lambda$ with each passing year, regardless of how many units are actually produced.

So, which is it? Is cost reduction driven by the hands-on experience of production (​​endogenous learning​​), or by the abstract march of scientific progress that happens in the background (​​exogenous learning​​)? The answer depends on the technology.

• For a technology like ​​solar photovoltaics​​, the story is overwhelmingly one of Wright's Law. Massive, policy-driven deployment created a virtuous cycle: increased production led to lower costs, which spurred even more demand and production. The learning was internal to the industry's own activity.

• Contrast this with a highly specialized scientific instrument. Its cost might fall over time not because many are being made, but because the lasers, processors, and materials science that it relies on are all improving due to R&D in other fields. This is a better fit for a time-based model.

The difference is not merely academic; it's a fundamental question of causality. Imagine two factories making the same product. Factory A ramps up production quickly, reaching a cumulative output of one million units in five years. Factory B is more cautious and takes ten years to produce the same one million units. At the moment each factory hits the one-million-unit mark, what will their costs be?

• ​​Wright's Law predicts:​​ Their costs will be the same. The crucial variable is the cumulative experience ($1,000,000$ units), not the time it took to get there.
• ​​Moore's Law predicts:​​ Factory B's costs will be lower, because more time has passed for exogenous innovation to occur.

This conceptual test reveals the core identifying assumption of Wright's Law: conditional on cumulative output $Q$, cost is invariant to the time $t$. This also exposes a subtle trap for analysts. If a technology's production happens to grow exponentially over time (say, $Q(t) \propto \exp(gt)$), then Wright's Law, $C \propto Q^{-\alpha}$, becomes $C(t) \propto (\exp(gt))^{-\alpha} = \exp(-\alpha gt)$. This looks identical to a time-based Moore's Law! Without understanding the causal driver, one could easily mistake endogenous learning for an exogenous "gift of time."

Few modern technologies are monolithic. A car is an assembly of an engine, a chassis, electronics, and thousands of other parts. If each component learns at its own pace, what is the learning rate of the car as a whole?

Suppose the system cost is the sum of its component costs: $C_{system} = \sum m_i c_i$, where $m_i$ is the number of units of component $i$ and $c_i$ is its cost. If each component $i$ follows its own Wright's Law, $c_i \propto N_i^{-\alpha_i}$, where $N_i$ is the cumulative production of that component and $\alpha_i$ is its learning exponent. The surprising truth is that a sum of different power-law functions is not, in general, a single power-law function itself.

This seems to shatter the elegant simplicity of Wright's Law for complex systems. But nature has a beautiful trick up its sleeve. As cumulative production of the system ($N_s$) becomes very large, the system's cost curve asymptotically begins to look like a single power law. And the exponent of this emergent system-level law is dictated by the ​​smallest learning exponent of its components​​ ($\alpha_{min}$).

This is a profound insight. The long-term cost reduction of a complex system is ultimately bottlenecked by its most stubborn, slowest-learning part. It doesn't matter if your microchips are getting cheaper by 30% with every doubling if the cost of the copper wiring is only decreasing by 3%. As production scales, the cost of the copper will come to dominate the system's cost profile and dictate its overall learning rate. An orchestra, in the long run, can only improve as fast as its slowest-learning musician. This tells us precisely where to direct our innovation efforts: not at the parts that are already learning quickly, but at the ones that are holding everything else back.

The simple model of Wright's Law assumes that experience, once gained, is permanent. But is it? A factory that is shuttered for a decade and then re-opened will have lost skilled workers, institutional memory, and supplier relationships. This phenomenon is often called ​​organizational forgetting​​.

We can refine our model to account for this. Imagine the stock of "effective experience," $E$, as water in a leaky bucket. The production rate, $q$, is the water flowing in from a tap. Forgetting is a leak, an outflow proportional to the amount of water already in the bucket, $\phi E$, where $\phi$ is the forgetting rate. The rate of change of experience is then:

If production continues at a constant rate $q$, the experience level doesn't grow to infinity. Instead, the water level rises until the inflow from the tap exactly balances the outflow from the leak. This happens at a steady-state experience level $E^* = \frac{q}{\phi}$.

The implication is startling. Because experience saturates at a finite level, the cost reduction also grinds to a halt. The cost doesn't fall forever toward some theoretical minimum; it gets stuck at an asymptotic level determined by the balance between learning and forgetting. To drive costs down further, a society or a company must either increase the rate of production ($q$) or find ways to plug the leak (reduce the forgetting rate $\phi$ through better knowledge management). This dose of realism tempers the utopian promise of infinite progress, showing that continuous effort is required just to maintain our hard-won experience.

If learning is so powerful, how should it shape our strategies? Should we only invest in technologies that are already cheap, or should we nurture new ones that are currently expensive but have high potential for learning?

Dynamic programming offers a rigorous answer through the concept of the ​​shadow value of experience​​. This concept reveals that the true "cost" of producing one unit of a new technology today is not just its sticker price. The true cost is the sticker price minus the value of the experience gained from making it, because that experience makes all future units cheaper.

This future benefit, the shadow value, is the discounted sum of all cost savings that will ever accrue from the small piece of experience gained today. For an optimal production path, this value, $V'(Q_t)$, can be expressed as:

where $\alpha$ is the learning elasticity, $\beta$ is the discount factor, and $c(Q_{t+k})$ and $x_{t+k}$ are the future costs and production levels.

The sign of this value is negative, meaning more experience decreases total future costs—experience is a valuable asset. The magnitude tells us exactly how valuable it is. This formalizes a powerful strategic idea: it can be perfectly rational for a society to subsidize an expensive new technology, like early solar power or electric vehicles. The initial "losses" are not losses at all; they are a strategic investment. They are the price we pay to "buy down the cost curve," an investment that pays dividends in the form of cheaper, better technology for all future generations. Wright's Law, therefore, is not just a descriptive tool for historians of technology; it is a prescriptive guide for architects of the future.

In our journey so far, we have uncovered a surprisingly simple and elegant pattern woven into the fabric of human endeavor: the more we do something, the better and more efficiently we get at it. Wright’s Law gives this rhythm of progress a mathematical voice, describing how costs decline with cumulative experience. We have seen its form, $C(Q) = C_0 (Q/Q_0)^{-\alpha}$, and explored the mechanisms that give rise to it. But this is far more than a historical curiosity or a neat formula. It is a powerful lens through which we can view the future, devise strategies, and even shape the world. Let us now explore the far-reaching consequences of this simple law as it echoes across industries, laboratories, and even the halls of government.

At its heart, Wright's Law is a tool for prediction. For decades, engineers and economists have used it to forecast the cost of everything from aircraft to solar panels. By observing the learning rate of a young technology, one can make remarkably good predictions about its future economic viability. You might think this applies only to factory assembly lines, but the same principle is at work in the most advanced laboratories on Earth.

Consider the field of precision medicine, where diagnostics are becoming as complex as therapeutics. The cost of Next-Generation Sequencing (NGS), a technology that allows us to read a person's genetic code, is a perfect example. Early on, sequencing a genome was a monumental and expensive task. But as laboratories processed more and more samples, they learned. They standardized workflows, automated bioinformatics pipelines, and built reusable libraries of curated genetic variants. Each of these improvements contributes to a learning curve. By applying Wright's Law, a health system can project how the per-test cost will fall as its cumulative volume of tests grows, allowing for more informed decisions about when and how to adopt these powerful new diagnostic tools.

This principle extends to the very frontier of medicine: gene-editing therapies like CRISPR. These treatments, which involve custom-engineering a patient's own cells, are fantastically expensive at their inception. A hypothetical but plausible first-in-class therapy might launch with a variable cost of over half a million dollars per patient. For most of humanity, a therapy at that price might as well not exist. But Wright's Law offers a glimmer of hope. As facilities treat more patients, they gain experience. They improve cell culture yields, shorten processing times, and reduce failures. Each doubling of cumulative experience shaves a fraction off the cost. Our model, based on realistic learning rates, shows that scaling production from one thousand to eight thousand patients a year could cut the average cost per patient by more than half.

However, the history of other advanced biologics, like monoclonal antibodies, teaches us a sobering lesson: meaningful cost declines can coexist with persistent inequities in access. The specialized facilities and expert workforce required for these therapies tend to be concentrated in a few locations, creating bottlenecks and leaving many regions behind. The journey from a scientific breakthrough to equitable global access is not guaranteed by Wright's Law alone; it remains a profound social and logistical challenge.

In a world governed by Wright's Law, standing still is falling behind. But moving too early can be expensive. This creates a beautiful strategic tension, forcing us to ask: when is the right time to invest in a new technology?

Imagine you are a planner tasked with building a nation's hydrogen production capacity to meet a growing demand over the next decade. The technology, electrolyzers, is subject to a learning curve—it will get cheaper over time as more are produced globally. Do you build all the capacity you'll need right now to get a head start, or do you build it piecemeal, just as it's needed? The mathematics of optimization, when combining the effects of technological learning with the time value of money (a dollar today is worth more than a dollar tomorrow), provides a surprising and wonderfully elegant answer. The optimal strategy is to build "just-in-time". By waiting until the last possible moment to add capacity, a planner benefits from both the lower capital costs produced by global progress and the financial advantage of delaying expenditure. There is no incentive for speculative, early overbuilding.

But what if you are not a benevolent central planner, but a firm in a competitive market? Here, the situation becomes far more complex and interesting. Wright's Law gives rise to one of the most powerful phenomena in economics: ​​path dependence​​. The future becomes a product of the contingent twists and turns of the past.

Consider two competing technologies, A and B. Technology A starts out slightly cheaper, but Technology B has a faster learning rate—it gets cheaper more quickly with experience. In the beginning, Technology A's cost advantage gives it a larger market share. This larger market share gives it more cumulative production experience, which, via Wright's Law, drives its costs down further. This creates a self-reinforcing feedback loop. Technology A's early, perhaps even accidental, advantage can "lock in" its dominance, potentially squeezing the superior long-run technology (B) out of the market entirely. Early moves, even small ones, can have massive and irreversible consequences, illustrating how markets can sometimes "choose" a suboptimal long-term outcome.

If firms can strategize with Wright's Law, so can governments. Indeed, the law is a cornerstone of modern energy and climate policy modeling.

The adoption of a new technology is not a linear march; it often follows an S-shaped diffusion curve. Initially slow, it accelerates rapidly before leveling off as the market becomes saturated. But the speed of this diffusion is not fixed. Because lower costs drive faster adoption, and faster adoption drives down costs, a virtuous cycle is born. Endogenous learning—progress driven from within the system—acts as a powerful accelerator for technological transitions. A policy that jump-starts this cycle can have an outsized impact.

But here is a deeper insight: policy doesn't just operate on a fixed learning curve; it can change the slope of the curve itself. The institutional environment matters. Consider two policies to support a new clean technology: a fixed-price Feed-in Tariff (FiT) that guarantees a certain payment, versus a competitive auction where firms bid to provide power at the lowest cost. The FiT provides security but can allow for slack, potentially slowing the rate of innovation. The auction, however, forces firms to relentlessly pursue cost reductions to outbid their rivals. This intense competition can steepen the learning curve, accelerating cost decline beyond what would happen on its own. The policy's design can tune the very rhythm of progress.

These individual pieces—the learning curve, investor choice, stock turnover, and policy design—are the building blocks of the large-scale simulation models that inform critical debates about climate policy. Should we implement a carbon tax or a technology mandate? Answering this question requires simulating the complex interplay of these forces. Such models can quantify the trade-offs: a carbon price might be more economically efficient, but a mandate might provide more certainty in achieving a specific deployment goal by a specific date. These tools allow us to distinguish what environmental science can predict about the outcomes of a policy from what environmentalism may demand as a normative goal, providing clarity in highly charged public debates.

As powerful as this framework is, we must wield it with care and humility. The real world is always messier than our models. The very act of measuring a learning curve can be fraught with difficulty.

Consider a surgeon learning a new, complex minimally invasive procedure. Her operative time will naturally decrease as she performs the procedure over and over—a perfect example of a personal learning curve. Now, imagine a clinical trial designed to see if this new technique is faster than the established standard. If the study is designed naively—for instance, by having the surgeon perform the new technique for the first 50 cases and the standard technique for the next 50—the comparison is hopelessly biased. You are not comparing the two techniques on a level playing field; you are comparing the surgeon's performance as a novice in the new technique to her performance as an expert in the old one. The learning effect itself becomes a confounding variable that threatens the internal validity of the study, making it impossible to draw a correct causal conclusion about the techniques themselves. This illustrates a profound challenge in empirical science: sometimes, the very phenomenon we are trying to understand can obscure our ability to measure other things accurately.

This brings us to a final, crucial lesson about the nature of knowledge itself. To treat progress as a mysterious, external force—an "exogenous" trajectory of costs that fall from the sky, independent of our actions—is to fundamentally misunderstand our own role in creating the future. It's not just a technical modeling choice; it's a worldview. An "endogenous" model, where progress is the result of our collective actions, investments, and policies, recognizes that the future is not something we merely predict, but something we actively build.

To ignore this feedback loop, where policy influences deployment and deployment drives down cost, is to create models that are fragile and advice that is flawed. It can lead one to underestimate the power of policy to accelerate change or to misjudge the long-term benefits of an early investment. This is a modern echo of the Lucas Critique from economics: a model that does not account for how agents respond to changes in their environment is bound to fail when used to evaluate new policies. The choice to see progress as endogenous is the choice to see ourselves as agents of change, capable of bending the arc of progress toward a more desirable future.