Why learning curves matter (and how to architect for them)

History shows learning rates matter, but not all technology has benefited. So how do you get on the right trajectory?

Reading Casey Handmer’s blog last week (which is worth checking out) reminded me why learning rates matter, and why they can be under-appreciated and somewhat misunderstood.

Casey suggested that the learning rate for solar panels has increased from roughly 26% to around 48%! That means that every time cumulative production of solar panels doubles, costs fall by roughly 48% - which is nuts!!

This remarkably consistent pattern of cost reduction has held for over four decades, transforming solar from an expensive curiosity into the cheapest form of electricity generation in history.

The same learning curve dynamics appear across multiple technologies. Lithium‑ion battery costs have fallen by ~90% since 2010. DNA sequencing costs likewise collapsed: NHGRI’s long‑running benchmark shows sequencing costs dramatically outpacing Moore’s Law starting around January 2008 with the shift to next‑generation sequencing - Illumina now market $200 “genomes”. Semiconductors historically delivered steep cost reductions, with the cost of manufacturing a transistor falling by roughly 20–30% per year on average during the Moore’s‑Law era. And LED lighting exhibits similar scale effects: DOE documented LED A‑lamp prices falling from $250/klm in 2008 to $40/klm in 2012, with mass-market bulbs reaching ~$2 by the mid‑2010s.

This phenomenon is known as Wright’s Law, or experience curves, and it can be critically important for hardware startups in particular.

The reason to write this blog is to highlight that this isn’t an exogenous static factor where only some technologies or companies can benefit: founders have agency to make choices that architect themselves onto new or better learning curves.

In other words, companies have a better chance of winning if they design their businesses to move down learning curves faster than their competitors.

Learning curves and Wright’s Law

Wright’s Law states that costs decline at a predictable rate as cumulative production increases. Every doubling of production volume tends to reduce costs by a relatively constant percentage - the “learning rate” or learning curve.

For instance, if a widget costs £100 today at 1m units of production, after one doubling to 2m widgets, a 20% learning rate brings the cost to £80. After the next doubling to 4m units, the cost is £64. After five doublings, the cost would be £32. You can see how powerful this can be if it takes hold.

The mechanisms behind learning curves are fairly tangible, but they operate at three distinct levels that compound together.

At the worker level, repetition directly improves speed and quality. Wright’s seminal 1936 study of aircraft assembly documented an 80% learning rate, meaning that each time cumulative production doubled, labour hours per unit dropped to 80% of the previous level. This 20% improvement with each doubling has since been observed across multiple industries: a comprehensive EPA review of 22 field studies covering 108 observations across manufacturing sectors found that most progress ratios clustered around 80%, while Uzumeri and Nembhard’s 1998 study fitted individual learning curves to 3,874 episodes of worker performance in a large US manufacturing firm, demonstrating that these improvements are both measurable and persistent across different workers and tasks.

At the organisational level, companies discover and solve bottlenecks that only become visible at scale. This is distinct from workers simply getting faster, it’s about the production system itself evolving. When solar panel manufacturers doubled production in the 1980s and 1990s, they discovered silicon processing bottlenecks invisible at low volumes. Solving these pushed yields from 60% to over 90%. Engineers optimised anti-reflective coatings through thousands of iterations, each refinement building on the last. Studies of assembly operations show that training methods and process improvements can significantly affect the learning curve trajectory, with effects differing between early learning phases and steady state performance. Unlike human learning, which eventually plateaus, organisational learning can compound indefinitely as cumulative production reveals new optimisation opportunities.

At the economic level, volume unlocks capabilities through pure scale effects. The £1m robotic assembly system becomes economical not because anyone learned anything new, but because the fixed cost per unit drops when amortised across millions of units. Automated testing equipment that seemed prohibitively expensive for small volumes became essential and economical at larger scale. Equipment suppliers see increased demand and invest in better tools. Materials suppliers scale up and reduce costs. Bulk purchasing power reduces input costs. These are economies of scale rather than learning per se, but they’re just as important to the overall cost trajectory.

These patterns are clear in battery manufacturing, with all three mechanisms operating simultaneously. Workers refined manual assembly techniques, becoming faster and more precise through repetition (human learning). Engineers developed roll to roll coating processes, automated electrode stacking, and precision electrolyte filling systems (organisational innovation). And equipment that required significant upfront investment paid dividends across millions of units (economies of scale). Each innovation required significant upfront investment but the returns scaled with volume.

Learning also happens at every level of the supply chain simultaneously. When semiconductor fabs double production, line operators become more skilled at equipment operation, the organisation discovers defects in photolithography equipment and optimises chemical vapour deposition processes, and increased volumes justify investments in better tools and bulk material purchases. The entire ecosystem improves together.

This is why Wright’s Law scales with cumulative production rather than time. You need the actual experience of making millions of units to unlock all three mechanisms. Computer simulations help, but nothing replaces the feedback loop of high volume manufacturing. Each doubling reveals new optimisation opportunities that weren’t visible or economically relevant at smaller scale.

Wright’s Law explains the divergent paths of solar and nuclear.

Until around 1970, nuclear power showed positive learning rates of 23% in the US and 30% in Britain. Then something broke. From 1970 onwards, learning rates became deeply negative across the Western world: negative 94% in the US, negative 82% in Germany, and negative 23-56% in Britain. The crucial difference was volume. Countries that kept building saw costs stabilise or even fall. Countries that stopped building saw costs explode.

South Korea exemplifies the power of sustained building. By constructing 8-12 identical reactors in sequence for each design, they maintained stable supply chains, preserved skilled workforces between projects, and refined construction techniques across multiple builds. Their cost: approximately £2.5 million per megawatt.

Britain took the opposite path (great overview here).

Plants constructed before 1995 averaged £4.79m per megawatt - roughly average internationally and half the cost of current projects. But the combination of stop-start construction patterns, regulatory complexity that prevented standardisation, and planning paralysis created a perfect storm.

After building 26 Magnox reactors between 1956 and 1971, followed by 14 Advanced Gas-Cooled Reactors (AGRs) between 1976 and 1989, construction stopped entirely for 28 years. The AGR programme itself was catastrophic - for political reasons, three different consortia built completely different detailed designs, preventing any learning transfer between projects. Dungeness B, ordered in 1965 with a planned 1970 completion, finally generated power in 1983, a whole 13 years late and four times over budget in inflation-adjusted terms.

The 28-year construction gap devastated Britain’s nuclear capabilities. Supply chains atrophied, specialist skills disappeared, and by the time Hinkley Point C began, Britain had to rebuild everything from scratch, but without any benefit from standardisation or repetition. Hinkley Point C took 10 years from proposal to construction start, compared to just three years in France for the identical reactor. Sizewell C’s environmental assessment runs to 44,260 pages.

The result is visible in stark cost comparisons. Hinkley Point C uses the identical EPR reactor design as plants in France, Finland, and China, yet costs five times more than China’s Taishan plant, 68% more than Finland’s Olkiluoto 3, and 39% more than France’s Flamanville 3.

The critique

This theory isn’t without some degree of push back. For instance, Matt Clancy wrote a few years ago about how the empirical relationship between cost and cumulative production isn’t as clear cut as it appears. He argues that cost declines often correlate with cumulative production simply because both rise over time, making it hard to isolate true “learning-by-doing” from broader technological progress. His review suggests learning explains only part of cost reduction (40–67%), so startups shouldn’t assume that scaling alone will deliver the headline learning rates seen in historical data. My view at least is that startups not doing i.e. iterating with customer feedback are unlikely to be around to reap the benefits of Matt’s technological progress over time anyway, so it’s somewhat of a moot point in this context.

But regardless of what the exact number is, for now it seems like there is sufficient evidence the phenomenon exists and is worth taking seriously for anyone building hard tech.

Designing for Learning

The important point is that startups can actively engineer themselves onto better learning curves through deliberate architectural and strategic choices. Founders are not stuck with the learning dynamics of their industry, but can choose them by following these principles.

Manufacturing over construction. Technologies manufactured in factories at scale show the strongest learning effects. Factories enable controlled experimentation, rapid iteration, and accumulation of tacit knowledge across shifts and teams. At scale, each production line processes thousands of units daily, generating immediate feedback on process changes. This is fundamentally different from site built infrastructure like bridges or dams, where each project is unique, teams disperse afterwards, and learning dissipates. Factory manufacturing also enables automation investments that only make sense at scale. The robotic assembly system that seems absurd for 100 units becomes essential for 100,000. Small modular reactors are essentially this bet: can we turn nuclear from a construction project into a manufacturing one? Break your system into standardised components that can be manufactured in factories, even if final assembly happens on site.

Standardisation over customisation. Learning compounds when you make the same thing repeatedly. The defence sector illustrates this starkly. Exquisite bespoke systems, custom designed for specific requirements, remain expensive despite decades of development. Each new variant requires fresh engineering, preventing systematic cost reduction. Early customers will want bespoke solutions. Each customisation kills your learning curve because you’re not making the same thing repeatedly. Ford’s Model T strategy, “any colour as long as it’s black”, isn’t just historical trivia. It’s what is required to benefit from learning curves. Say no to customisation, even if it costs you early sales.

Technology stability. The core technology must be stable enough that improvements compound. Crystalline silicon solar has maintained fundamentally the same architecture for decades, allowing thousands of incremental improvements to accumulate. PERC cells, anti reflective coatings, bifacial modules, these innovations built on a stable foundation. Compare this to early stage technologies where the architecture keeps changing, PERC versus thin film, organic versus perovskite, making it impossible for learning to compound effectively. Semiconductors tell the same story: CMOS has remained the dominant transistor architecture for over fifty years, enabling continuous gains through lithography, materials, and process control without repeatedly resetting the learning curve. Pick an architecture with a long runway of incremental improvements rather than constantly pivoting to completely new approaches.

Control critical learning curves. Complex products with many components create opportunities for learning across the value chain, but the strategic choice is whether to capture that learning internally or rely on suppliers. Vertical integration makes sense when components are critical to your cost structure and learning rates directly determine competitiveness. For less critical components, deep supplier partnerships work if your volumes justify their investment in process improvements. The risk is depending on commodity suppliers for important parts. If you’re buying standard components at market prices, you benefit from industry wide learning but have no influence over the pace of improvement. Either vertically integrate critical components or commit to volumes that let suppliers invest in their own learning curves.

Design for volume from day one. Use abundant materials, not exotic ones, even if exotic performs better initially. Design production processes that can scale 100x without fundamental changes. Avoid manual assembly steps that can’t be automated. Resource constraints kill learning curves. If your technology depends on scarce materials or ideal locations, learning curves flatten quickly. Geothermal power faces this challenge. The best geothermal sites, high temperature reservoirs close to the surface, are limited. Even if drilling techniques improve, later projects must tackle progressively more difficult geology.

Optimise for iteration speed. Long development cycles mean slow learning. James Dyson built 5,127 prototypes over four years before reaching his final vacuum design. While ultimately successful, traditional manufacturing meant each iteration took weeks. Contrast this with the evolution of drones in Ukraine, where designs are modified, field-tested, lost, and improved in quick succession, compressing build–test–learn cycles into days. Modern startups using rapid prototyping increasingly operate closer to this model, breaking systems into subsystems that can be tested independently and iterated continuously. How quickly can you go from design to real-world feedback? What limits the frequency of your learning cycles? The faster you complete build–test–learn loops, the quicker you move down the learning curve.

Choose your beachhead carefully. Find initial markets where customers will pay premium prices, not because revenue drives learning directly, but because it allows you to deploy enough units to learn while costs remain high. Solar started with space applications where price was irrelevant, moved to off-grid applications, and eventually reached grid scale as costs fell. The beachhead must provide enough volume to generate learning while accepting higher initial costs.

The Strategic Choice

The key point is that you get to choose which learning curve you’re on. Different design choices put you on fundamentally different trajectories.

It’s important to recognise that technological superiority isn’t enough. You need to architect your solution so that each unit you build makes the next one cheaper and better. You need to design systems where learning compounds rather than resets with each project.

This is hard. It often means making tradeoffs against pure technical performance. It means saying no to lucrative customisation opportunities. It means choosing abundant materials over optimal ones. But these architectural choices, made early, determine whether you’re building a company that gets cheaper and better with scale, or one where each project is as expensive as the last.

Wright’s Law isn’t destiny. It’s a choice encoded in your technology architecture and business model. The question for deep tech founders isn’t “will we benefit from learning curves?” It’s “which learning curve do we want to be on, and how do we get there?”

—

Thanks for reading! Please do subscribe and share…